Answers from four API configurations of AI assistants, collected on October 8, 2026 (11:49 to 12:06 UTC), with Google AI Overviews pulled the same day. This article reports how answers changed across repeated calls in a larger study of 11 executive and consulting cost questions, published as 4 AI Assistants, 11 Executive and Consulting Cost Questions.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This article links to the author’s fractional COO and fractional CMO service pages, which may benefit from search traffic it earns. The author’s own domain appears among the references presented by Perplexity and Google’s AI Overviews for some of the study’s questions, as reported in the full study. This analysis does not test what those services cost or are worth. The question set, the calls, and the analysis rules were fixed before the main data pull. Five post-collection changes are listed under Method: a parser fix, a calibration check coded by the study’s AI producer instead of a person, one preregistered check that was not met as written, a rounding implementation difference that changed no figure, and the title confirmation.

These are AI-generated statements, not market prices. The study does not test whether any answer is accurate.

How the repeat test worked

Each of 11 cost questions went to four API configurations ten times: OpenAI (openai/gpt-5.6-terra), Google (google/gemini-3.8-flash), Anthropic (anthropic/claude-sonnet-5.5), and Perplexity (perplexity/sonar-pro). The question wording, settings, and configuration were identical across the ten calls, and the calls were made in a shuffled order within one 17-minute window.

From each answer, a parser took the first price range and its unit. A cell is one configuration, one question, and one unit with at least 8 closed ranges among its 10 calls. There were 27 such cells.

Two preset tests measured how much a cell varied:

• Max ÷ min: the highest midpoint divided by the lowest. The threshold was 1.25.
• Share differing: the share of ranges that differed from the cell’s most common range. Endpoints were first rounded to the nearest $10 below $1,000 and to the nearest $100 at or above $1,000. The threshold was 30%.

A cell counted as varying if it met either test. The rounding grid changes at $1,000. Under the preset sensitivity check with a single $100 grid, the count is 19 of 27.

Results

21 of the 27 cells varied under at least one test. Two cells, Google on Q07 and Google on Q08, had no single most common range, and both met the max ÷ min test.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

These results were first reported in the full study, where the questions are listed. Q01 to Q11 cover fractional CMO, fractional COO, CMO, COO, and business consultant costs.

The widest and narrowest cells

The widest cells:

• Anthropic, Q07 (“How much does a fractional CMO cost per month?”): lowest monthly midpoint $3,500, highest $11,500, a max ÷ min of 3.29.
• Anthropic, Q04 (“How much does a fractional CMO charge?”): lowest $3,500, highest $9,000, 2.57.
• Google, Q08 (“How much does a fractional COO charge?”): lowest $4,000, highest $10,000, 2.50, with no single most common range.
• Anthropic, Q05 (“How much does a business consultant usually cost?”): lowest hourly midpoint $75, highest $187.50, 2.50.

The narrowest cells:

• OpenAI, Q10 (“What is the average cost of a fractional COO?”): lowest $10,000, highest $10,500, 1.05.
• Google, Q01 and Perplexity, Q07: both 1.09.

No two answers were word for word the same

Within each of the 27 cells, no two of the ten answers had identical text. The preset duplicate share was zero in every cell.

As a descriptive count added after the main results, the number of different first ranges a configuration gave to one question, across its ten answers, ran from 2 to 9. Google gave 9 different first ranges in its 10 answers to “How much does a CMO cost?”

The unit could change between calls

Some configurations switched units from one call to the next on the same question.

• Anthropic:
• Q08 (“How much does a fractional COO charge?”): an hourly range in 6 answers and a monthly range in 4.
• Q10: hourly in 8 and monthly in 2.
• Q11: monthly in 8 and hourly in 2.
• OpenAI:
• Q04 (“How much does a fractional CMO charge?”): hourly in 6 and monthly in 4.
• Q11: monthly in 8 and hourly in 2.

Separately, Perplexity’s Q02 answers gave an hourly range 8 times and a range whose unit could not be resolved twice. The open-ended form could also change. For example, Perplexity ended its first range for Q08 with “+” in 4 of 10 answers and gave a closed range in the other 6.

What the numbers mean (operating judgment)

In this data, one response did not characterize all the responses observed from that configuration during the collection window. The same question could return a different figure, a different unit, or an open-ended range on the next call.

As operating judgment, buyers using an AI tool for a first estimate may find that repeating the question reveals variation, and should note the unit each time. This study does not establish how many repetitions are enough. For example, in the widest cell, the midpoints of the first closed ranges observed across ten calls had a lowest value of $3,500 and a highest of $11,500, a max ÷ min of 3.29. The study does not establish the range of values a future call could return.

What this data does not show

• These are not market prices, and the study does not test accuracy.
• The calls went to API configurations at default settings, not consumer chat apps.
• All calls came from one window, about 17 minutes on October 8, 2026. Variation over longer periods was not tested.
• The study does not test why answers varied.

Questions about this analysis can be sent through the contact page.

Method

The data comes from the full study, 4 AI Assistants, 11 Executive and Consulting Cost Questions, where the complete method, audit, and exclusion tables appear.

• Collection. From 11:49 to 12:06 UTC on October 8, 2026, each of 11 People Also Ask cost questions went to four API configurations ten times, producing 440 API answers, with provider-default settings and no researcher-added system message. There were also 33 Google AI Overview pulls, three per question, made the same day.
• Extraction. A rule-based parser took the first price range, its unit, and an open-ended flag from each answer. The parser matched the AI-arbitrated coding on all fields in 100 of 100 randomly sampled answers (95% lower bound 0.963). Two auditor models from families not under test did the coding and a third model arbitrated. This was not an independent human audit.

Descriptive counts. Counts labeled as added after the main results are computed from the public normalized-records CSV. Distinct first ranges are the unique combinations of low figure, high figure, unit, and open-ended flag within one configuration and question. Identical text is detected with the text-hash column. The registered results in this article were first reported in the full study.

Five post-collection changes, each described in full in the study:

• Calibration coder. The 30-answer calibration was coded by the study’s AI producer (Claude, made by Anthropic, whose model is one of the four tested), not a person.
• Parser fix. A parser fix for escaped dollar signs affected 5 of 440 answers and changed no headline figure.
• Kill rule. One preregistered kill rule required auditor agreement (κ) of at least 0.60 on whether an answer contained a price range. The κ statistic could not be computed because both auditors marked every audited answer as containing a range, with raw agreement of 115 of 115. The kill rule therefore was not met as written. After seeing the data, the producer decided that the study would proceed and recorded that decision as a protocol deviation.
• Rounding. The analysis code rounded tie endpoints to the even value instead of half-up. A half-up check changed no cell and no figure.
• Title. The study title was shortened to fit the site’s limit.

Data.

• Normalized records (CSV)
• Question selection (CSV)
• Public protocol page

Related analyses:

• COO Cost Questions: How 4 AI Assistants Answered
• Business Consultant Fees: How 4 AI Assistants Answered
• Which Sites AI Tools List for Executive and Consulting Cost Questions

Google did not review or endorse this analysis. OpenAI, Google, Anthropic, and Perplexity did not review or endorse it. Model names are used only to identify the configurations tested.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This analysis does not test what those services cost or are worth.

Answers from four API configurations of AI assistants, collected on October 8, 2026 (11:49 to 12:06 UTC), with Google AI Overviews pulled the same day. This article reports the reference domains from a larger study of 11 executive and consulting cost questions, published as 4 AI Assistants, 11 Executive and Consulting Cost Questions.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This article links to the author’s fractional CMO and strategy service pages, which may benefit from search traffic it earns. The author’s own domain appears among the references presented by Perplexity and Google’s AI Overviews for some of these questions (counts in the section on the author’s domain). This analysis does not test what those services cost or are worth. The question set, the calls, and the analysis rules were fixed before the main data pull. Five post-collection changes are listed under Method: a parser fix, a calibration check coded by the study’s AI producer instead of a person, one preregistered check that was not met as written, a rounding implementation difference that changed no figure, and the title confirmation.

A reference count shows which domains were listed. It does not show what a page said, whether the answer relied on it, or whether the page is accurate. The study did not read or grade the pages.

What was counted

The study asked 11 People Also Ask cost questions about fractional CMOs, fractional COOs, CMOs, COOs, and business consultants. Of the four API configurations, only Perplexity (perplexity/sonar-pro) supplied the reference data analyzed here. It was asked each question 10 times, for 110 runs. Google AI Overviews were a separate collection, with three pulls per question, for 33 pulls, and 31 of them returned an overview. Reference domains from the OpenAI, Google Gemini, and Anthropic configurations were not analyzed.

For each run or pull, every listed domain was counted once.

The most frequently listed domains

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

Perplexity repeated its sources

On 9 of 11 questions, every reference domain that appeared in any Perplexity run appeared in all 10 runs of that question. The two exceptions:

• Q01 (“How much does a fractional CMO charge per hour?”): salary.com appeared in 7 runs and chatterbuzzmedia.com in 3.
• Q08 (“How much does a fractional COO charge?”): eca-partners.com appeared in 7 runs.

The following are descriptive counts added after the main results:

• Perplexity listed 13 to 17 domains per run.
• AI Overviews listed between 0 and 21 domains per pull.
• Across all questions, Perplexity listed 68 distinct domains and AI Overviews 107.
• 30 domains appeared in both.

Question by question

The table shows how many distinct reference domains each source listed for each question. The Perplexity columns are registered reference frequencies. The AI Overview column is a descriptive count added after the main results.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

AI Overviews spread their references across more domains than Perplexity on most questions, because each pull could list a different set. On Q10, the two pulls that returned an overview listed only three domains between them.

Different questions drew different sources

These groupings are descriptive counts added after the main results. They sum runs or pulls within each group of questions and are not compared across groups.

Descriptive counts added after the main results.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

No Perplexity domain appeared in all four groups. In AI Overviews, only linkedin.com did. The domains are reported by name only. The study did not classify them by type.

The author’s domain

The domain kamyarshah.com is the author’s own site, as stated in the disclosure at the top.

• Perplexity: listed in all 10 runs of each of the four COO questions (Q08, Q09, Q10, and Q11), which is 40 of 110 runs. It was not listed for the CMO or consultant questions.
• AI Overviews: listed in 7 of 31 pulls (Q02 pull 1, Q05 pull 1, Q08 pulls 1 to 3, Q11 pulls 2 and 3).

This shows only that the domain was listed. It does not show what was cited from it or whether any answer used it.

What the numbers mean (operating judgment)

As operating judgment, firms may choose to make relevant service information clear and directly responsive to the questions buyers ask. This study did not test whether any page characteristic causes a domain to be listed, and it provides no method for securing inclusion.

What this data does not show

• Reference counts do not show what a page said or whether an answer used it.
• Pages were not read, graded, or classified.
• All pulls come from one day, October 8, 2026. Sources may change over time.
• The Perplexity results come from its API, not its consumer app.
• Of the four API configurations, only Perplexity supplied the reference data analyzed here. Google AI Overviews were a separate collection.

Questions about this analysis can be sent through the contact page.

Method

The data comes from the full study, 4 AI Assistants, 11 Executive and Consulting Cost Questions, where the complete method and audit appear.

• Collection. From 11:49 to 12:06 UTC on October 8, 2026, each of 11 People Also Ask cost questions went to four API configurations ten times, producing 440 API answers, with provider-default settings and no researcher-added system message. There were also 33 Google AI Overview pulls, three per question, made the same day.
• Extraction. A rule-based parser took the first price range, its unit, and an open-ended flag from each answer. The parser matched the AI-arbitrated coding on all fields in 100 of 100 randomly sampled answers (95% lower bound 0.963). Two auditor models from families not under test did the coding and a third model arbitrated. This was not an independent human audit.
• Reference extraction. Reference domains were taken from the listed references and reduced to the registered domain (for example, www.linkedin.com and linkedin.com count as one). Reference extraction separately matched an AI auditor model’s coding on 20 of 20 checked records.

Descriptive counts. Counts labeled as added after the main results are computed from the public normalized-records CSV. For each record, split the reference-domain field into individual registered domains and remove duplicates within that record. For Perplexity, use the 110 run records. For AI Overviews, use the 31 records that returned an overview, keep a zero where an overview listed no domain, and exclude the two pulls with no overview. Distinct-domain counts are the unique individual domains in each source set. Overlap is the intersection of the two sets. References per run or pull is the number of unique domains in that record. Group totals sum runs or pulls within each question group. The registered results in this article were first reported in the full study.

Five post-collection changes, each described in full in the study:

• Calibration coder. The 30-answer calibration was coded by the study’s AI producer (Claude, made by Anthropic, whose model is one of the four tested), not a person.
• Parser fix. A parser fix for escaped dollar signs affected 5 of 440 answers and changed no headline figure.
• Kill rule. One preregistered kill rule required auditor agreement (κ) of at least 0.60 on whether an answer contained a price range. The κ statistic could not be computed because both auditors marked every audited answer as containing a range, with raw agreement of 115 of 115. The kill rule therefore was not met as written. After seeing the data, the producer decided that the study would proceed and recorded that decision as a protocol deviation.
• Rounding. A rounding implementation difference changed no figure.
• Title. The study title was shortened to fit the site’s limit.

Data.

• Normalized records (CSV)
• Question selection (CSV)
• Public protocol page

Related analyses:

• COO Cost Questions: How 4 AI Assistants Answered
• Business Consultant Fees: How 4 AI Assistants Answered
• The Same AI Price Question, Asked 10 Times

Google did not review or endorse this analysis. OpenAI, Google, Anthropic, and Perplexity did not review or endorse it. Model names are used only to identify the configurations tested.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This analysis does not test what those services cost or are worth.

Answers from four API configurations of AI assistants, collected on October 8, 2026 (11:49 to 12:06 UTC), with Google AI Overviews pulled the same day. This article reports the business consultant questions from a larger study of 11 executive and consulting cost questions, published as 4 AI Assistants, 11 Executive and Consulting Cost Questions.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This article links to the author’s strategy service page, which may benefit from search traffic it earns. The author’s own domain appears among the references presented by Google’s AI Overviews for these questions (counts in the references section). This analysis does not test what those services cost or are worth. The question set, the calls, and the analysis rules were fixed before the main data pull. Five post-collection changes are listed under Method: a parser fix, a calibration check coded by the study’s AI producer instead of a person, one preregistered check that was not met as written, a rounding implementation difference that changed no figure, and the title confirmation.

These are AI-generated statements, not market prices. The study does not test whether any answer is accurate.

The two questions

Each question went to four API configurations ten times: OpenAI (openai/gpt-5.6-terra), Google (google/gemini-3.8-flash), Anthropic (anthropic/claude-sonnet-5.5), and Perplexity (perplexity/sonar-pro). A configuration was eligible if at least 8 of its 10 answers led with a closed range in the question’s most common unit, which was hourly for both questions. All four configurations were eligible on both.

Results

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

Figures are hourly medians of the midpoints of each configuration’s eligible lead ranges.

On both questions, Anthropic gave the lowest median and Google the highest. OpenAI and Perplexity sat between them. Perplexity’s Q02 median is computed from its 8 hourly lead ranges. The other 2 of its 10 answers gave a range with no resolved unit.

Asking the same question ten times

Each configuration was asked each question ten times. Some answered with the same range almost every time, and others did not.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

Anthropic’s ten answers to Q05 spanned midpoints from $75 to $187.50. Google’s answers stayed within $25 of each other on both questions.

What Google’s AI Overviews showed

AI Overviews appeared in all six pulls for these two questions. The first range and unit were recorded from each, and no AI Overview text is reproduced.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

These are descriptive and are not compared statistically.

What a 2.67× gap means in dollars (illustration)

The following arithmetic is an illustration, not a finding. Applied to a hypothetical 40-hour engagement, the lowest and highest Q02 medians imply very different totals:

• At Anthropic’s hourly median of $75, 40 hours comes to $3,000.
• At Google’s hourly median of $200, the same 40 hours comes to $8,000.

A buyer who asked only one configuration would have seen only one of those totals. Neither total is a price that any consultant charges.

How the AI Overview ranges sat against the medians

As a descriptive comparison added after the main results, the closed AI Overview ranges had midpoints of $175 and $250 on Q02, and $225 and $250 on Q05. The other two pulls were open-ended and have no midpoint. Three of those four midpoints were at or above the highest configuration median for the same question. The AI Overview ranges were also wider than the spread of the configuration medians, starting between $50 and $150 and ending between $250 and $400.

References presented

Perplexity and AI Overviews list reference domains. The study counted which domains were listed and did not read or grade the pages.

Perplexity. It listed 15 domains in all 10 runs of Q02 and 17 domains in all 10 runs of Q05, with the set fixed within each question. As a descriptive count added after the main results, ten domains were listed in every run of both questions: agiled.app, alexberman.com, clutch.co, consultingsuccess.com, indeed.com, mailchimp.com, paperbell.com, squareup.com, thumbtack.com, and upwork.com.

AI Overviews. The most frequent reference domains across the six pulls were phoenixmi.org (5 pulls), consultingsuccess.com (4), nrcpas.com (4), thumbtack.com (4), and talkspresso.com (3).

The author’s domain. The domain kamyarshah.com was listed in the first AI Overview pull for each of Q02 and Q05. It was not listed by Perplexity for these two questions. This is the author’s own site, as stated in the disclosure at the top. A reference count shows what was listed, not what the page said or whether the answer relied on it.

What the numbers mean (operating judgment)

An AI answer to a consultant fee question is a starting point for further research, not a market price or a provider quote. In this data, the configuration asked made more than a twofold difference in the median, and some configurations varied widely from one call to the next.

As operating judgment, buyers should define the deliverable, the time frame, and how success will be measured before comparing any fee. For example, a hypothetical $15M manufacturer might compare a fixed-fee diagnostic with an hourly advisory arrangement. Those engagements can produce very different totals even at similar hourly figures.

What this data does not show

• These are not market prices, and they are not what any consultant charges.
• The study does not test accuracy.
• The calls went to API configurations, not consumer chat apps.
• The 440 API answers in the full study came from one window, about 17 minutes on October 8, 2026. The AI Overview pulls were made the same day.
• Only four configurations were tested, all at default settings.

Questions about this analysis can be sent through the contact page.

Method

The answers come from the full study, 4 AI Assistants, 11 Executive and Consulting Cost Questions, where the complete method, audit, and exclusion tables appear.

• Collection. From 11:49 to 12:06 UTC on October 8, 2026, each of 11 People Also Ask cost questions went to four API configurations ten times, producing 440 API answers, with provider-default settings and no researcher-added system message. There were also 33 Google AI Overview pulls, three per question, made the same day.
• Extraction. A rule-based parser took the first price range, its unit, and an open-ended flag from each answer. The parser matched the AI-arbitrated coding on all fields in 100 of 100 randomly sampled answers (95% lower bound 0.963). Two auditor models from families not under test did the coding and a third model arbitrated. This was not an independent human audit.

Descriptive counts. Counts labeled as added after the main results are computed from the public normalized-records CSV. For each Perplexity record, split the reference-domain field into individual domains and remove duplicates within the record. The domains listed in every run of both questions are those present in all 10 Q02 records and all 10 Q05 records. The registered results in this article were first reported in the full study.

Five post-collection changes, each described in full in the study:

• Calibration coder. The 30-answer calibration was coded by the study’s AI producer (Claude, made by Anthropic, whose model is one of the four tested), not a person.
• Parser fix. A parser fix for escaped dollar signs affected 5 of 440 answers and changed no headline figure.
• Kill rule. One preregistered kill rule required auditor agreement (κ) of at least 0.60 on whether an answer contained a price range. The κ statistic could not be computed because both auditors marked every audited answer as containing a range, with raw agreement of 115 of 115. The kill rule therefore was not met as written. After seeing the data, the producer decided that the study would proceed and recorded that decision as a protocol deviation.
• Rounding. A rounding implementation difference changed no figure.
• Title. The study title was shortened to fit the site’s limit.

Data.

• Normalized records (CSV)
• Question selection (CSV)
• Public protocol page

Related analyses:

• COO Cost Questions: How 4 AI Assistants Answered
• Which Sites AI Tools List for Executive and Consulting Cost Questions
• The Same AI Price Question, Asked 10 Times

Google did not review or endorse this analysis. OpenAI, Google, Anthropic, and Perplexity did not review or endorse it. Model names are used only to identify the configurations tested.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This analysis does not test what those services cost or are worth.

Answers from four API configurations of AI assistants, collected on October 8, 2026 (11:49 to 12:06 UTC), with Google AI Overviews pulled the same day. This article reports the COO questions from a larger study of 11 executive and consulting cost questions, published as 4 AI Assistants, 11 Executive and Consulting Cost Questions.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This article links to the author’s fractional COO service page, which may benefit from search traffic it earns. The author’s own domain appears among the references presented by Perplexity and Google’s AI Overviews for these questions (counts in the references section). This analysis does not test what those services cost or are worth. The question set, the calls, and the analysis rules were fixed before the main data pull. Five post-collection changes are listed under Method: a parser fix, a calibration check coded by the study’s AI producer instead of a person, one preregistered check that was not met as written, a rounding implementation difference that changed no figure, and the title confirmation.

These are AI-generated statements, not market prices. The study does not test whether any answer is accurate.

The four COO questions

Each question went to four API configurations ten times: OpenAI (openai/gpt-5.6-terra), Google (google/gemini-3.8-flash), Anthropic (anthropic/claude-sonnet-5.5), and Perplexity (perplexity/sonar-pro). Three of the questions name a fractional COO. The fourth, “How much does a COO cost?”, does not, and its answers are reported separately and never treated as fractional pricing.

A configuration was eligible on a question if at least 8 of its 10 answers led with a closed range in the question’s most common unit. A cross-configuration ratio was computed only where three or more configurations were eligible.

Results

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

Figures are median midpoints of eligible configurations. Exclusion counts in the table are mutually exclusive: an answer in another unit is counted there even if it was also open-ended. Only Q10 had enough eligible configurations for a ratio. Its 1.53× falls at the lower end of the five ratios at or above 1.5× in the full study.

The unit was the first disagreement

Before any dollar figure, the configurations disagreed on the unit. These unit counts are registered lead-unit results, first reported in the full study.

• On Q08, OpenAI led with an hourly range in all 10 answers, and every one of those ranges was open-ended. Google and Perplexity led with monthly ranges in all 10.
• On Q08, Anthropic led with an hourly range in 6 answers and a monthly range in 4.
• On Q10, Anthropic led with an hourly range in 8 of 10 answers. Its hourly median midpoint across the 6 closed ranges was $325, a descriptive count added after the main results.
• On Q11, Anthropic led with a monthly range in 8 answers and an hourly range in 2.

A buyer asking “How much does a fractional COO charge?” could receive an hourly figure from one assistant and a monthly figure from another, on the same day.

Open-ended ranges

An open-ended range, such as “$10,000+ per month”, has no upper figure, so the fixed rules gave it no midpoint.

• OpenAI’s answers to Q11 ended their first range with “+” in 9 of 10 cases: 7 monthly and 2 hourly.
• Perplexity’s answers to Q09 did so in 8 of 10.
• Perplexity’s answers to Q08 did so in 4 of 10.

A check set before the data pull counted open-ended ranges at their stated figures. Under that check, Q11 becomes computable at 1.52×. This is a sensitivity result, not a finding.

Asking the same question ten times

Repeated calls did not always return the same range.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

After endpoints were rounded to the nearest $10 below $1,000 and the nearest $100 at or above $1,000, Google’s ten closed Q08 ranges had no single most common range.

What Google’s AI Overviews showed

AI Overviews appeared in 7 of the 9 pulls for the three fractional COO questions. The first range and unit were recorded from each, and no AI Overview text is reproduced.

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

These AI Overview ranges are descriptive and are not compared statistically with the API results. The third Q09 pull is not shown, because the extraction audit disagreed on its unit and range. That disagreement is listed in the full study’s Method.

References presented

Perplexity and AI Overviews list reference domains. The study counted which domains were listed and did not read or grade the pages.

On Q08, Q10, and Q11, Perplexity listed fractionus.com in every run, and gofractional.com in every run of Q08 and Q11. It also listed, in every run of all three, linkedin.com, mentorcruise.com, kore1.com, hirechore.com, and fractionalpulse.com. On Q09, it also listed crv.com and founderpath.com in every run.

The author’s domain. The domain kamyarshah.com was listed in all 10 Perplexity runs of each of Q08, Q09, Q10, and Q11, which is 40 runs. It was also listed in AI Overview pulls 1 to 3 for Q08 and pulls 2 and 3 for Q11. This is the author’s own site, as stated in the disclosure at the top. A reference count shows what was listed, not what the page said or whether the answer relied on it.

What the numbers mean (operating judgment)

An AI answer to a fractional COO cost question is a starting point for further research, not a market price or a provider quote. In this data, the unit alone varied by assistant, and the hourly answers were often open-ended.

As operating judgment, buyers should compare scope, days per month, and decision rights rather than a single figure. For example, a hypothetical $25M distributor might compare one proposal at two days a month with another at two days a week. Those are different purchases, and a single monthly number hides the difference.

What this data does not show

• These are not market prices, and they are not what any provider charges.
• The study does not test accuracy.
• The calls went to API configurations, not consumer chat apps.
• The 440 API answers in the full study came from one window, about 17 minutes on October 8, 2026. The AI Overview pulls were made the same day.
• Only four configurations were tested, all at default settings.

Questions about this analysis can be sent through the contact page.

Method

The answers come from the full study, 4 AI Assistants, 11 Executive and Consulting Cost Questions, where the complete method, audit, and exclusion tables appear.

• Collection. From 11:49 to 12:06 UTC on October 8, 2026, each of 11 People Also Ask cost questions went to four API configurations ten times, producing 440 API answers, with provider-default settings and no researcher-added system message. There were also 33 Google AI Overview pulls, three per question, made the same day.
• Extraction. A rule-based parser took the first price range, its unit, and an open-ended flag from each answer. The parser matched the AI-arbitrated coding on all fields in 100 of 100 randomly sampled answers (95% lower bound 0.963). Two auditor models from families not under test did the coding and a third model arbitrated. This was not an independent human audit.

Descriptive counts. Counts labeled as added after the main results are computed from the public normalized-records CSV. A closed median in a non-eligible cell is the median midpoint of that configuration’s closed first ranges in the stated unit for that question. The registered results in this article were first reported in the full study.

Five post-collection changes, each described in full in the study:

• Calibration coder. The 30-answer calibration was coded by the study’s AI producer (Claude, made by Anthropic, whose model is one of the four tested), not a person.
• Parser fix. A parser fix for escaped dollar signs affected 5 of 440 answers and changed no headline figure.
• Kill rule. One preregistered kill rule required auditor agreement (κ) of at least 0.60 on whether an answer contained a price range. The κ statistic could not be computed because both auditors marked every audited answer as containing a range, with raw agreement of 115 of 115. The kill rule therefore was not met as written. After seeing the data, the producer decided that the study would proceed and recorded that decision as a protocol deviation.
• Rounding. A rounding implementation difference changed no figure.
• Title. The study title was shortened to fit the site’s limit.

Data.

• Normalized records (CSV)
• Question selection (CSV)
• Public protocol page

Related analyses:

• Business Consultant Fees: How 4 AI Assistants Answered
• Which Sites AI Tools List for Executive and Consulting Cost Questions
• The Same AI Price Question, Asked 10 Times

Google did not review or endorse this analysis. OpenAI, Google, Anthropic, and Perplexity did not review or endorse it. Model names are used only to identify the configurations tested.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This analysis does not test what those services cost or are worth.

Answers from four API configurations of AI assistants, collected on October 8, 2026 (11:49 to 12:06 UTC), with Google AI Overviews pulled the same day. Each configuration was asked 11 cost questions that Google shows under People Also Ask, ten times each.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This article links to the author’s fractional COO, fractional CMO, and strategy service pages, which may benefit from search traffic it earns. The author’s own domain appears among the references presented by Perplexity and Google’s AI Overviews for some of these questions (counts in the references section). This analysis does not test what those services cost or are worth. The question set, the calls, and the analysis rules were fixed before the main data pull. Five post-collection changes are listed under Method: a parser fix, a calibration check coded by the study’s AI producer instead of a person, one preregistered check that was not met as written, a rounding implementation difference that changed no figure, and the title confirmation.

These are AI-generated statements, not market prices. The study does not test whether any answer is accurate.

The questions and the configurations

The 11 questions are People Also Ask cost questions found on Google results for fractional CMO, fractional COO, and business consultant searches. A fixed selection rule picked them before any assistant was called. The count column shows how many of 124 Google result pages, captured October 7, 2026, listed the question under People Also Ask. The page-level selection file lists all 124 pages and which selected questions each one showed. The rule kept price questions seen on at least two pages, or on one page if they named the COO or CMO role, after the exclusions listed on the protocol page.

Each question went to four API configurations, ten times each, with no researcher-added system message and with the providers’ default settings.

Every call returned the requested model ID, and the returned IDs are in the data package. That made 440 API answers. Add 33 Google AI Overview pulls, three per question, and the total is 473 records, with no record missing or discarded (two AI Overview pulls returned no overview, as reported below). A fifth configuration was dropped before collection, because its provider’s terms do not allow the study to name it.

API configurations are not the consumer chat apps. The apps can add instructions, memory, search, and settings that these calls did not have.

How the answers were compared

From each answer, the parser took the first price range it gave, along with that range’s unit (per hour, per month, or per year) and its midpoint. If a range ended in “+”, such as “$10,000+ per month”, it has no upper figure, so it got no midpoint.

For each question, the modal unit is the unit most answers led with. A configuration was eligible on a question if at least 8 of its 10 answers gave a closed range in that unit. Its median midpoint is the median of those midpoints.

Where three or more configurations were eligible, the cross-configuration ratio is the highest eligible median divided by the lowest.

For the repeated-call analysis only, a cell is any configuration and unit with at least 8 closed ranges among its 10 calls, whether or not that unit is the question’s modal unit.

Results by question

AI-generated statements collected through an API on Oct 8, 2026; not market prices.

The full exclusion counts for every question and configuration are in the appendix.

Six questions had a ratio. Five of the six were at 1.5× or more. The other five questions had too few eligible configurations to compute a ratio, so they carry no headline.

Where the configurations differed most

Fractional CMO per month (Q07). Three configurations were eligible. Their medians were $4,125 (Anthropic), $8,500 (Google), and $12,500 (Perplexity), a ratio of 3.03×. OpenAI’s answers were excluded: 8 of its 10 ended their first range with “+”. The data does not show why.

Fractional CMO, no unit stated (Q04). Medians were $4,500 (Anthropic), $9,000 (Perplexity), and $10,000 (Google), a ratio of 2.22×. OpenAI led with an hourly range in 6 of 10 answers.

Business consultant (Q02 and Q05). All four configurations were eligible on both questions. Anthropic gave the lowest hourly medians, $75 and $100. Google gave the highest, $200 and $225. The ratios were 2.67× and 2.25×. These were the only two questions where all four configurations were eligible.

Fractional CMO per hour (Q01). Medians were $275 to $325, a ratio of 1.18×. This was the only computed ratio below 1.5×.

Fractional COO (Q10). Medians were $8,500 (Google), $10,000 (OpenAI), and $13,000 (Perplexity), a ratio of 1.53×. Anthropic led with an hourly range in 8 of 10 answers.

Asking the same question ten times

Repeated calls with the same question did not always return the same closed range.

In 21 of 27 eligible configuration-unit cells, the closed ranges available from ten calls had a midpoint max/min ratio of at least 1.25, or, where a unique modal normalized range existed, at least 30% of the normalized ranges differed from it. For this test, each range’s endpoints were rounded to the nearest $10 below $1,000 and to the nearest $100 at or above $1,000, so small formatting differences do not create different ranges. The rounding grid changes at $1,000. Under the preset sensitivity check with a single $100 grid, the count is 19 of 27. Two of the 27 cells, Google on Q07 and Google on Q08, had no single most common range. Both count on the max/min test, at 1.33 and 2.50. One of the 27 cells is Perplexity on Q03, which qualified in the monthly unit rather than the question’s modal annual unit.

The last row is the narrowest eligible cell, included for contrast.

The unit changes the answer

The same question did not always get the same unit.

• On Q08 (“How much does a fractional COO charge?”), OpenAI led with an hourly range in all 10 answers, while Google and Perplexity led with monthly ranges.
• On Q10, Anthropic led with an hourly range in 8 of 10 answers.
• On Q03 (“How much does a CMO cost?”), OpenAI, Google, and Anthropic mostly led with annual-unit ranges. Perplexity led with a monthly range in all 10 answers.

Q03 and Q09 do not say “fractional”. Their annual-unit answers are reported separately here and are never treated as fractional pricing.

Open-ended ranges

The share of answers whose first range ended in “+” differed by configuration.

Open-ended ranges have no midpoint under the fixed rules, which is why OpenAI was excluded so often.

A check set before the data pull counted open-ended ranges at their stated figures. With that change, Q06 (2.38×) and Q11 (1.52×) become computable. The six computed ratios above are unchanged. The check also adds OpenAI and Perplexity medians on several questions, for example OpenAI $9,000 on Q07 and Perplexity $275,000 on Q09. Under this check, the count of questions at 1.5× or more would rise from five to seven (Q06 at 2.38× and Q11 at 1.52× become computable). These figures are a sensitivity result, not a finding.

What Google’s AI Overviews showed

AI Overviews appeared in 31 of 33 pulls. The first range and unit from each were recorded. The table shows the six questions with a computed ratio. These are descriptive and are not compared statistically.

References presented

The study recorded and analyzed reference domains only for Perplexity and AI Overviews. The study counted the domains each one listed. It did not read or grade the pages.

Perplexity. On 9 of 11 questions, every reference domain that appeared in any run appeared in all 10 runs of that question. On Q01, salary.com appeared in 7 runs and chatterbuzzmedia.com in 3. On Q08, eca-partners.com appeared in 7 runs. Across all questions, the most frequent domains were mentorcruise.com and fractionus.com, at 90 runs each, followed by linkedin.com at 70.

AI Overviews. Across 31 pulls, the most frequent reference domains were gofractional.com (21), linkedin.com (19), fractionus.com (18), graystoneconsulting.co (11), rankedcmo.com (11), fractionalcxo.to (10), and matthewdeal.com (10).

The author’s domain. The domain kamyarshah.com appeared in 7 of 31 AI Overview pulls (on Q02, Q05, Q08, and Q11) and in all 40 Perplexity runs of Q08 to Q11. This is the author’s own site, as stated in the disclosure at the top. A reference count shows what was listed, not what the page said or whether the answer relied on it.

What the numbers mean (operating judgment)

An AI price answer is a starting point for further research, not a market price or a provider quote. In this data, the answer depended on which configuration was asked, which unit it chose, and which of ten calls it was. A buyer who asks once receives one draw from that configuration’s possible responses. In 21 of the 27 eligible cells, the repeated answers met at least one of the study’s preset variability thresholds.

As operating judgment, buyers should compare scope, expected hours, and responsibilities rather than relying on the title alone. A fractional CMO at two days a month and one at two days a week are different purchases. A single number hides that difference.

How to handle it (operating judgment)

Buyers should define the outcomes and hours before comparing any quotes. One possible operating template is a Balanced Scorecard with two or three selected measures in each of four perspectives: financial, customer, internal process, and learning and growth. A scope worksheet then fixes days per month, decision rights, and reporting.

For example, a hypothetical $30M services firm might compare three fractional CMO proposals. If each proposal is priced against the same scorecard and the same days per month, the firm compares like with like. If it starts from an AI estimate, it compares a range nobody scoped.

Buyers asking an assistant should ask it to state the unit, the scope, and the source behind the figure.

What this data does not tell you

• These are not market prices, and they are not what any provider charges.
• The study does not test accuracy.
• The calls went to API configurations, not consumer chat apps.
• The 440 API answers came from one collection window, about 17 minutes on October 8, 2026. The 33 Google AI Overview pulls were made the same day.
• Only four configurations were tested, all at default settings.
• A reference count does not show what the cited page says.

Questions about this analysis can be sent through the contact page.

Method

Selection and collection. A fixed rule selected the 11 questions from People Also Ask data retrieved through a licensed search-data provider. The question manifest, call schedule, and code were archived before the main data pull. All 440 calls were shuffled into one sequence with a fixed random seed and made through an API router, pinned to each model’s own provider, with fallbacks disabled.

Extraction and audit. A rule-based parser extracted the first range, unit, currency, and open-ended flag from each answer. Two auditor models from families not under test coded a random sample of 100 answers independently, and a third model arbitrated disagreements. The auditor models are not named, under the site’s rule against naming tools.

• The parser matched the arbitrated audit on all fields in 100 of 100 sampled answers (95% lower bound 0.963).
• Auditor agreement on the unit was κ 0.984.
• The parser matched 45 of 45 answers in a preset check set.
• AI Overview extraction matched the audited coding on 29 of the 31 pulls that returned an AI Overview. The two differences were the third pulls for Q03 and Q09, both on questions without a computed ratio. Neither pull appears in the AI Overview table, and reference domains were checked separately.
• Reference extraction was correct on 20 of 20 checked.

Changes after collection, each dated and archived:

• Calibration coder. The protocol called for a person to code 30 answers as a calibration check. At the author’s instruction, the study’s AI producer (Claude, made by Anthropic, whose model is one of the four tested) coded them instead. The parser matched on 29 of 30. This is not an independent human check.
• Parser fix. Five of 440 answers wrote dollar signs as `\$`, and the parser missed their upper figure. All three audit models had coded these five as ranges. The fix was applied and everything was re-run. No headline figure changed. Eligible cells rose from 26 to 27, and cells with notable within-cell spread rose from 20 to 21.
• One kill rule not met as written. One preregistered kill rule required auditor agreement (κ) of at least 0.60 on whether an answer contained a price range. Across the 100-answer random sample and 15 additional audited diagnostic records, both auditors marked every answer as containing a range, for raw agreement of 115 of 115. Because there was no variation in that classification, κ could not be computed. The kill rule therefore was not met as written. After seeing the data, the producer decided that the study would proceed and recorded that decision as a protocol deviation.
• Rounding implementation. The preset endpoint-rounding rule used half-up rounding, but the analysis code rounded ties to the even value. Fifteen records had tie endpoints. All 27 eligible cells were checked with half-up rounding. No cell classification or reported figure changed, and the count stayed at 21.
• Title. The title was confirmed after results were available. It is a shortened form of the title set before collection, cut to fit the site’s 70-character limit, and makes no new claim.

The public data package is a CSV of all 473 normalized records. Each record has the configuration, question, requested and returned model ID, lead range, unit, midpoint, reference domains, and a hash of the answer text. It does not include raw answer text. Because raw answer text is not public, outside readers can recompute the reported summaries from the normalized records but cannot independently verify the extraction from the public package alone. Exclusion counts are in the appendix below. The question-selection rule and a summary of every amendment are on the protocol page. The protocol is summarized on the public protocol page.

Companion analyses from the same data:

• COO Cost Questions: How 4 AI Assistants Answered
• Business Consultant Fees: How 4 AI Assistants Answered
• Which Sites AI Tools List for Executive and Consulting Cost Questions
• The Same AI Price Question, Asked 10 Times

Appendix: all 27 eligible cells

Each row is one configuration and unit with at least 8 closed ranges among its 10 calls. The last column is the share of closed ranges that differed from the cell’s most common range, after rounding endpoints to the nearest $10 below $1,000 and the nearest $100 at or above. The rounding grid changes at $1,000. A single $100-grid sensitivity is reported in the repeated-calls section.

Appendix: exclusion counts per question

Each cell shows mutually exclusive dispositions out of 10, as valid / other unit / open-ended in the modal unit. An answer that was both in another unit and open-ended is counted under other unit. The configuration-level percentages in the open-ended section count every first range ending in “+”, regardless of unit, so they cannot be obtained by summing the open-ended column here.

“Other unit” includes answers with no stated unit. Perplexity’s 10 “other unit” answers on Q03 were monthly ranges. Nine of them were closed, so they form an eligible within-configuration cell in the monthly unit. That cell counts toward the 27 cells in the repeated-calls section, but not toward any cross-configuration ratio.

Google did not review or endorse this analysis. OpenAI, Google, Anthropic, and Perplexity did not review or endorse it. Model names are used only to identify the configurations tested.

The author sells fractional COO, fractional CMO, post-merger advisory, and strategy services. This analysis does not test what those services cost or are worth.

The author sells fractional COO, CMO and post-merger advisory services. This analysis does not test whether those services change any outcome.

The latest year runs September 2025 to August 2026, the first September 2022 to August 2023. Estimates for fractional CFO, CTO, COO, and CHRO were all higher in the latest year. The ranking measures change, the latest twelve months divided by the first twelve, not size. Among the five, the latest fractional CMO estimate of 38,200 was second only to fractional CFO.

The ranking

No Google Ads estimate was returned for “fractional cro”, so it is not part of the ranking. The table compares relative change across the five terms rather than their absolute search volume.

The fractional CMO line, year by year

Google’s estimate for “fractional cmo” and its close variants was 41,400 in the first year. It rose to 58,000 in the second year, then fell to 51,900 in the third. The latest year came in at 38,200, below both the first year and the peak. Every one of the four other fractional role estimates rose from the first year to the latest.

Google reports these monthly figures in six rounded levels, from 1,900 to 5,400 for this keyword. The levels sit roughly 20 to 27 percent apart, and Google publishes no uncertainty for them. The rounded sums place Google’s latest point estimate below the first, without any measure of uncertainty. So the main fractional CMO result appears as a rank and a ratio, never a percentage.

Sensitivity checks run after the main result

These checks were added after the main result was seen, and they carry that label on purpose. March to August 2026 summed about 5.4 percent more than September 2025 to February 2026. Google’s monthly estimates are rounded, so this small difference is not treated as a rebound. No conclusion about the recent direction of the keyword rests on that six-month comparison.

A second check compared the second year with the latest, and fractional CMO again ranked last. A third check compared the third year with the latest, and fractional CMO ranked fourth of five. Fractional CFO’s third-year sum includes two months above 30,000, in February and April 2025. Those two months lower CFO’s ratio in that check, and no month was excluded or adjusted.

Related terms

Google’s estimate for “chief marketing officer” and close variants fell from 51,800 to 38,900. Its estimate for “marketing agency” and close variants rose from 638,500 to 756,300, a 1.18 ratio. Both comparisons use rounded Google estimates whose close-variant sets may overlap or change over time. They are therefore reported here as context and are never used as a headline result.

The ratio of fractional CMO to chief marketing officer estimates did not rise every year. It moved from 0.80 to 1.05, then to 1.02, and finally to 0.98 in the latest year. Against marketing agency, the ratio moved from 0.065 to 0.088, then to 0.064, and finally to 0.051. The “interim cmo” estimate moved from 1,490 to 1,980, but interim placement is a different model.

The “outsourced cmo” estimate moved from 4,030 to 2,970, and “part time cmo” from 4,050 to 4,820. These nearby phrases give context only, and none of them changes the five-term ranking above.

Google Ads estimated top-of-page bid ranges

The figures below are Google Ads estimated top-of-page bid ranges, which may indicate advertiser competition. They are not direct measures of competition or prices paid, and they describe one retrieval date.

What the numbers mean (operating judgment)

In operating practice, the label typed into a search box is rarely the real bottleneck. The friction sits in matching a leadership problem to a role that can own it. Many organizations treat a falling search label as a market verdict, which wastes attention on vocabulary. The calm reading is narrower, and it starts with a measured diagnosis of what the data covers.

These estimates do not show who searched, whether searchers were buyers, or why the figures moved. Signaling theory explains the trap: a label signals intent, but the signal is noisy and shared. Treat a rising or falling label as evidence about language rather than evidence about the work. Do not panic over one keyword, and diagnose the job before choosing any title or model.

How to handle it (operating judgment)

The Balanced Scorecard separates the marketing leadership job from the title attached to it in practice. Its four perspectives are financial, customer, internal process, and learning and growth, each with measures. Write down two or three measures the role must own in each perspective before any hiring conversation. That discipline builds a shared definition the leadership team can trust and refine over time.

Jobs to be done offers a second check: name the outcome the company hires for. A RACI chart then makes ownership visible, so every process has one accountable leader. Consider a hypothetical $40M distributor whose pipeline stalls because nobody owns pricing and channel decisions. In this illustration, a scorecard would help the team test whether unclear ownership causes the stall.

Consider a hypothetical founder who wants brand work, pipeline work, and team coaching at once. That founder should rank the three outcomes and test each engagement model against the same measures. Full-time hires, fractional engagements, interim placements, and agencies can all be judged on one scorecard. Coherence between the scorecard and the hire gives the organization a consistent basis for later reviews.

Every system that defines outcomes first gives later reviews a fixed baseline rather than a fresh debate. Measured rigor compounds in practice, because each review can refine the same agreed measures over time.

What this data does not tell you

• Searches are not buyers, hires, demand for services, or spend.
• Each figure is Google’s estimate for a keyword together with its close variants, and it is rounded.
• The data covers US Google searches only.
• It says nothing about the outcomes of any engagement.
• Google may revise its estimates after the retrieval date.
• No Google Ads estimate was returned for “fractional cro”.
• The bid ranges are a snapshot from one retrieval date.

Growth figures for single fractional-role keywords have also been published elsewhere, using a different data source. Those figures describe individual terms, while this analysis compares five role keywords side by side. This analysis compares twelve-month changes across five fractional role keywords on the same Google estimates. Its figures are therefore not comparable with that publisher’s numbers, and no novelty claim is made here.

Method

Figures are Google Ads estimates from a licensed search-data provider, retrieved October 6, 2026. Data covers monthly US English searches on Google Search without partner networks, September 2022 to August 2026. Each keyword is reported on its own, and figures are never summed across different keywords. Each twelve-month figure is the exact sum of Google’s rounded monthly values for that keyword.

The keyword list and main registered measures were fixed before the study pull, after feasibility checks ran. Sensitivity measures reported above were added after the main result and carry a post-outcome label. Before the list was fixed, the author saw twelve-month average monthly estimates for twelve terms, listed below. Feasibility data also showed four years of monthly estimates and low and high bids for three terms.

• “fractional cmo”: 2,900
• “fractional chief marketing officer”: 2,900
• “fractional coo”: 1,600
• “fractional cfo”: 6,600
• “fractional cto”: 1,600
• “fractional executive”: 880
• “outsourced cmo”: 260
• “interim cmo”: 170
• “part time cmo”: 390
• “chief marketing officer”: 3,600
• “marketing agency”: 60,500
• “hire a cmo”: 30

The keywords “fractional chro” and “fractional cro” were not seen before the keyword list was fixed. Four-year feasibility data and bid values covered “fractional cmo”, “fractional chief marketing officer”, and “chief marketing officer”. No registered growth ratio or cross-role ranking was computed before the list was fixed for this study. The study protocol is summarized on the public protocol page for anyone checking the method.

Appendix: twelve-month sums used in the ranking

Each figure sums twelve rounded monthly Google Ads estimates for the keyword and close variants. The sensitivity ratios above use these sums, and the six-month check uses the same monthly values. Monthly values are not republished here, so the six-month check cannot be recomputed from this page alone. The ranking ratios in the first table divide the last column by the first column.

Google did not review or endorse this analysis. The figures are Google Ads estimates reported through a licensed search-data provider. They are rounded, cover each keyword together with close variants, and Google may change how variants and history are estimated, so changes over time cannot be attributed to a single literal search phrase.

The author sells fractional COO, CMO and post-merger advisory services. This analysis does not test whether those services change any outcome.

Questions about this analysis can be sent through the contact page.

Friction Appears After the Signature

Most acquirers plan the thesis, model the value, and negotiate the terms with precision. Integration plans look complete on paper. Then the deal closes and the two operations meet for the first time under shared ownership.

Friction appears immediately. Redundant roles surface, workflows collide, and handoffs between the two companies stall because no single person owns the outcome. The acquirer adds hours, headcount, and meetings to solve the confusion. More capacity without ownership only moves the chaos faster.

Tasks duplicate across both teams. Decisions reverse because no one knows who holds final authority. Metric definitions differ between the two companies and reporting becomes a negotiation instead of a diagnostic.

The bottleneck is not effort. The bottleneck is the absence of a structural rule that assigns one owner to every handoff before the integration starts. Firms that succeed at post acquisition integration separate the ownership question from the capacity question.

They name the owner first. Then they decide what resources the owner needs to deliver the combined operation. Strategy breaks when accountability is diffuse.

Ownership Maps Assign One Person Per Handoff

Integration creates dozens of handoffs that did not exist before the deal. Payroll runs on two systems. Customer lists live in separate databases. Vendor contracts carry different terms.

Pricing models diverge. Each difference is a handoff that someone must resolve, standardize, or explicitly leave separate. Without that answer, every handoff becomes a negotiation and the integration stalls.

Theory of constraints applies here. The constraint is not the complexity of the systems. The constraint is the diffused accountability that prevents any single decision from closing. Integration magnifies that break because two organizations bring two sets of assumptions about who decides what.

An ownership map resolves the constraint by naming the decider before the question arises. One person per handoff holds accountability for the outcome and coordinates everyone else. That person answers when the handoff breaks. Clarity protects the team from scrambling and prevents the concealed bottleneck where two people think the other one is responsible.

Staffing Structure Separates Baseline from Surge

Integration creates a temporary spike in operational demand. Two accounting systems must reconcile. Two HR platforms must merge.

Product lines must share a warehouse. That work is real and it requires capacity.

The acquirer makes the mistake of treating surge capacity as permanent. The company hires full-time staff to handle the integration and then carries that overhead after the combined operation stabilizes. Cost never recovers and the team becomes bloated.

Baseline capacity is the payroll the combined business needs permanently to run operations, serve customers, and grow. Surge capacity is the temporary resource that handles the integration work and then leaves. Fixed payroll funds the baseline. Contract resources, fractional operators, and project staff cover the surge.

Firms that run integration well build the baseline team first. They model the combined operation at steady state and staff for that model. Then they layer in the surge resources to handle the integration tasks. Surge resources report to the baseline owners, execute the handoff work, document the new processes, and train the permanent team.

A Fractional COO Holds the Handoffs

A fractional COO works inside the integration as an operator, not as an advisor. The role owns the execution of the combined operation during the surge period. Kamyar Shah holds the handoffs, aligns the departments, and standardizes the processes. The deliverable is a running system, not a recommendation deck.

The first step is to map every handoff between the two companies. Mapping starts with interviews of the department heads, traces the workflows, and identifies where the two operations must connect. Each handoff gets a named owner, a success metric, and a timeline. That map becomes the single source of truth for the integration.

The second step is to standardize the metric set using a balanced scorecard. Both companies arrive with different definitions for revenue recognition, customer count, utilization, and margin. The fractional COO selects one definition per metric, documents the choice, and trains both teams on the standard. Reporting aligns and the combined business becomes measurable.

The third step is to design the shared processes using jobs to be done as the organizing principle. The fractional COO takes the best elements from both companies and builds the process the combined operation will run permanently. Processes cover order to cash, procure to pay, and hire to onboarding. Documentation follows, training completes, and monitoring continues until the process runs without intervention.

Integration Execution Is Not a Document

A post-merger integration consulting engagement produces the plan: the sequencing decision, the owners, and the timeline. The plan is usually correct. The plan does not execute itself.

Someone must hold the handoffs, make the calls when the two companies disagree, and drive the work to completion. That someone is the operator. The operator role is not advisory. The operator role is accountable for the combined operation running.

Integration breaks in execution, not in planning. The acquirer knows what needs to happen. The acquirer does not have the capacity or the structure to make it happen while running the existing business. A fractional COO provides that capacity.

Consider a mid-market brand that acquired a smaller competitor to expand geographic reach. The integration plan was clear. Execution stalled because the leadership team was running the core business and could not dedicate time to the integration.

The acquirer brought in a fractional COO to own the integration. The fractional COO held the handoffs, aligned the operations, and handed a running combined business back to the baseline team.

Surge Ends and Baseline Runs

Integration is not a permanent state. The combined operation reaches a point where the handoffs are standardized, the metrics align, and the processes run without intervention. At that point, the surge capacity should leave. Baseline teams take full ownership and the combined business operates as one entity.

The acquirer makes the mistake of extending the surge resources beyond the integration period. The company keeps the fractional COO on retainer or converts contract staff to full-time employees because the resources are performing well. That decision adds permanent cost without permanent need. The combined operation does not require the surge capacity once the integration completes.

Define the end state before the integration starts. Model the combined operation at steady state and identify the baseline team that will run it permanently. Surge resources know from day one that their engagement is temporary. The fractional COO builds the operation with the goal of handing it off.

Firms that execute integration well measure the handoff completion rate. They track how many handoffs are running without intervention, how many metrics are standardized, and how many processes are documented and trained. When those numbers reach full coverage, the integration is complete. Surge resources exit and the baseline team owns the combined operation.

Owners Decide Before Integration Kicks Off

Bringing in a fractional COO happens before the deal closes or immediately after. Waiting until the integration stalls wastes time and money. Early weeks after close are when the two companies set the patterns that will persist. Patterns that form without clear ownership and structure drift into chaos.

The acquirer should model the combined operation during diligence. Identify the handoffs, estimate the surge demand, and decide who will hold the integration execution. Most mid-market companies do not have a dedicated integration function. For those companies, the choice is between asking the existing management team to absorb the integration or bringing in temporary operational capacity.

Management teams are already running the core business. Adding the integration load creates risk. An interim COO model fits the structure.

The role is temporary, accountable, and operational. Engagement ends when the integration completes.

Match the resource to the demand. Integration is a surge event. Surge events require surge resources.

Permanent staff handle baseline operations. Temporary resources handle temporary spikes. That separation protects the business from cost bloat and keeps the integration pointed at the value the thesis promised.

Combined Operations Run as One System

Post acquisition integration aims to deliver a combined operation that runs as one system. Handoffs are standardized. Metrics align.

Processes are documented and the team knows who owns what. The combined business serves customers, reports one set of numbers, and grows without the friction that marked the first weeks after close.

Measure success by operational coherence. The combined business operates with the same discipline the acquirer brought to the core business before the deal. Reporting is consistent.

Decisions are clear. Accountability is assigned. Leadership teams can diagnose performance using VRIO and make adjustments without debating definitions or chasing down who is responsible.

Every successful integration follows the same pattern. The owner names the deciders before the work starts. The owner separates baseline capacity from surge capacity.

That owner brings in temporary operational resources to handle the integration and then removes those resources when the integration completes. Baseline teams take ownership of the combined business and run it permanently.

Firms that execute integration well protect their teams, serve their customers, and build systems that compound over time. Integration is not a distraction. Integration is the operational work that turns the thesis into a running business.

Discipline required to execute that work is the same discipline required to run any mid-market operation. Integration compresses the timeline and magnifies the consequences of diffused accountability. Ownership maps and surge capacity structures prevent those consequences and deliver the combined operation the acquirer intended to build.

Kamyar Shah sells fractional COO and CMO services.

Two counts of the same year

Two Census programs offer different measures related to job endings and employment contraction at young firms.

The Quarterly Workforce Indicators (QWI) are built from state unemployment insurance wage records. They count jobs, quarter by quarter, and record a separation whenever a job present in one quarter is absent in the next. Census defines it exactly: “A separation is indicated when a job is present in one quarter, but is not present in the following quarter.” It also warns what the count cannot do: “Voluntary separations (retirement, new job) and involuntary separations (layoff, firing) cannot be separately identified.”

The Business Dynamics Statistics (BDS) are the Census program used here for job destruction and jobs lost at closing establishments. They count jobs lost from contracting and closing establishments during the last 12 months. Census defines the measure as “Job destruction is the sum of all employment losses from contracting establishments from year t–1 to year t including establishments shutting down.” The data service labels the subset from closures as “Number of jobs lost from closing establishments during the last 12 months”.

Both programs classify by firm age. In the BDS, “The age of the parent firm is based on the age of the oldest establishment in the firm, in the year the firm came into existence, and does not change in subsequent years as a result of establishment entry or exit.” In the QWI, “Firm age and size are defined at the national level, rather than the state level.” This study takes the youngest group each program offers that lines up with the other: QWI firm age 0 to 1 years, and BDS ages 0 and 1 combined.

The registered side-by-side

All industries, private employers, QWI quarters 1 to 4 of 2023 and BDS data year 2023, covering 48 states and the District of Columbia (every jurisdiction for which both programs returned usable figures, since Alaska and Michigan returned no QWI data).

Sources: U.S. Census Bureau, Quarterly Workforce Indicators, state files, 2023 Q1 to Q4, private ownership, all industries, not seasonally adjusted, firm age groups 1 and 5. Business Dynamics Statistics, state by firm age, 2023, age codes 0 years, 1 year and 11+ years. Both retrieved from the Census data service on October 4, 2026 and archived before the protocol was frozen.

The QWI rate is the year’s separations (5,212,670 at firms aged 0 to 1) divided by the average over the four quarters of jobs held at the start of a quarter (4,027,491). It can exceed 100 without any error in the data, because the numerator is an annual count of separations and the denominator is the average beginning-of-quarter job count. The BDS rate is jobs destroyed (611,207) divided by the Census denominator for the age group (3,625,836), which is the average of employment in the current and prior year. Of those destroyed jobs, 345,091 were at establishments that closed.

Nothing is subtracted. The two rates count different events (a worker-job relationship ending, as against gross year-to-year employment loss at contracting and closing establishments), on different denominators (average beginning-of-quarter employment, as against average current- and prior-year employment), with different timing (the quarter a job ended, as against a year-to-year comparison), and with firm age set on different program schedules (QWI reports its 0 to 1 group for each quarter, while BDS determines firm age annually from the year the firm came into existence, using the age of its oldest establishment), so matching the calendar year, the age labels and the jurisdictions does not make the two observations one cohort. There is also a known asymmetry between them: QWI separations are built from unemployment insurance wage records and can be unobserved for a firm’s final quarter, so observed QWI separations at exiting firms may be understated, while BDS job destruction from deaths is designed to include employment losses from closing establishments. That affects how the two columns should be read beside each other. It does not make them parts of one total, and no remainder between them is reported here.

By sector

The same three measures by sector appear below for firms aged 0 to 1 and, as the prespecified contrast, for firms aged 11 years or more. Each row uses the jurisdictions where QWI returned usable figures for that sector in every quarter and BDS returned the matching annual cells. Sectors with fewer than 40 jurisdictions are marked as limited coverage.

Firms aged 11 years or more, by sector, on the same construction. Where a sector’s jurisdiction count differs between the two age tables (mining 32 against 48, utilities 27 against 49, management of companies 42 against 49), the two age rows rest on different jurisdiction sets and are not comparable with each other across age. The all-industries rows and every sector at the same count in both tables are on the same set.

The sector columns do not move together. Information had one of the lower QWI separation rates, at 96.0 separations per 100 beginning-of-quarter jobs, and the second-highest BDS job destruction rate, at 23.6 jobs destroyed per 100 jobs on the Census annual denominator, behind only management of companies. Finance and insurance had the second-highest QWI separation rate, at 244.0 separations per 100 beginning-of-quarter jobs, and the sixth-highest BDS job destruction rate, at 20.7 jobs destroyed per 100 jobs on the Census annual denominator. A reader looking for the sectors where young firms “churn the most” will get a different list from each column, which is the point of showing both.

What the numbers mean for an operator

These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.

For an operator, the two 49-jurisdiction aggregate figures identify different things to monitor. QWI records worker-job separations of every kind, which it cannot sort into quits, layoffs, firings or exit-related losses. BDS records gross year-to-year employment losses at establishments that contracted or closed. In the BDS aggregate, job destruction equaled about 17 per 100 jobs on the Census annual denominator, and more than half of that was at establishments that closed. As operating judgment, the first figure points at hiring and retention, which an owner can measure monthly, and the second at survival and sizing, which an owner can also track internally. The data cannot say how much of the first is connected to the second, and this article does not estimate it.

How to handle it: keep two ledgers

The steps below are an operating recommendation. This study did not test them, and the Census Bureau does not endorse them.

1. Track worker-job endings and establishment employment reductions separately. Record worker-job endings throughout the year and, separately, compare establishment employment over a consistent annual period. The first is an internal counterpart to QWI separations, and the second is an internal counterpart to BDS job destruction. Neither internal count is the federal measure, and a company that tracks only period-end headcount sees neither clearly.

2. Use 129.4 as aggregate context, not a company forecast. Across firms aged 0 to 1 in the 49-jurisdiction frame, annual QWI separations equaled 129.4 per 100 average beginning-of-quarter jobs. The aggregate does not show what a particular company, or a firm in its first year alone, should expect. As operating judgment, a budget that assumes a stable team in the first two years is assuming something the aggregate does not show for young firms in general.

3. Read the sector column that matches the decision. For a hiring plan, the separations column. For a question about whether the business model is holding, the job destruction column. The two rank sectors differently, so the right comparison depends on the question.

4. Measure the first ninety days of each job. The QWI count includes jobs that started and ended inside a single quarter. Whether a company’s separations are concentrated there is something its own records can show and the public data cannot. Employment decisions, including terminations and reductions in force, carry legal obligations that vary by jurisdiction, so an owner should take them with qualified legal and HR advice.

What this data does not tell you

• It does not separate quits from layoffs or from jobs lost when firms close (QWI), and it does not say why an establishment contracted or closed (BDS).
• It does not show what share of separations came from job destruction. The two measures cannot be subtracted, and no such figure is reported here.
• Both are national firm-age classifications applied to state records, and the sector tables cover different jurisdiction sets.
• QWI values are protected with noise and carry status flags. Only values flagged 1 (OK) or 6 (calculated from other released measures, no significant distortion) were used, and Census publishes no sampling error for either program.
• QWI separations can be unobserved in a firm’s final quarter, so separations at exiting firms may be undercounted.

Method

Already published by the agency: Census publishes QWI separations and employment by firm age and state every quarter, and BDS job destruction, job destruction from deaths and the denominator by firm age, state and sector every year. New here: the two programs built on matched jurisdiction frames for one year (one 49-jurisdiction frame for all industries, a complete-case frame for each sector) and reported beside each other at firms aged 0 to 1 and 11 or more.

Sources. U.S. Census Bureau, Quarterly Workforce Indicators via the Census data service (qwi/sa), 2023 quarters 1 to 4, private ownership (A05), all industries and each two-digit sector, firm age groups 1 (“0-1 Years”) and 5 (“11+ Years”), not seasonally adjusted, indicators Emp (“Beginning-of-Quarter Employment: Counts”) and Sep (“Separations: Counts”) with their status flags. U.S. Census Bureau, Business Dynamics Statistics via the Census data service, 2023, state by firm age and state by sector by firm age, age codes 010 (“0 years”), 020 (“1 year”) and 075 (“11+ years”), variables JOB_DESTRUCTION, JOB_DESTRUCTION_DEATHS and DENOM. All responses retrieved October 4, 2026 and archived before the protocol was frozen.

Frames. A jurisdiction is in a frame when its four QWI cells for that age group and industry carry status 1 or 6 on both Emp and Sep, and its BDS row(s) carry the three variables. The all-industries frame is 49 jurisdictions: every state except Alaska and Michigan, plus the District of Columbia. Each sector has its own frame, with counts in the table. Sectors at 49 use that same set. For firms aged 0 to 1, agriculture lacks the District of Columbia (48). Mining (32) lacks Connecticut, Delaware, the District of Columbia, Hawaii, Iowa, Maine, Maryland, Massachusetts, Minnesota, Nebraska, New Hampshire, New Jersey, North Carolina, Oregon, Rhode Island, South Dakota and Vermont. Utilities (27) lacks Alabama, Arkansas, Delaware, the District of Columbia, Hawaii, Indiana, Iowa, Kansas, Louisiana, Maine, Mississippi, Nevada, New Hampshire, New Mexico, North Dakota, Rhode Island, South Dakota, Vermont, Washington, West Virginia, Wisconsin and Wyoming. Management of companies (42) lacks Delaware, Hawaii, New Mexico, Rhode Island, South Dakota, Vermont and Wyoming. For firms aged 11 or more, agriculture and mining lack the District of Columbia (48) and every other sector is at 49.

Rates. QWI: the sum of separations over four quarters and frame jurisdictions, divided by the mean over the four quarters of beginning-of-quarter employment summed over the frame, times 100. BDS: job destruction summed over frame jurisdictions and the age codes, divided by the summed denominator, times 100, and the same for job destruction from closures.

Protocol. The age mapping, frame rule, measures and the rule that no difference or ratio between the programs is computed were frozen in a dated protocol on October 4, 2026, after the source responses had been archived with hashes and before the registered quantities were computed. Before freezing, the QWI status flags and the Census API’s structure had been inspected, and national BDS rows by firm age, including 2023 national job destruction rates by firm age, had been seen, but the 49-jurisdiction rates reported here had not been computed. No QWI value had been read. The protocol (wave 3 section) and its dated amendments are at https://kamyarshah.com/public-data-protocol/

Check. A second program, written separately from the registered definitions by a different author from the analysis code and run against the raw files without access to this study’s code or results, recomputed all three claim groups, covering every frame list, count and rate in the tables. Every figure in the tables and text was compared and matched. The programs, archived inputs, manifests and hashes are kept with the protocol record.

The agencies did not review or endorse this analysis. Kamyar Shah sells fractional COO and CMO services.

Questions about this analysis can be sent through https://kamyarshah.com/contact/.

Kamyar Shah sells fractional COO and CMO services.

The ratio the IRS table does not print

For tax year 2023, the Internal Revenue Service’s Statistics of Income division published an income statement for nonfarm sole proprietorships, by industrial sector, built from a sample of Schedule C returns. The table carries a line for “Advertising expenses” and a line for “Business receipts, total [1]”. Dividing one by the other gives the share of the sector’s receipts that went to advertising. That share is the number this article is about. It is simple arithmetic on two published lines, and the IRS does not publish it.

For tax year 2023, across all nonfarm sectors, advertising expenses were 1.18% of aggregate estimated business receipts on sample-weighted sole proprietorship returns, a receipts-weighted ratio in which the largest filers carry most of the weight: about 1.2 cents per aggregate receipts dollar. The table note says what the figures are: “[All figures are estimates based on samples–money amounts are in thousands of dollars]”.

Advertising by sector, and three other bought-in costs

Source: IRS, Statistics of Income Division, Sole Proprietorship Returns 2023, Table 2 “Nonfarm Sole Proprietorships: Income Statements, by Industrial Sectors, Tax Year 2023”, released March 2026, downloaded October 4, 2026. Shares are the sector’s aggregate deduction divided by the sector’s aggregate receipts. The all-industries row includes 521,885 unclassified returns that are not assigned to one of the 18 matched sector rows. Receipts are rounded to the nearest million.

Three other patterns in that table stand beside the advertising column. First, in every sector but retail trade, sole proprietors reported more contract labor than advertising, and in construction the gap is more than twenty to one: 11.89 cents of each receipts dollar to contract labor, 0.51 to advertising. Second, the two sectors with the heaviest wage bills, accommodation and food services (15.73%) and health care (13.13%), sit in the middle of the advertising ranking. Third, legal and professional services ran between 0.50 and 2.57 cents on the dollar, and in six sectors (agriculture, mining, utilities, transportation and warehousing, professional, scientific and technical services, and health care and social assistance) it exceeded the advertising share.

The registered comparison: a near-zero rank correlation

The study fixed one question in the frozen protocol before the IRS advertising values were read and before the registered comparison was computed: across sectors, does the 2023 advertising share rank the way employer-firm job growth ranks? Growth is the Census Bureau’s net job creation rate for 2021 to 2023, from the Business Dynamics Statistics, recomputed as 100 times net job creation summed over those years divided by the Census DHS denominator summed over the same years, for the same 18 sectors. Because both numerator and denominator are sums of three annual values, the result is the denominator-weighted mean of the three annual rates, a rate per year, not a three-year cumulative change. One IRS column, administrative and support and waste services, includes management of companies among its subsectors, so its Census counterpart is sectors 55 and 56 combined. That correction was made after the first result was seen and is recorded as Amendment 1 to the protocol: under the earlier, wrong mapping the coefficient was minus 0.10. The year 2020 was left out, as registered, as a pandemic outlier.

With 2020 excluded as registered, the Spearman rank correlation across 18 sectors between the 2023 advertising share on sole proprietorship tax returns and the 2021 to 2023 net job creation rate at employer firms was minus 0.09, and because the growth window largely precedes the tax year in which advertising is measured, the correlation carries no temporal direction. That is a near-zero descriptive rank correlation: in these matched sector rows, the sectors where sole proprietors put the largest share of receipts into advertising were not, as a group, the sectors where employer firms added jobs faster. No inferential test was run, so this is not a statistical null, and it is not evidence that advertising has no effect. Information had the highest advertising share and a job growth rate of 0.41% a year, the fourth-lowest. Transportation and warehousing had the second-lowest advertising share and the highest job growth, 4.16% a year. Construction, with the third-lowest advertising share, grew at 2.05%.

The advertising measure covers sole proprietorship tax returns, while the growth measure covers jobs at employer firms. The growth window, 2021 to 2023, largely precedes the tax year in which advertising is measured, so the correlation carries no temporal direction, and no causal reading is available either way. The result is reported because the comparison was registered. Two values are reported: minus 0.10 under the crosswalk as originally registered, and minus 0.09 under the crosswalk corrected by Amendment 1 after the first result was seen. The correction rests on a fact visible in the IRS table’s own column headers (the administrative column lists management of companies as a subsector), not on the coefficient, which it barely moved.

Source for the growth column: U.S. Census Bureau, Business Dynamics Statistics, sector by year, variables NET_JOB_CREATION and DENOM, 2021 to 2023, retrieved October 4, 2026.

What the share is, and is not

The advertising share is advertising dollars per dollar of aggregate estimated receipts on sample-weighted returns. It is receipts-weighted, which means the largest sole proprietorships in a sector carry most of the weight. It is not the mean or median return’s share, not a benchmark for a business of any particular size, and the table does not classify returns by whether the business has employees. IRS estimates are based on a sample and may differ from a full count. No coefficient of variation table accompanies this income statement on the IRS site, so no interval is shown and the shares should be read as sector-level descriptions, not precise constants.

Two descriptive companions. The first was registered as the advertising share by sector and subsector, and its extension to the detail tier is recorded in Amendment 1, made after results were seen. The second was registered as written. First, Appendix A, available as a downloadable CSV at Appendix A (CSV), reports the advertising share for every labelled column of the IRS table, 153 rows at the sector, subsector and detail tiers (two detail rows, formal wear and costume rental and video tape and disc rental, are suppressed in the IRS file). Second, computed from the published return counts in the IRS’s Table 1 rather than printed by the IRS, the share of returns with net income:

What the numbers mean for an owner deciding what to spend

These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.

The useful fact in this table is not the 1.18 cents. It is that aggregate contract labor exceeded aggregate advertising in 17 of the 18 sectors, that the all-industries legal and professional services share was 1.01% against 1.18% for advertising, and that the sectors buying the most advertising per receipts dollar did not rank higher on employer-firm job growth in the registered comparison. An owner setting a marketing budget by asking “what does my sector spend” is asking a question this table can answer only in aggregate, and this study found no rank alignment between that aggregate and employer-firm job growth.

How to handle it: budget from the funnel, not from the sector

The steps below are an operating recommendation. This study did not test them, and the Internal Revenue Service and the Census Bureau do not endorse them.

1. Start from cost per acquired customer, not from a percentage of receipts. A share of receipts is a reporting category. A budget is the number of customers the business needs, times what it has cost to acquire one, measured on the business’s own last twelve months.

2. Put contract labor and advertising on the same page. In most sectors in this table, aggregate contract labor exceeds aggregate advertising. If outside labor is doing work that wins customers (referral partners, subcontracted sales, agency retainers booked as contract labor), the real marketing spend is larger than the advertising line shows, and the budget review should count it.

3. Treat a sector average as a prior, not a target. The information sector’s 3.26 cents and construction’s 0.51 cents describe what those businesses reported in aggregate. Neither number says what a given firm should spend, and the near-zero rank correlation above is a reason not to copy a heavy-advertising sector on the assumption that it grows faster.

4. Review the deduction categories with an accountant before comparing across years. Where a cost is booked (advertising, contract labor, commissions, other expenses) changes the shares. Whether a particular outlay is deductible, and in which category, depends on the taxpayer’s facts and applicable law, so a qualified tax professional should review the classification.

What this data does not tell you

• It does not say what a typical sole proprietor spends on advertising. The shares are aggregate and receipts-weighted.
• The IRS advertising table does not cover corporations or partnerships, and it does not say whether a sole proprietorship return has employees.
• It does not connect advertising to growth: the growth measure is a different population (employer firms) over an earlier window, and the registered comparison found no alignment.
• IRS figures are sample estimates without a published coefficient-of-variation table for this income statement.
• Deduction categories reflect how taxpayers classified costs, not a marketing accounting standard.

Method

Already published by the agency: the IRS publishes advertising expenses and business receipts by sector as dollar totals (Table 2) and return counts with and without net income by sector (Table 1). Census publishes net job creation by sector in the Business Dynamics Statistics. New here: the advertising, contract labor, legal and professional, and wage shares of receipts by sector, the advertising share for every subsector and detail column (Appendix A), the share of returns with net income computed from the published counts, and the registered rank comparison with Census job growth.

Sources. IRS, Statistics of Income, Nonfarm Sole Proprietorship Returns, Tax Year 2023: Table 2 (file 23sp02is.xls) and Table 1 (file 23sp01br.xls), both released March 2026 and downloaded October 4, 2026. U.S. Census Bureau, Business Dynamics Statistics, sector by year, retrieved October 4, 2026 and archived before the protocol was frozen.

Crosswalk. Sixteen of the IRS’s 18 sector columns correspond directly to one NAICS sector each. The IRS nonfarm agriculture column is paired with Census sector 11, but the coverage is not identical: the IRS table is for nonfarm returns and excludes farm proprietorship activity, while Census sector 11 covers the full employer-firm agriculture sector. That row stays in the comparison as registered, with this difference disclosed. The administrative and support and waste services column includes management of companies among its subsectors, so it corresponds to NAICS 55 and 56 together, and its Census rate is computed from the two sectors’ summed counts (Amendment 1, made after the first result was seen). The IRS unclassified establishments column has no counterpart and is excluded from the comparison. For each matched row, the Census rate is 100 times net job creation summed over 2021 to 2023, divided by the Census denominator summed over the same years.

Comparison. Spearman rank correlation across the 18 matched sectors between the 2023 advertising share and the 2021 to 2023 net job creation rate: minus 0.09 (minus 0.10 under the pre-amendment crosswalk). Coefficient and count only, with no probability value and no causal reading.

Protocol. The sources, share formulas, growth window, crosswalk and comparison were frozen in a dated protocol on October 4, 2026, after the source files and responses had been archived with hashes and before the registered comparison was computed. Before freezing, the IRS file’s labels and the Census API’s structure had been inspected, and some Census rows had been seen (national rows by firm age, the all-sector net job creation row, and the agriculture and mining sector rows). No IRS value had been read. The crosswalk was corrected after the first result was seen (Amendment 1, above). The protocol (wave 3 section) and its dated amendments are at https://kamyarshah.com/public-data-protocol/

Check. A second program, written separately from the registered definitions by a different author from the analysis code and run against the raw files without access to this study’s code or results, recomputed the 9 claim groups of the original analysis and, after Amendment 1, all 12 amended claim groups, including every row of Appendix A. Every figure in the tables and text was compared and matched. The programs, archived inputs, manifests and hashes are kept with the protocol record.

The agencies did not review or endorse this analysis. Kamyar Shah sells fractional COO and CMO services.

Questions about this analysis can be sent through https://kamyarshah.com/contact/.

Kamyar Shah sells fractional COO and CMO services.

A person buying a business with a government-guaranteed loan is making two bets at once: on the business, and on the sector it sits in. The Small Business Administration’s loan records show recorded charge-off dates for loans coded as a change of ownership, which this study tests against the registered five-year window. The Census Bureau’s Business Dynamics Statistics show, for the same sectors, how often employer firms exited. Neither agency publishes the two beside each other. This study does, for 13 sectors, and then reports one registered comparison: whether the sector ranking of loan charge-offs matched the sector ranking of firm deaths.

The sector table

Loans are non-cancelled SBA 7(a) records with an approval date from October 1, 2017 through September 30, 2020 (fiscal 2018 to 2020) and a business-age code of “Change of Ownership”. A loan counts as charged off within five years when SBA’s charge-off date falls on or before the fifth anniversary of approval. The firm death rate is the Census count of employer firms that exited, divided by the count of employer firms, averaged over calendar 2019 to 2023.

Sources: U.S. Small Business Administration, 7(a) FOIA loan-level files as of June 30, 2026 (fiscal 2010 to 2019 and fiscal 2020 to present). U.S. Census Bureau, Business Dynamics Statistics, sector by year, retrieved October 4, 2026. The six sectors with fewer than 200 loans are shown for completeness and were excluded from the registered comparison, as fixed in advance. The sold-on-the-secondary-market column matters because the public file may not record later charge-off performance for every sold loan, so a sector’s charge-off rate may be understated, and more so where more loans were sold. No loan had a blank, non-numeric or unresolvable NAICS code, and no sector in the loan file fell outside the Census sector list.

The Census counts behind the last column, by year (firms, then firm deaths, from the Business Dynamics Statistics, with the mean annual death rate calculated as the mean of the five yearly ratios):

Across all 19 sectors the cohort was 14,517 loans and 295 charge-offs, a five-year rate of 2.03%. The loan unit is a loan, not a business, and SBA’s borrowers include firms with no employees. The Census unit is an employer firm. A charge-off is SBA’s accounting event, defined in its data dictionary as “Date SBA charged off loan (if applicable)” for the date, and a firm death is Census’s “Number of firms that exited during the last 12 months”. The two populations overlap without being the same population, and the article never divides one by the other.

The registered comparison

The study fixed one comparison before any loan was counted: a rank correlation across eligible sectors between the five-year charge-off rate and the mean annual firm death rate. Eligible meant at least 200 change-of-ownership loans and complete Census figures for every year from 2019 to 2023. Thirteen sectors qualified.

The descriptive, ecological Spearman rank correlation was 0.68 across those 13 sectors, comparing five-year charge-off rates for SBA change-of-ownership loans, whose borrowers can include firms with no employees, with Census death rates for employer firms. In words: the sectors whose acquisition loans charged off most often were, broadly, the sectors where employer firms exited most often. Transportation and warehousing had the second-highest charge-off rate (3.93%) and the highest firm death rate (10.76%) among eligible sectors. Finance and insurance had the lowest charge-off rate (1.03%), while manufacturing had the lowest firm death rate (6.25%). Health care and social assistance had the second-lowest charge-off rate (1.21%) and the third-lowest firm death rate (6.93%). Arts, entertainment and recreation had the highest charge-off rate of all (4.38%) and the third-highest firm death rate (9.08%) among eligible sectors.

This is a descriptive, ecological result. Thirteen points support a ranking, not an inference, and no probability value is attached. The loans’ five-year windows run from late 2017 to late 2025, while the Census years are 2019 to 2023, so the two measures overlap in time without matching exposure. Nothing here says whether a charge-off happened because of the sector, the deal, the buyer, the lender or the year.

What is in the loan count, and what was set aside

The cohort started from 15,938 change-of-ownership loans approved in the window, after applying the registered rule that BusinessAge had to be exactly “Change of Ownership”. In the same approval window, 1,220 records had a blank BusinessAge value and 21,889 were coded “Unanswered”, so they were outside the cohort by construction. Cancelled loans (1,421) were removed. No record in any sector was inconsistent under the registered rules (a charge-off date before approval, an invalid date, or a charged-off status with no date), so the inconsistent count was 0 and the inconsistent share was 0.0%. No loan was excluded because its fifth anniversary fell after the June 30, 2026 as-of date: the count was 0. One undisbursed loan with status “COMMIT” stayed in the cohort under the registered definition.

A further 262 loans carry a charge-off date after their fifth anniversary. Under the registered endpoint they count as not charged off within five years, because they had not been. Over the whole fiscal 2018 to 2020 cohort, then, 557 loans show a charge-off status in the June 30, 2026 file, of which 295 fall inside the five-year window.

Of the 14,517 loans, 8,174 (56.3%) carry SBA’s indicator for a loan sold on the secondary market, which the dictionary describes as “An indicator if the loan was sold on the secondary market.” The public file may not record later charge-off performance for every sold loan. If unrecorded post-sale charge-offs occurred, the charge-off rates here may be understated, especially in sectors where more loans were sold. The sold count for every sector is in the sector table. Among the 13 eligible sectors, the share sold ranged from 45% in wholesale trade (209 of 462) to 66% in accommodation and food services (1,924 of 2,904).

SBA’s status “EXEMPT”, carried by 5,794 of the loans, means in the dictionary’s words “The status of loans that have been disbursed but have not been cancelled, paid in full, or charged off are exempt from disclosure under FOIA Exemption 4”. Those loans count as not charged off, because they have no charge-off date and no charge-off status.

The five-year rate here is the proportion observed in the complete eligible extract cohort, not an estimate from a probability sample. No sampling interval is shown, and the wave 1 study’s intervals for the same file are not repeated. Variation across loans that might be made in the future is not estimated.

What the numbers mean for a buyer or a seller

These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.

A charge-off rate of 2.03% over five years, with sector rates from about 1% to about 4% among the large groups, is the backdrop against which a buyer is negotiating. The sector ranking says something a buyer can use without over-reading it: broadly, in sectors where employer firms exited more often, the loans that financed a purchase also charged off more often. That is consistent with sector conditions showing up in both records, and consistent with several other explanations the data cannot separate.

For a seller, the same table supplies historical sector-level context that a buyer may consider when pricing financing risk. It does not measure any particular buyer’s risk.

How to use it: three checks before the letter of intent

The steps below are an operating recommendation. This study did not test them, and the Small Business Administration and the Census Bureau do not endorse them.

1. Price the sector, then the business. Before valuing the specific company, a buyer can read the sector’s employer-firm death rate as sector context, and treat a target whose own history appears better than that context as a claim to be verified in diligence, not accepted from the listing.

2. Do the underwriting as if no one else will. More than half of these loans were sold on the secondary market, which is normal practice and says nothing in this data about any loan’s risk. Whoever ends up holding the loan, this data does not establish who bears the legal, financial or operating consequences of a charge-off. A buyer should review those consequences with qualified legal and lending advisers, and the buyer’s own underwriting of the business has to stand on its own.

3. Separate the deal risk from the sector risk in the plan. The sectors with high charge-off rates in this table also have high employer-firm death rates, and those exits are measured across employer firms generally, not only bought firms. A buyer’s first-year operating plan in such a sector should stress-test customer, staffing and supplier assumptions, rather than treating the purchase as inheriting a stable position. Decisions about loan structure, personal guarantees and tax treatment depend on the buyer’s circumstances and applicable law, and should be reviewed with qualified legal, lending and tax advisers.

What this data does not tell you

• It does not say why any loan charged off, or that the sector caused it.
• A charge-off is a loan outcome, not the failure of the purchased business, and a firm death is an exit from the Census frame, not necessarily a failure.
• The loan unit includes businesses with no employees, while the Census unit is employer firms only.
• The five-year clock starts at approval, and loans charged off later than that are not counted here.
• Charge-off performance may be unobserved for loans sold on the secondary market.
• Thirteen sectors support a ranking, not a statistical test, and no interval or probability value is claimed.

Method

Already published by the agency: SBA publishes industry-level charge-off rates for all 7(a) loans in its portfolio performance reports, without a split by the business-age code. Census publishes firm counts and firm death counts by sector in the Business Dynamics Statistics. The wave 1 study on this site reported the all-sector five-year charge-off rate for change-of-ownership loans against existing-business loans. New here: the change-of-ownership cohort’s five-year charge-off rate by sector, set beside the Census firm death rate, and the registered rank comparison.

Sources. U.S. Small Business Administration, 7(a) FOIA loan-level files, FOIA_7a_FY2010_FY2019_asof_260630.csv and FOIA_7a_FY2020_Present_asof_260630.csv, data as of June 30, 2026. The files are public records. In SBA’s data dictionary, BusinessAge is a “Categorical description of the age of the business” and NaicsCode is the “North American Industry Classification System (NAICS) code”. U.S. Census Bureau, Business Dynamics Statistics, sector by year, variables FIRM (“Number of firms”) and FIRMDEATH_FIRMS (“Number of firms that exited during the last 12 months”), retrieved from the Census data service on October 4, 2026 and archived before the protocol was frozen.

Cohort and outcome. Approval dates were from October 1, 2017 through September 30, 2020, BusinessAge was exactly “Change of Ownership”, and cancelled loans were excluded. The fifth anniversary is the same month and day five years after approval (a February 29 approval maps to February 28). A loan is charged off within five years when its charge-off date is valid, not before approval, and not after that anniversary. Sector is the first two digits of the NAICS code, with 31 to 33, 44 to 45 and 48 to 49 combined as NAICS does. No borrower, lender or franchise brand is named.

Comparison. Spearman rank correlation across sectors with at least 200 eligible loans and complete Census figures for 2019 to 2023, between the five-year charge-off rate and the mean of the five annual firm death rates. No quotient is formed between loan events and firm events.

Protocol. The cohort, outcome, exclusions, eligibility threshold and comparison were frozen in a dated protocol on October 4, 2026, after the source files and responses had been archived with hashes and before the registered comparison was computed. Before freezing, the SBA file’s field list and the Census API’s structure had been inspected, and some national Census rows had been seen. No loan had been counted. The protocol (wave 3 section) and its dated amendments are at https://kamyarshah.com/public-data-protocol/

Check. A second program, written separately from the registered definitions by a different author from the analysis code and run against the raw files without access to this study’s code or results, recomputed all 13 claim groups, including every row of the sector tables. Every figure in the tables and text was compared and matched. The programs, archived inputs, manifests and hashes are kept with the protocol record.

The agencies did not review or endorse this analysis. Kamyar Shah sells fractional COO and CMO services.

Questions about this analysis can be sent through https://kamyarshah.com/contact/.

The Bureau of Labor Statistics measures nonfatal workplace injuries and illnesses through the Survey of Occupational Injuries and Illnesses (SOII). BLS describes its scope this way: “The SOII excludes all work-related fatalities as well as nonfatal work injuries and illnesses to the self-employed; to workers on farms with 10 or fewer employees; to private household workers; to volunteers; and to federal government workers.” Each year, according to BLS, “approximately 200,000 employers report for establishments in private industry and the public sector (state and local government).” This article uses private industry only.

BLS describes SOII as “an establishment-based survey used to estimate incidence rates and counts of workplace injuries and illnesses,” and places each establishment in a size class that is “based on an establishment’s average annual employment.” The unit is therefore the establishment, generally a single workplace, not a company. Rates are cases per 100 full-time equivalent workers. BLS computes them as the number of cases times 200,000, divided by hours worked, and explains: “(The 200,000 hours in the formula represents the equivalent of 100 employees working 40 hours per week, 50 weeks per year, and provides the standard base for the incidence rates.)”

The national pattern, which BLS has already published

Total recordable case rates, private industry, by establishment size (cases per 100 full-time equivalent workers):

Source: U.S. Bureau of Labor Statistics, SOII, quartile table 1 (total recordable cases) for each year, private industry, downloaded October 3, 2026. The 2020 and 2021 rates may reflect pandemic-period changes in employment and reporting.

In every year shown, the 50-to-249 class had the highest rate and the 1-to-10 class the lowest. BLS described the same shape for 2023 in November 2025: “rates range from 1.0 cases per 100 full-time workers among establishments with 1–10 employees to 3.1 cases for establishments with 50–249 employees.” The question this study asks is whether that national shape also holds inside individual industries.

Inside industries: where the highest rate falls

Before any detailed-industry size-class rate was read, the study listed the most detailed private industries in the 2024 table that do not overlap one another: 652 industries. Industries whose injury data BLS takes from mine-safety or railroad regulators were excluded first, because BLS states those records are not comparable with other industries. For each of the 652, the study checked whether BLS published a 2024 mean rate for all five size classes. BLS published all five for 17. In the other 635, at least one size class was not published.

Among those 17 industries, the class with the highest total recordable case rate in 2024:

The study’s single registered result is the count in the two middle classes: 8 of 17, or 47.1%. BLS does not publish standard errors for industry-by-size rates, so no significance is claimed. The 17 industries are not a sample of the economy. They are the 17 industries for which BLS published rates in all five size classes, and they span manufacturing, wholesale and retail trade, restaurants, health care and social assistance, construction, transportation, oil and gas support, and civic organizations (see Appendix A). The result describes them and nothing wider.

The same count for two narrower measures, in the same 652 industries:

The industry-level picture is more spread out than the national one. The national rate peaked at 50 to 249 in every year shown, but within these industries the peak falls in every size class, including the smallest.

Averages and the middle of the distribution

For the larger size classes, BLS also publishes quartiles of establishment rates. In 2024, private industry:

In each class, the published mean is above the median. The class mean and the rate of the establishment in the middle are different summaries, and the mean and median alone do not establish the full shape of the distribution. Quartiles are not published for the two smallest classes.

A known limitation: small establishments are less likely to keep records

Lower rates at small establishments may partly reflect recordkeeping rather than safety. A BLS follow-up survey of SOII respondents found that “small establishments are less likely than midsized and large establishments to keep records” (Elizabeth Rogers, “The Survey of Occupational Injuries and Illnesses Respondent Follow-Up Survey,” Monthly Labor Review, May 2020). That finding concerns recordkeeping practice, and it does not measure how many cases go unrecorded at each size. It is still reason to treat the 1-to-10 figures, nationally and inside industries, with care.

What the numbers mean for an operator

These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.

The national size pattern is a useful starting point and a poor benchmark for any one establishment. Within the few industries where a comparison is possible, the size class with the highest rate varies, and in four of the seventeen it was the smallest. A growing company in the middle classes should not assume it is in the most dangerous size band, and a small one should not assume it is in the safest. The relevant comparison is each workplace’s own industry and its own establishment size class, where BLS publishes it. A company with several workplaces should compare each one separately.

The gap between mean and median in the larger classes points the same way. When the mean sits far above the median, the class average is higher than the rate of the workplace in the middle, so the two should be read as different summaries. An operator who only checks the class average can be below it and still above the median workplace.

How to handle it: count properly, then compare

The steps below are an operating recommendation. This study did not test them, and BLS does not endorse them.

1. Calculate your own rate the BLS way. For comparison with these figures, use nonfatal recordable injury and illness cases in the year times 200,000, divided by total hours worked by all employees. BLS excludes fatalities from these rates. Use the OSHA recordkeeping rules that apply to the business to decide which cases are recordable, and ask a qualified safety or compliance professional if it is unclear which rules apply. A rate calculated another way cannot be compared with the published figures.

2. Compare against your industry and size, not the national average. For each workplace, find the most detailed industry row that matches it and the size class that matches that establishment’s average annual employment, not the company’s current or total employment. If BLS does not publish that cell, use the industry’s all-sizes rate and say so. Track the comparison every year rather than once.

3. Fix the records before reading the rate. If reliable records of cases and hours are missing or inconsistent, do not read a calculated rate as low. Treat it as unknown until the records are fixed. Where an OSHA log is required, or kept voluntarily, assign one person to own it, review it monthly with supervisors, and record near misses separately so they can be acted on before they become cases.

4. Look for the cases behind the number. Group recorded cases by task, location, shift and length of service. A small number of tasks or a single location often accounts for a large share of cases, and that is where controls pay off first.

5. Expect the rate to move as headcount grows. Moving from 40 to 80 employees usually means new supervisors, new shifts and less direct oversight by the owner. Put the safety routines in writing before that happens, not after the first serious case.

What this data does not tell you

• It does not show why injury and illness rates differ by size, or that any practice lowers them.
• The industry-level result covers only 17 industries where BLS published all five size classes. It does not describe the other 635.
• No standard errors are published for industry-by-size rates, so differences between classes inside an industry may not be meaningful.
• Small-establishment rates may be understated, because small establishments are less likely to keep records.
• Establishment size is the size of one workplace, not the company.

Appendix A: the 17 industries with all five size classes published (2024, total recordable cases)

Appendix B (CSV download), published with this article, lists all 652 industries in the analysis set, including the 635 without all five classes, and marks every size class for which BLS did not publish a rate.

Method

Already published by the agency: BLS publishes rates by industry and establishment size, with quartiles, in its SOII quartile tables, and has published the national size pattern. New here: a count, across the most detailed industries, of which size class has the highest rate, and the same count for two narrower case types.

Source. U.S. Bureau of Labor Statistics, Survey of Occupational Injuries and Illnesses, 2024 quartile tables 1 (total recordable cases), 2 (days away, restriction or transfer) and 3 (days away from work), and quartile table 1 for each year from 2019 to 2023, national, private industry, downloaded October 3, 2026.

Analysis set. Built from industry labels and NAICS codes only, before any detailed-industry size-class rate was read: the 1,020 private-industry rows with a six-digit NAICS code, less the 31 rows BLS footnotes as drawn from mine-safety (MSHA) or railroad (FRA) data, reduced to the 652 lowest-level rows that do not overlap. A size class with no published mean is treated as missing, never as zero. An industry’s highest class counts only if one class is strictly higher than all others. Otherwise the industry is counted as tied.

Protocol. The analysis set, the measure and the counting rule were fixed in a dated protocol before any rate was computed. The protocol (wave 2 section, frozen October 4, 2026) and its dated amendments are at https://kamyarshah.com/public-data-protocol/

Check. All figures were recomputed from the raw files by a separately written program, built without access to this study’s results. The 2023 private-industry size-class rates match those BLS published in November 2025 (1.0, 2.1, 3.1, 2.8, 2.7).

The agencies did not review or endorse this analysis. Kamyar Shah sells fractional COO and CMO services.

Questions about this analysis can be sent through https://kamyarshah.com/contact/.

The Census Bureau’s Quarterly Workforce Indicators (QWI) are built from state unemployment insurance wage records linked to Census business data. This article uses QWI average monthly earnings of new hires into stable jobs, based on covered UI wages. Census defines the measure as “Average monthly earnings of newly stable employees (i.e., full-quarter employees who were new hires with a firm in the previous quarter).”

Two features of that definition matter. First, it counts only new hires who stayed long enough to work a full quarter. Hires who left quickly are not included, so this is a selected group, and the share of hires who stay may differ by firm age. Second, it is earnings, not a wage rate: it includes no information on hours, and it covers UI-reported wages only, not benefits or other compensation.

Firm age is measured for the whole national firm: “Firm age and size are defined at the national level, rather than the state level,” according to Census, and are reported for private-sector firms only.

Earnings by firm age

Private sector, all industries, both sexes, all worker ages, nominal dollars per month, not seasonally adjusted. The recent-window columns are means of four quarterly values, 2024Q3 to 2025Q2. The last column is the mean of 12 quarterly ratios, 2017 to 2019:

Source: U.S. Census Bureau, LEHD Quarterly Workforce Indicators, national files by firm age, release R2026Q3, downloaded October 3, 2026. Ratios are means of the four quarterly ratios, so they can differ slightly from dividing the dollar means shown.

New-hire earnings rise with firm age in the recent window. Against firms aged 11 or more, the ratio for every younger class was between 78.9% and 85.9%.

The registered comparison

The study set one comparison in advance: for each quarter, new-hire earnings at firms aged 0 to 1 year divided by the same measure at firms aged 11 or more, averaged across the latest four quarters, at or before 2025Q3, in which every value used had a QWI status of 1 or 6.

Dividing the four-quarter dollar averages instead gives the same 78.9%. In 2017 to 2019 the mean ratio was 82.1%.

Census says QWI uses “a system of multiplicative noise infusion, whereby all released data are ‘fuzzed.'” Every value carries a status flag, and this analysis used only values with status 1 (OK) or 6 (calculated from other released measures, no significant distortion). Census publishes no sampling error or other margin of error for QWI, and a covariance-aware interval for this ratio is not available from the public tables. No statistical significance is claimed, and values can be revised in later releases.

New hires against all stable employees

The fourth column of the table compares new hires with all stable employees at firms of the same age. At firms aged 0 to 1, new hires averaged 81.9% of what stable employees earned. At firms aged 11 or more, they averaged 62.1%. Put differently, the earnings gap between young and established firms is wider for stable employees than for new hires: dividing the four-quarter averages, stable employees at firms aged 0 to 1 earned about 60% of what stable employees at firms aged 11 or more earned, against 79% for new hires. Workforces at firms aged 0 to 1 have, by definition, shorter possible tenure with the firm than workforces at firms aged 11 or more. The data does not isolate why the two gaps differ.

What the numbers mean for an owner

These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.

A young company recruits against established employers whose new hires, on average, earn more. The data cannot say whether that reflects different jobs, different industries, fewer hours or lower pay for comparable work, and the table should not be used as a pay benchmark. What the data show is an aggregate earnings difference between these groups. As operating judgment, a young company should expect some candidates to compare its offer with established employers, and should be ready to compete on the whole offer.

The second pattern is useful in a different way. New hires at the youngest firms earned much closer to what all stable employees there earned (81.9%) than new hires at firms aged 11 or more did (62.1%), a difference that may reflect tenure, seniority, career progression and other factors the data does not isolate. A young company that wants to keep the people it hires will eventually need the same thing: a visible path from entry earnings to something better.

How to handle it: compete on the whole offer, and plan the second year

The steps below are an operating recommendation. This study did not test them, and the Census Bureau does not endorse them.

1. Benchmark pay against the right comparison. National averages across all industries say little about a specific role. Use local pay data for the actual occupation and level, and decide deliberately whether to pay at, above or below that market.

2. Put the whole offer in writing. Base pay, any variable pay, benefits, flexibility, and what the role will lead to. A young company that cannot match an established employer on salary often can on scope, speed of advancement and direct access to leadership, but only if the candidate can see it.

3. Define what the second year looks like. Write down, for each core role, what someone must achieve to earn the next step in pay and responsibility. It turns a modest starting figure into a credible trajectory, and it gives the company a retention tool that costs nothing until it is earned.

4. Track who stays past the first quarter. The Census measure counts people who were new hires in one quarter and full-quarter employees in the next. A young company should define its own checkpoint, such as the share of hires still employed after ninety days, and track it with what those employees earn.

5. Revisit pay when the company’s position changes. As revenue becomes predictable, the case for paying below the market weakens. Set a date, such as the close of each fiscal year, to review pay against the market and against the company’s own results.

What this data does not tell you

• It does not show what any firm pays for a given job, or the hourly rate behind these earnings.
• It covers only new hires who stayed a full quarter, a selected group.
• Earnings are nominal UI-covered wages and exclude benefits and other compensation.
• Differences by firm age reflect industry, occupation, location, hours and worker mix as well as pay practice, and none of these is isolated.
• QWI is noise-protected administrative data with no published sampling error, and values can be revised.

Method

Already published by the agency: Census publishes QWI earnings of new hires into stable jobs and of stable employees by firm age every quarter. New here: four-quarter and pre-pandemic averages across firm-age classes, the registered young-versus-established ratio quarter by quarter, and the comparison of new hires with stable employees within each firm age.

Source. U.S. Census Bureau, LEHD Quarterly Workforce Indicators, national firm-age file, release R2026Q3, data through 2025Q3, downloaded October 3, 2026. Selection: private ownership, all industries, both sexes, all worker ages, not seasonally adjusted.

Window rule. The current window is the latest four quarters, at or before 2025Q3, in which every value used had a status of 1 (OK) or 6 (calculated from other released measures, no significant distortion): 2024Q3 to 2025Q2. Values with any other status are treated as missing, never as zero. The baseline is 2017Q1 to 2019Q4.

Protocol. The comparison, the window rule and the baseline were fixed in a dated protocol before any value was computed. The protocol (wave 2 section, frozen October 4, 2026) and its dated amendments are at https://kamyarshah.com/public-data-protocol/

Check. All figures were recomputed from the raw files by a separately written program, built without access to this study’s results. As a second check, the ratio was rebuilt from state-level values for the 48 states and the District of Columbia that the Census data service returned for these quarters, weighting each state by its number of all new hires, because the Census data service does not serve the count of new stable hires that matches this earnings measure. Alaska and Michigan returned no data. Both limits make this a partial, approximate check. The quarterly ratios from that rebuild (79.8%, 81.7%, 78.5% and 77.5%) agree with the national file to within 0.7 point.

The agencies did not review or endorse this analysis. Kamyar Shah sells fractional COO and CMO services.

Questions about this analysis can be sent through https://kamyarshah.com/contact/.