A persona that carries more than nine attributes describes, statistically, nobody. That is not a rhetorical flourish. It was measured across six real customer survey datasets, and the result held in five of the six, including every consumer dataset.

The deeper problem was named earlier, in 2006, by two Microsoft researchers who set out to evaluate the method honestly:

“Personas cannot be adequately verified or falsified and therefore have no demonstrable validity.”

And their reasoning, which is hard to argue with:

“Because personas are admittedly fictional, it would seem that there is no way to falsify them; no data can disprove a fictional construction. Therefore, they are outside the scientific method and cannot be verified.”

They added, and nobody has contradicted it since: “To our knowledge, no critical tests have been proposed in the persona literature, nor is there an example where a persona was refuted by data.”

This article is about how to build something you could be wrong about instead. Which requires knowing what the research actually says about interviews, about attributes, and about what a room full of colleagues can and cannot produce.

The arithmetic that kills the workshop persona

Take a published persona of the kind every agency produces. The 2006 paper takes one from the literature, a small business owner, and counts at least 21 attributes: age, role, company size, budget, technical comfort, decision style, three pain points, and so on.

Under independence and generous base rates of one in two per attribute, the share of the population matching all 21 is 0.5 to the power of 21. That is 0.000048 percent, or roughly 134 people in the United States.

Their conclusion: “Without that data, one cannot know whether a persona represents a million people or zero.”

Two years later the same team put numbers on it properly. They took six real survey datasets collected for actual market segmentation work, ranging from 268 to 10,307 respondents, four American consumer surveys, one French consumer survey and one survey of IT administrators. They then generated 10,000 randomly constructed persona-like descriptions per dataset, sampling each attribute level at its observed prevalence so as not to stack the deck toward rare combinations.

The finding, verbatim:

“the 99th percentile level fails to match anyone when 9 or more attributes are combined in 5 out of 6 survey datasets, including all consumer datasets.”

Even in the one dataset where descriptions survived longer, only in the IT survey did the 99th percentile match “more than 0.3% of real respondents with 7 or more attributes.”

Their own honest limit is worth reproducing, because it is the strongest defense of personas anyone has offered: “it is possible that the method could be useful for inspirational purposes even if the information claims are wrong, i.e., even if personas do not actually describe people.”

That is a defensible use. It is just a completely different one from the use most companies think they are making.

How the prevalence of a persona description falls as the number of attributes increasesHow the prevalence of a persona description falls as the number of attributes increases, based on two studies. The first, published in two thousand six, took a published persona of a small business owner from the persona literature and counted at least twenty one attributes. Under an assumption of independence and generous base rates of one in two per attribute, the share of the population matching all twenty one attributes is one half raised to the power of twenty one, which equals zero point zero zero zero zero four eight percent of the population, or approximately one hundred and thirty four people in the United States. The authors concluded that without supporting data one cannot know whether a persona represents a million people or zero. The second study, published in two thousand eight by the same team, used six real survey datasets collected for actual market segmentation work, ranging from two hundred and sixty eight to ten thousand three hundred and seven respondents, comprising four American consumer surveys, one French consumer survey and one survey of information technology administrators, plus two simulated datasets of two thousand respondents each. They generated ten thousand randomly constructed persona like descriptions per dataset, varying the number of attributes from two to eleven and the number of levels per attribute from two to eleven, sampling each attribute level at its observed prevalence so as not to bias toward rare combinations. They found that the ninety ninth percentile description fails to match anyone once nine or more attributes are combined, in five of the six survey datasets including all consumer datasets, and that only in the information technology dataset did the ninety ninth percentile performance level match more than zero point three percent of real respondents with seven or more attributes. Their declared limitation is that the work does not address the utility of personas, and that the method could be useful for inspirational purposes even if its information claims are wrong.Every attribute you add removes peopleA published persona, 21 attributes~134 peoplein the entire United States, under independenceand base rates of one in two per attributeMeasured on six real datasets9 attributesis where the 99th percentile descriptionstopped matching anyone, in five of sixThe method, so you can judge itSix survey datasets from real segmentation projects, 268 to 10,307 respondents. 10,000 randomly generatedpersona-like descriptions per dataset, each attribute level sampled at its observed prevalence.The authors’ own caveatThe method “could be useful for inspirational purposes even if the information claims are wrong.”
Six real datasets, 10,000 generated descriptions each. Past nine attributes, the 99th percentile description matched nobody in five of six. Source : Chapman, Love, Milham, ElRif and Alford, Quantitative evaluation of personas as information, HFES 52nd Annual Meeting, 2008; Chapman and Milham, HFES 50th Annual Meeting, 2006 (2008)

How many interviews, honestly

The number everyone quotes is twelve. Its source is a 2006 study of 60 interviews with women in Nigeria and Ghana, on HIV risk among sex workers, where saturation occurred within the first twelve interviews.

Read the conditions the authors attach: a homogeneous corpus, one occupation, one gender, a narrow research objective, and thematic analysis by coding. They make no claim to generality.

What the field has established since:

  • A systematic review of 23 empirical tests of saturation, published in 2022, reports that studies reached saturation “within a narrow range of interviews (9-17)”, specifically for “relatively homogenous study populations and narrowly defined objectives”.
  • A 2017 study separated two things that get confused. Code saturation was reached at nine interviews, at which point the range of topics was known. Meaning saturation took 16 to 24. The authors put it well: code saturation indicates when researchers have “heard it all”, but meaning saturation is needed to “understand it all”.
  • Across multiple sites, the numbers multiply. A 2017 study across four sites found up to 16 interviews sufficed per site for local themes, but 20 to 40 per site were needed for themes that cut across sites.
  • And the uncomfortable one. A 2024 secondary analysis of five federally funded studies found that true saturation arrived only after 30 to 67 interviews, while what those authors call near saturation arrived at 15 to 23. On that reading, the famous twelve is near saturation, not saturation.

That last study contains a detail worth carrying. Studies with a tightly structured interview guide and deductive coding saturate faster. Which means saturation partly measures how closed your questions were, not how much the market had to tell you.

Two respected methodologists reject the concept altogether, arguing that judgments about when to stop “are inescapably situated and subjective, and cannot be determined (wholly) in advance of analysis”. The author of the original 2006 study now recommends abandoning the magic number in favor of three declared parameters: how many interviews form your base, how long a run without new information counts, and what threshold of new information you treat as meaningful.

The practical version: for one narrow segment, plan on the mid-teens, expect to keep going if the answers keep surprising you, and write down the rule you used to stop.

Published interview counts required to reach saturation and what each figure measuresPublished interview counts required to reach saturation in qualitative research, and what each figure actually measures. The widely cited figure of twelve interviews comes from a two thousand six study of sixty in depth interviews with women in Nigeria and Ghana on HIV risk among sex workers, where saturation occurred within the first twelve interviews, but the authors condition the result on a homogeneous corpus with one occupation, one gender, a narrow research objective and thematic analysis by coding, and make no claim to generality. A systematic review published in two thousand twenty two covering twenty three empirical tests of saturation reports a range of nine to seventeen interviews, specifically for relatively homogeneous study populations and narrowly defined objectives, and four to eight focus group discussions. A two thousand seventeen study separated code saturation, reached at nine interviews when the range of topics is known, from meaning saturation, which took sixteen to twenty four interviews to develop a richly textured understanding, with the authors noting that code saturation indicates when researchers have heard it all while meaning saturation is needed to understand it all. A two thousand seventeen multi site study of one hundred and thirty two respondents across four sites found that up to sixteen interviews sufficed per site for site specific themes but twenty to forty interviews per site were required for themes cutting across sites. A two thousand twenty four secondary analysis of five federally funded studies with thirty to seventy interviews each found that true saturation arrived only after thirty to sixty seven interviews while near saturation arrived at fifteen to twenty three, meaning the famous figure of twelve corresponds to near saturation rather than saturation. That study also found that research with a tightly structured interview guide and deductive coding saturates faster, so saturation partly measures how closed the questions were rather than how much the field had to say.Every published number, with its condition12One 2006 study, 60 interviews, one occupation, one gender, one narrow question.The authors claim no generality. This is the number everyone repeats.9 to 17Systematic review of 23 empirical tests, 2022. For homogeneous populationsand narrowly defined objectives only.9 vs 16 to 24Code saturation at 9, meaning saturation at 16 to 24. “Heard it all”against “understand it all”.20 to 40Per site, for themes that cut across several markets or segments.Four-site study, 132 respondents, 2017.30 to 67True saturation, 2024 analysis of five studies. Near saturation at 15 to 23,which is where the famous twelve actually sits.
The famous twelve is one dataset with a homogeneous population. Nothing in this table is a constant. Source : Guest, Bunce and Johnson, Field Methods 18(1), 2006; Hennink and Kaiser, Social Science and Medicine 292, 2022; Hennink, Kaiser and Marconi, Qualitative Health Research 27(4), 2017; Hagaman and Wutich, Field Methods 29(1), 2017; Squire et al., JMIR 26, 2024 (2024)

Why a workshop cannot produce what an interview produces

The persona built in a meeting fails for a reason that has been measured, and the measurement is clean.

In a randomized split-sample experiment after the 2008 election, people were asked what issue mattered most in their vote. Half got a closed question with five options. Half got the same question open-ended.

  • 58 percent chose the economy when it was on the list.
  • 35 percent volunteered it when it was not.
  • And 43 percent of the open-ended group gave an answer that appeared nowhere on the closed list, against 8 percent in the closed condition.

A workshop runs exclusively in closed mode. The team proposes plausible motivations and converges. It mechanically produces the 58 percent, never the 35 percent, and never the 43 percent of answers nobody in the room thought to list. That last group is the only content that would have justified doing research at all.

There is a second, harder problem, and it applies to interviews too if you ask the wrong question. A foundational 1977 paper argued that people have little or no direct introspective access to their own higher-order reasoning, and that when they explain their decisions they draw not on introspection but on plausible causal theories. A 2005 experiment in Science made it vivid: participants chose between faces, were secretly handed the face they had not chosen, failed to notice, and then confidently explained why they had chosen it. The authors named it choice blindness.

So asking a buyer “why did you choose this supplier” produces their theory of why, not the reason. Asking what happened, in order, with dates and names, produces something you can check.

And sensitive topics get edited. A 2007 review in Psychological Bulletin concluded that misreporting on sensitive topics is common and situational, and that respondents “edit the information they report to avoid embarrassing themselves in the presence of an interviewer”. In a business interview, “why did you leave your previous supplier” and “who actually decided” are both sensitive questions.

Results of a randomized comparison between a closed and an open version of the same questionResults of a randomized split sample comparison between a closed and an open ended version of the same question, conducted after the two thousand eight United States presidential election. Respondents were asked what one issue mattered most to them in deciding how they voted for president. One randomly assigned half received a closed ended version offering five options with the possibility of volunteering an option not on the list, and the other half received an open ended version. When explicitly offered the economy as a response option, more than half of respondents, fifty eight percent, chose that answer. Among those who received the open ended version, only thirty five percent volunteered the economy, a gap of twenty three percentage points on the same population with random assignment. Furthermore, among those asked the closed ended version, fewer than one in ten, eight percent, provided a response other than the five they were read, whereas fully forty three percent of those asked the open ended version provided a response that did not appear on the closed list at all. The implication for building marketing personas is that a workshop operates exclusively in closed mode, since the team proposes plausible motivations and converges on them, so it mechanically produces the equivalent of the fifty eight percent figure, never the thirty five percent figure, and never the forty three percent of answers that nobody in the room thought to put on the list, which is precisely the only content that would have justified conducting research in the first place.The same question, asked two waysRandomly assigned halves of the same sampleChose the economy, offered in a list58%Volunteered it, open-ended35%Answers that were not on the list at allOpen-ended version43%Closed version8%
The 43 percent of off-list answers is the only part of the research that a workshop can never generate. Source : Pew Research Center, Writing Survey Questions, post-2008 election split-sample experiment (2019)

Size the segment with public data, not with a character sketch

The part of a persona that is genuinely knowable is the part nobody bothers to look up, and it is free.

The Census Bureau’s Statistics of U.S. Businesses gives firm counts, establishment counts, employment and payroll by employment size band, by industry code, by state and by county. The most recent annual tables cover reference year 2022 and were published in April 2025.

The national picture: 6,395,635 employer firms, 8,298,562 establishments, 135,748,407 employees.

Now the part that matters if you sell to mid-size companies. Firms employing 100 to 999 people number 103,741. That is 1.62 percent of all employer firms, and 18.8 percent of private employment, or 25,524,994 people.

By sector, at that same size band:

SectorFirmsEmployees
Health care and social assistance20,2604,718,066
Manufacturing14,2073,123,651
Accommodation and food services10,8232,414,244
Professional, scientific and technical services10,4602,134,934
Administrative and support services10,2242,276,492
Wholesale trade8,9271,477,843
Retail trade8,5741,479,617
Construction7,8811,584,398
Finance and insurance4,5621,085,857
Information2,825643,528

Two cautions so you use this correctly. The sector rows do not sum to the total, because a firm operating in several sectors is counted in each one where it has establishments. And Statistics of U.S. Businesses counts firms, while County Business Patterns counts establishments. If you sell to a plant, a branch or a store, you want the second file. Confusing the two is the classic market-sizing error.

The same file goes down to state and county. Manufacturers with 100 to 999 employees, by state: California 1,867, Texas 1,425, Ohio 1,217, Pennsylvania 1,056, Illinois 1,001, Michigan 988, Wisconsin 894.

That is a segment you can count, reach, and be wrong about. It is the opposite of a character with a stock photo. Checking whether you were wrong only works when the same definition runs through the positioning, the site and the campaigns, which is why we run the three as one system rather than as three briefs.

United States employer firms and employment by enterprise employment size bandUnited States employer firms and employment by enterprise employment size band, from the Census Bureau Statistics of U.S. Businesses annual data tables for reference year two thousand twenty two, published in April two thousand twenty five and revised in July two thousand twenty five, using the two thousand seventeen North American Industry Classification System. The national totals are six million three hundred and ninety five thousand six hundred and thirty five employer firms, eight million two hundred and ninety eight thousand five hundred and sixty two establishments, and one hundred and thirty five million seven hundred and forty eight thousand four hundred and seven employees. By size band, firms with fewer than twenty employees number five million seven hundred and twenty thousand and ninety three and employ twenty one million nine hundred and fifty thousand one hundred and eighty four people. Firms with twenty to ninety nine employees number five hundred and sixty one thousand one hundred and sixty and employ twenty one million nine hundred and seventy six thousand seven hundred and one people. Firms with one hundred to nine hundred and ninety nine employees number one hundred and three thousand seven hundred and forty one, representing one point six two percent of all employer firms, and employ twenty five million five hundred and twenty four thousand nine hundred and ninety four people, representing eighteen point eight percent of private employment. Firms with one thousand or more employees number ten thousand six hundred and forty one and employ sixty six million two hundred and ninety seven thousand five hundred and twenty eight people. Two cautions apply to using this data. Sector rows do not sum to the national total because a firm operating in several sectors is counted in each sector where it has establishments. And Statistics of U.S. Businesses counts firms while the separate County Business Patterns program counts establishments, so a company selling to individual plants, branches or stores should use the establishment file instead, since confusing the two is a common market sizing error.6,395,635 employer firms, sorted by sizeReference year 2022, published April 2025Under 20 employees5,720,09320 to 99561,160100 to 999103,7411,000 and over10,641The 100 to 999 band, in proportion1.62 percent of employer firms. 18.8 percent of private employment, or 25,524,994 people.Two trapsSector rows do not sum to the total. And this file counts firms; a separate file counts establishments.
Free, official, and revisable. Firms with 100 to 999 employees are 1.6 percent of companies and 18.8 percent of employment. Source : U.S. Census Bureau, Statistics of U.S. Businesses, 2022 Annual Data Tables by Establishment Industry, published April 2025 (2025)

The buying committee number you should stop citing

While we are auditing inputs, one more.

The claim that a complex business purchase involves six to ten decision makers appears in thousands of marketing articles. It was published on its originator’s own page for years, worded exactly that way and attributed to one of its executives.

It never carried a methodology. The full page contains no sample size, no fieldwork date, no definition of what counts as a decision maker, and no link to an underlying report.

And it is gone. Comparing successive archived captures of that page, the sentence is present on 8 October 2023 and absent on 25 October 2023. The page still exists, still runs to about twenty thousand characters, and still discusses buying groups at length. Only the number was removed.

The measurement it derives from was 5.4 people, from a 2014 survey of more than 3,000 stakeholders. The commercial successor of that research now publishes a progression of 5.4, then 6.8, then 10.2, and most recently 11.2, on a study of 667 buyers of complex purchases. The original publisher’s current position, from a survey of 632 buyers conducted in August and September 2024, is that buying groups range “from five to 16 people across as many as four functions”, which is a range, not an average.

No public institution measures this. Every available figure comes from a commercial research firm, on a self-selected sample of enterprise software buyers, with the underlying report behind a client wall.

Timeline of the six to ten decision makers claim and the measurements around itTimeline of the widely repeated claim that a complex business to business purchase involves six to ten decision makers, together with the actual measurements around it. In two thousand fourteen a survey of more than three thousand stakeholders involved in a typical business purchase produced a stated average of five point four people formally involved in a purchase decision. From two thousand twenty until October two thousand twenty three, the publisher’s page on the business to business buying journey stated that the typical buying group for a complex business to business solution involves six to ten decision makers, each armed with four or five pieces of information gathered independently and requiring deconfliction with the group, attributed on the page to one of its executives. The full text of that page contained no sample size, no fieldwork date, no definition of what counts as a decision maker, and no link to an underlying report, so the claim never carried a published methodology. Comparing successive archived captures of the page shows the sentence present on the eighth of October two thousand twenty three and absent on the twenty fifth of October two thousand twenty three, while the page itself continued to exist at around twenty thousand characters and continued to discuss buying groups at length, meaning only the number was removed. The commercial successor of the original research now publishes a progression of five point four, then six point eight, then ten point two, and most recently eleven point two people, the last from a study of six hundred and sixty seven buyers of complex purchases of which eighty one percent exceeded fifty thousand dollars. The original publisher’s current position, from a survey of six hundred and thirty two buyers conducted in August and September two thousand twenty four, is that buying groups range from five to sixteen people across as many as four functions, which is an observed range rather than an average. No public institution measures buying committee size, and every available figure comes from a commercial research firm on a self selected sample with the underlying report behind a client wall.A number that was never measured, and is now gone2014Survey of 3,000+stakeholders5.4 people2020 to 2023Published claim, nomethodology at all”six to 10”Oct 2023Present on the 8th,absent on the 25thremoved2024 survey632 buyers, a rangerather than a mean5 to 16What the page contained while the claim was liveNo sample size. No fieldwork date. No definition of “decision maker”. No link to a report.No public institution measures this. Every figure comes from a commercial firm, report behind a client wall.
Present on 8 October 2023, gone on 25 October. The page is still there. Only the number was removed. Source : Archived captures of the publisher's B2B buying journey page; CEB 2014 survey as reported by its commercial successor; publisher press release of 7 May 2025 (2025)

What to build instead

Count the segment before you describe anyone in it. Public data will tell you how many firms exist, where, at what size, with what payroll. If the segment is too small to matter, you have saved yourself the interviews.

Keep the description under a handful of attributes. Past nine, the research says you are describing nobody. Three or four that genuinely predict behavior beat twelve that make a nice slide.

Interview, and count the interviews honestly. Mid-teens for one narrow segment, more if answers keep surprising you, 20 to 40 per site if the segment spans several markets. Write down the rule you used to stop.

Ask what happened, not why. Sequence, dates, who was in the room, what was tried first, what the alternative was. People reconstruct their reasons and cannot help it. They remember events.

Make the thing refutable. Write the description as a set of claims that could turn out to be false: this segment buys after an external trigger, evaluates two or three suppliers, decides in under a quarter, and involves the finance function. Then go and check. If nothing in your persona could ever be shown wrong, it is not research, and it will not survive contact with the first customer who does not fit.