Three families. Nineteen sources.

Not a menu. A checklist.

The grid is not a menu; it is a checklist. The more of it a project reads, the more immersive the insight and the more impactful the measurement, for three reasons:

Triangulation.

An insight that appears in all three families — asked, observed, assembled — has been found three different ways. That is the cheapest confidence you will ever buy, and it raises t (transferability) without a single extra interview.

Immersion.

Each source adds a dimension: what people say, what they do, what already exists about them. Together they produce a picture you can walk around in, which is what personas need.

Measurement.

Every source read for Build becomes a baseline for Prove. Coverage now is measurement later.

Coverage was always desirable and rarely affordable, because someone had to read it all and hold it in one head. That is the constraint AI removes: it reads the whole grid together and keeps the sources in one frame, which is what “making sense of all of it together” means in practice. The human role moves to designing the coverage and judging the synthesis.

Sample size is how you buy probability.

The reframe. The classic justification for small qualitative samples is saturation. The empirical literature is narrow: saturation is reached within 9–17 interviews or 4–8 focus groups, mainly with homogeneous populations and narrow objectives.^S8 Finer cut: code saturation (the range of issues) around 9 interviews; meaning saturation (fully understanding them) around 24.^S9 Focus groups: four to identify the issues, more to understand them.^S10 But small samples were never only about saturation. They were also about analyst throughput — Kvale's “1,000-page question”: you cannot analyse data you cannot read.^S11 AI removes the throughput constraint. The question changes from “when can I stop” to “what job is the qual doing.”

Three jobs, sized per cell

Quant per cell: ~100 for a directional comparison; ~385 total for ±5 at 95%.

Starter numbers: quantitative thresholds

Margin of error at 95% confidence (approximate, 1/√n; exact figures in^{S24, S25}):

Doubling from 1,000 to 2,000 buys about one point^S26. The margin applies to the total sample only; every subgroup has a wider one^S26.

Starter targets by population

ScaleQ2 defaults; adjust to the decision

Two rules over the table. Coverage before count: a sample that misses a market, a role or a demographic is wrong at any n. Never report below n=25. It protects anonymity and it keeps a chart honest about its width.

Total = cells × per-cell minimum. Count the cells first. Four usage states × 30 = 120 interviews to compare them. Three persuadability tiers × two audiences × 10 = 60 to discover.

Norms with real support

• Human first, then scale: ~10 well-designed interviews, review, then 50 or 100.^S14
• Read the raw edges: AI synthesis is the starting point; read 3–5 transcripts yourself.^S14
• Report code saturation and meaning saturation separately.^S9
• Scale does not fix recruitment. 500 AI interviews of a convenience sample is a bigger convenience sample.
• Quantify themes only with a sample built for it; below that, report prevalence within the sample and say so.^S13
• Synthetic respondents pretest instruments; they do not produce findings.
• Rule of thumb for the room: scale a cell until the next ten interviews change nothing; then stop and spend on recruiting the cell you are missing.

What doesn't scale

Human moderation on contested or emotionally complex topics. Executive interviews. Ethnography. Anything where the interviewer is the instrument.

The paper names its own limits so the “why not 500?” critique is answered before it is asked.

Prove it. Then bring it to life.