How to turn customer conversations into proof a skeptical buyer (and an AI answer engine) will actually believe

Here’s a confession from the marketing world. A lot of “customer research” is one founder, three friendly customers, and a whiteboard that says CUSTOMERS LOVE SPEED. That line goes on the homepage. A month later a skeptical buyer asks, “Says who?” and the room gets very quiet.

Qualitative research gets dismissed as anecdote because most of it is run like one. The fix is boring in the best way: a deliberate sample, a disciplined interview guide, real transcripts, a written codebook, honest counting, and a defensible reason for stopping. Do those six things with twenty people and you’ll out-argue two hundred hallway conversations. Below are the questions I hear most, answered with the actual studies, their numbers, and their charts.

Q1. Why does qualitative research get written off as “just anecdotes”?

Because the receipts are usually missing. Janice Morse, longtime editor of Qualitative Health Research, put it bluntly in 2015: saturation “is present in all qualitative research, but unfortunately, it is evident mainly by declaration.” The evidence backs her up:

1. Francis et al. (2010, Psychology & Health) reviewed 16 months of Social Science & Medicine. Of 18 articles that mentioned saturation, 15 claimed they reached it, and it was unclear how any of them defined or demonstrated it.
2. Carlsen & Glenton (2011, BMC Medical Research Methodology) reviewed 220 focus group studies. Of those that justified their sample size, 83% pointed to saturation, often with nothing to show for it, and sometimes while sticking to the number they picked before fieldwork started.
3. Mason (2010, Forum: Qualitative Social Research) analyzed 560 PhD studies built on interviews. Mean sample size: 31. The most popular sizes were 20 and 30, then 40, 10 and 25. Apparently saturation has a thing for round numbers.

If PhD committees let this slide, imagine what happens during a two-week messaging sprint. The good news is that the fix is procedural, and procedures can be copied.

Q2. How many interviews do I actually need?

Fewer than you fear and more than you’d like. Across very different fields, the discovery curve bends the same way: steep at first, then flat.

Look at the very first interview in the Hennink study: it produced 53% of all 45 codes. By interview 9 they had 91%, and the final 16 interviews added just four codes. Griffin and Hauser got to a similar place for product needs three decades ago. A caveat from usability land keeps everyone humble, though. Laura Faulkner (2003) re-ran the famous five-user test and found that individual groups of five caught anywhere from 55% to 95% of problems. Small samples average out nicely and vary wildly one at a time.

So why twenty? For one focused B2B segment, 12 to 15 interviews usually map the territory. Twenty buys room for two or three subgroups plus the deeper layer of meaning (more on that in Q8). It’s big enough to be taken seriously and small enough that you’ll actually finish.

Sources: Hennink, Kaiser & Marconi (2017), cumulative codes by interview; Guest, Bunce & Johnson (2006), as summarized in Guest, Namey & Chen (2020); Nielsen & Landauer model via Nielsen Norman Group; Hennink & Kaiser (2022) systematic review.

Table 1. What the sample-size research actually found

Study	What they examined	Headline finding
Griffin & Hauser, 1993 (Marketing Science, MIT)	Customer-needs interviews for product design	20 to 30 interviews capture 90% to 95% of needs; their 30 interviews got an estimated 89.8%
Nielsen & Landauer, 1993 (NN/g)	Usability tests	5 users find about 85% of problems, but only if each problem hits 31% of users
Guest, Bunce & Johnson, 2006 (Field Methods)	60 interviews, 114 themes	70% of themes by interview 6; 92% by interview 12
Mason, 2010 (FQS)	560 PhD interview studies	Mean of 31; suspicious clustering at 20 and 30
Hennink, Kaiser & Marconi, 2017 (Qual. Health Research)	25 interviews, 45 codes	Code saturation at 9; meaning saturation at 16 to 24
Hagaman & Wutich, 2017 (Field Methods)	132 interviews across four sites	Cross-site themes need roughly 20 to 40
Hennink & Kaiser, 2022 (Social Science & Medicine)	Review of 23 empirical saturation tests	9 to 17 interviews (mean 12 to 13) or 4 to 8 focus groups

Q3. Who belongs in the twenty?

This is purposive sampling: you choose people for the information they carry, and you write down why. For a positioning study, a mix like this works well:

1. 8 recent wins who bought in the last six months, while the memory of the decision is still fresh.
2. 5 lost deals who picked a competitor, a spreadsheet, or nothing at all. They know exactly where your story fell flat.
3. 4 churned or downgraded customers. Painful calls, priceless data.
4. 3 long-tenured power users who can tell you what the product turned into for them.

A few guardrails. Recruit from the CRM rather than the customer-marketing list of superfans; if all twenty gave you a 10 on NPS, you’re studying a fan club. Take one person per account. Plan for subgroups up front: Nielsen suggests at least 3 people per user group, and Hennink’s 2019 focus-group study found you need at least two groups per stratum before the deeper understanding shows up. Good planning pays off. When Hennink’s team added a second patient population after 16 interviews, it contributed exactly one new code, because the guide was already built for it.

Q4. What makes a good interview guide?

1. Anchor every question in a real event. “Walk me through the week you started looking” beats “What do you value in a vendor?” Past behavior tells the truth; predictions are wishful thinking with a microphone.
2. Keep it to 8 to 12 open questions for a 45-minute call, each with probes: “What happened next?” “What did that cost you?” “Who else cared?”
3. Ask about every alternative, including Excel, an intern, and doing nothing. That’s your real competitive set.
4. Collect their words. “When you explained this to your boss, what did you say?” Nine times out of ten that sentence is better copy than anything written in a conference room.
5. Save your product for the end, if it comes up at all. Pitching in the middle contaminates everything after it.
6. Pilot on two people, fix the guide, then freeze the core questions. Hennink’s team made 78% of their codebook definition changes by interview 6. Early churn is normal. Changing the questions at interview 14 wrecks your ability to count.

Q5. Do I really need transcripts, or are my notes good enough?

Notes record what you found interesting in the moment. Transcripts record what the customer said. Since the whole point of messaging research is to borrow customers’ exact language, that difference matters a lot. The old objection was cost, and Figure 4 shows why.

Sources: audiotranskription.de measured a mean of 6.3 hours of typing per audio hour (SD 1.5) and about 18 hours for GAT2 basic transcripts; a 2025 arXiv study of 12 doctoral interviews found AI drafts plus manual correction took 2 hours 19 minutes per audio hour, a 53.8% to 76.9% time savings; BrassTranscripts (2026) reports 30 to 40 minutes of verification per audio hour.

My recommendation: record with consent, let an AI tool draft the transcript, then listen at 1.25x speed while you read and correct every line. That pass does double duty as Braun and Clarke’s first phase, familiarization. Twenty 45-minute interviews goes from roughly 95 hours of typing to somewhere between 9 and 35.

Q6. How do I code for themes without fooling myself?

Start with Braun and Clarke’s “Using Thematic Analysis in Psychology” (2006). According to Google Scholar, as cited on Wikipedia, it has over 59,000 citations and is the most cited academic paper published in 2006. Its six phases: (1) get familiar with the data, (2) generate initial codes, (3) search for themes, (4) review them, (5) define and name them, and (6) write it up.

Then do three practical things. First, build a codebook: code name, definition, when to apply it, when to leave it off, and one example quote. In Hennink’s study, 44% of code definitions never changed at all, and 76% of codes were inductive, meaning customers raised them without being asked. Second, remember that a code is a label (“manual month-end reporting”) and a theme is a claim (“finance leads feel exposed when numbers arrive late”). Headlines are made from themes. Third, have a second person code about 20% of the transcripts (4 of 20) and measure agreement.

Both figures: Hennink, Kaiser & Marconi (2017), Qualitative Health Research, Table 1 and Table 2 / Figure 5. Figure 2 is discussed in Q8.

Table 2. How to read your inter-coder agreement (Cohen’s kappa)

Kappa	Landis & Koch (1977)	McHugh (2012)	Roughly how much data is reliable (McHugh)
0.21 to 0.39	Fair	Minimal	4% to 15%
0.40 to 0.59	Moderate	Weak	15% to 35%
0.60 to 0.79	Substantial	Moderate	35% to 63%
0.80 to 0.90	Almost perfect	Strong	64% to 81%
Above 0.90	Almost perfect	Almost perfect	82% to 100%

Notice that 0.65 earns a gold star on the 1977 scale and a shrug from McHugh. If a claim is going on your website, aim for 0.80 or better. One honest note: Braun and Clarke’s later “reflexive” version of thematic analysis explicitly rejects reliability scores. For marketing proof you plan to count and publish, a codebook approach with a second coder is the better fit. Just say which one you used.

Q7. How do I report frequency honestly?

This is where most decks go off the rails. With twenty people, one person equals five percentage points, so “70%” suggests a precision you don’t have.

1. Give counts with the denominator: “14 of 20 customers.”
2. Break it out by segment: “8 of 8 wins, 2 of 5 losses.” The split usually tells the real story.
3. Report the negative cases. Who didn’t mention it, and why?
4. Define your words. If you say “most,” say it means 11 or more of 20.
5. Keep the counts inside the sample. Twenty interviews tell you what exists and why. A survey tells you how many.

Table 3. Same finding, reported two ways (illustrative example)

Theme	Oversold version	Honest version
Month-end reporting pain	“70% of the market hates reporting”	“14 of 20 described month-end reporting as the trigger; 12 of those 14 used the word ‘scramble'”
Price sensitivity	“Price doesn’t matter to buyers”	“Price came up in 6 of 20; all 5 lost deals raised it”
Integrations	“Customers want integrations”	“3 of 20 mentioned integrations; each was a deal blocker”

That last row matters. Hennink’s team found no pattern linking how often a code appeared to how fast it was understood, and Morse (1995) warned that “it is often the infrequent gem that puts other data into perspective.” Three mentions that each killed a deal can outrank fourteen mild complaints.

Q8. When do new interviews stop teaching me anything?

There are two finish lines. Hennink calls them code saturation, when you’ve “heard it all,” and meaning saturation, when you “understand it all.” Figure 2 shows the gap. Concrete codes like “time” were fully understood by interview 9. Conceptual codes, the emotional stuff that makes great messaging, took 16 to 24 interviews, and “responsibility for health” was still producing new dimensions at interview 25. Guess which kind your positioning depends on.

To make stopping defensible, use the rule from Guest, Namey & Chen (2020, PLOS ONE). Pick a base (the themes from your first several interviews), a run length (the next 2 or 3 interviews), and a new-information threshold (they propose 5% or less, or zero). Francis et al. (2010) offer a simpler version: 10 interviews, then stop after 3 in a row with nothing new. Keep a log like this one:

Table 4. A saturation log you can paste into a methods note (illustrative example)

Run	New themes	New vs. base (32 themes)	Verdict
Interviews 1 to 6 (base)	32	n/a	Base set
Interviews 7 to 8	5	16%	Keep going
Interviews 9 to 10	3	9%	Keep going
Interviews 11 to 12	2	6%	Close
Interviews 13 to 14	1	3%	Code saturation (5% or less)
Interviews 15 to 20	1, then 0	3%, then 0%	Meaning checks on conceptual themes; stop

Q9. How does all of this become positioning and messaging proof?

1. Lift customer language straight into headlines, then test them. Borrowed words tend to beat invented ones.
2. Publish a short methods note with every claim: “20 interviews with U.S. finance leaders, June to August 2026; recorded, transcribed, double-coded (kappa 0.82); stopped after two consecutive runs under 5% new themes.” One paragraph turns an opinion into a citable finding, and AI search tools are far more likely to quote a claim that carries its own method.
3. Attribute quotes by role and segment, with permission.
4. Hand the top themes to a survey. Hennink notes that code saturation is often enough to write good survey items. Qual finds the why, quant sizes it.

A quick illustration. Picture a payroll software team convinced its story is “fastest setup in the category.” Twenty coded interviews tell a different tale: 13 of 20 buyers, unprompted, described the dread of Sunday-night payroll checks, and setup speed came up in 4. The new headline writes itself, and now it has receipts.

Bottom line: twenty interviews won’t make you statistically significant. They will make you specific, credible, and very hard to argue with, which is what proof is supposed to do.

Sources

1. Braun, V. & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology. https://doi.org/10.1191/1478088706qp063oa
2. Wikipedia: Victoria Clarke (citation count of Braun & Clarke 2006). https://en.wikipedia.org/wiki/Victoria_Clarke_(psychologist)
3. Carlsen, B. & Glenton, C. (2011). What about N? BMC Medical Research Methodology. https://doi.org/10.1186/1471-2288-11-26
4. Francis, J. et al. (2010). What is an adequate sample size? Psychology & Health. https://doi.org/10.1080/08870440903194015
5. Griffin, A. & Hauser, J. (1993). The Voice of the Customer. Marketing Science. https://mitsloan.mit.edu/shared/ods/documents?PublicationDocumentID=5259
6. Guest, G., Bunce, A. & Johnson, L. (2006). How Many Interviews Are Enough? Field Methods. https://journals.sagepub.com/doi/10.1177/1525822X05279903
7. Guest, G., Namey, E. & Chen, M. (2020). A simple method to assess and report thematic saturation. PLOS ONE. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0232076
8. Hennink, M., Kaiser, B. & Marconi, V. (2017). Code Saturation Versus Meaning Saturation. Qualitative Health Research. https://pmc.ncbi.nlm.nih.gov/articles/PMC9359070/
9. Hennink, M. & Kaiser, B. (2022). Sample sizes for saturation in qualitative research. Social Science & Medicine. https://www.sciencedirect.com/science/article/pii/S0277953621008558
10. Sample Size in Qualitative Research: Moving from Data Saturation to Theoretical Saturation (2025), summarizing Hagaman & Wutich (2017) and Hennink et al. (2019). Journal of Global Marketing. https://www.tandfonline.com/doi/full/10.1080/08911762.2025.2590757
11. Landis, J. & Koch, G. (1977). The measurement of observer agreement for categorical data. Biometrics. https://doi.org/10.2307/2529310
12. McHugh, M. (2012). Interrater reliability: the kappa statistic. Biochemia Medica. https://pmc.ncbi.nlm.nih.gov/articles/PMC3900052/
13. Mason, M. (2010). Sample Size and Saturation in PhD Studies Using Qualitative Interviews. FQS. https://www.qualitative-research.net/index.php/fqs/article/view/1428
14. Morse, J. (2015). Data were saturated… Qualitative Health Research (editorial). https://doi.org/10.1177/1049732315576699
15. Research Design Review (2020), quoting Morse (1995) on the “infrequent gem.” https://researchdesignreview.com/2020/07/07/sample-size-qualitative-research-risk-of-relying-saturation/
16. Nielsen, J. (2000). Why You Only Need to Test with 5 Users. Nielsen Norman Group. https://www.nngroup.com/articles/why-you-only-need-to-test-with-5-users/
17. Faulkner, L. (2003). Beyond the five-user assumption. Behavior Research Methods, Instruments, & Computers. https://doi.org/10.3758/BF03195514
18. audiotranskription.de. How long does it take to transcribe interviews? https://www.audiotranskription.de/en/how-long-does-it-take-to-transcribe-interviews/
19. Halving transcription time: AI-assisted transcripts for content analysis (2025). arXiv. https://arxiv.org/abs/2503.13031
20. BrassTranscripts (2026). Research Interview Transcription: Complete Guide. https://brasstranscripts.com/blog/interview-transcription-qualitative-research-guide