Three terms you need first.
Everything else in this Deep Dive builds on these. Any other technical term is explained where it first appears.
Synthetic persona
An AI model asked to answer a survey as if it were a type of customer, described in a prompt. For example: "a 35-year-old parent in Manchester on a mid-range income".
A group of them is a synthetic panel.
Digital twin
An AI copy of one specific, real person, built from that person's own survey answers, reviews or past choices. A persona imagines a type of person. A twin is modelled on a real one.
Several studies here test twins, and we say so when they do.
Van Westendorp
The classic four-question pricing survey: at what price would this be too cheap, a bargain, expensive, too expensive? Where the answer curves cross, it produces one "optimal" price.
Introduced by Peter van Westendorp at the 1976 ESOMAR Congress.1
The short version, in five points.
- 01
The classic pricing survey guesses
Van Westendorp asks people to name prices for a product they have never bought, with no competitors to compare against. The "optimal price" it produces often lands far from what people actually pay.
- 02
AI personas make it worse
Their answers sit much closer to the middle, about half as spread out as real people's. The customers who would pay a premium, and the ones who would walk away, are missing. The answers also move when you switch AI model, and the "same" persona behaves like a different person when you change the price.
- 03
Ask people to choose, not to name a price
Surveys that show a few options at set prices and ask which one people would pick, such as discrete choice or brand-price trade-off (BPTO), hold up better, for real people and for AI, as long as the product category already exists. Even then, an AI result is an early signal to test with people. It is not proof.
- 04
Feed the AI real behaviour, then check it with real people
Digital twins built from what people actually did, such as their past reviews, reached 87.73% accuracy on conjoint (choice-based) tasks, far better than personas built from a description. Checking an AI study against a human sample of just 50 to 300 people cut its bias by 83% to 99.6% in published tests.
- 05
Report the spread, not the median
The customers at the extremes decide how sharply demand falls when you raise the price, and they are exactly the ones AI panels erase. Run diagnostics before any synthetic number goes into a deck or a pricing decision.
Synthetic personas are a triage layer, not a verdict.
Use them to decide what to test. Let real people decide the price.
Coming nextThat's only the beginning. Deep Dives No. 02 is on the way. Subscribe to get it first, or keep reading for the full story and the evidence behind every number.
The old workhorse: Van Westendorp.
For years it has been the default for early-stage pricing research. Four questions, a clean chart, and an "optimal price point" where two lines cross. It is fast, cheap and easy to explain, which is why it is still where most pricing studies start, including ours.
We have used this survey on our own projects, and we have started testing AI personas alongside it. Because we want to offer the best to our team and our clients, we went to the research to see what the actual results are.
People can't price what they don't know
In an established category, respondents know roughly what comparable things cost. In a new category they don't, so the survey ends up asking them to make up numbers and then treating those numbers as signals. Better product descriptions don't fix awareness.
Nobody buys in isolation
Every real purchase happens against a consideration set: rivals, substitutes, the cost of complementary products, even doing nothing. Van Westendorp strips that out for neatness, so it measures price perception in a setting that doesn't exist.
Stated limits aren't checkout behaviour
The gap between "I would not pay more than X" and what people pay at the till is well documented and persistent. The industry has known this for decades and kept using the method anyway. So did we.
Why synthetic personas make it worse.
Ask a large language model (LLM) playing a persona "at what price would this seem too expensive?" and it will give you a confident, well-reasoned, sensible-looking number. Tested against reality, that number is usually wrong, in four specific ways.
AI answers are too similar to each other
This is the most important finding in the research, and the one firms selling synthetic-persona tools mention least. AI-generated answers sit far closer to the average than real people’s do. Aher, Arriaga and Kalai named it "hyper-accuracy distortion" in 2023, and study after study has confirmed it since.
What goes missing are the outliers: the people who'll pay a premium and the ones who abandon at the first sign of price friction. In pricing, those are exactly the respondents who shape your elasticity curve.
If a synthetic pricing study doesn't say how it dealt with this narrowing, ask.
Richer personas don't fix it. Actually, they can make it worse.
Adding demographic detail ("you are a 35-year-old mother of two with a household income of X") does not reliably improve pricing answers. For underrepresented groups it can actively degrade them.
Taday Morocho and colleagues8 tested two open-weight chat models against 70,000+ respondent-item instances from the World Values Survey. Persona prompting "does not yield a clear aggregate improvement" and "in many cases, significantly degrades performance". (Not yet peer-reviewed.)
Digital twins hit the same limit. The Columbia mega-study7 built twins from hundreds of each person's own survey answers, and overall accuracy did not improve compared with twins built from demographics alone. And at ACL 2026, Tan and colleagues13 found GPT-4.1 predicted subgroup preferences correctly 57.4% of the time, emulating young, male, Chinese and Christian personas more faithfully than others.
Personas still earn their keep on qualitative work, such as mapping a customer journey. The problem is quantitative inference.
The numbers depend on the model
The most-cited paper on synthetic pricing, Brand, Israeli and Ngwe's Using LLMs for Market Research3 (Harvard Business School, 2023, revised 2025), re-ran the same study on GPT-4o instead of GPT-3.5. Willingness to pay (WTP) came out at roughly a third of the original. Same prompt, same method, only the model changed.
If the number moves with the model, it is measuring the model, not the market. And once a model is deprecated, you cannot replicate last quarter's results, even with identical inputs.
Tigre and Souto add a second twist: reliability changes with the task. The same model tracked humans well on product pricing and poorly on classic thinking traps.
Validation is per-task, and it expires. You can't calibrate a synthetic panel once and treat it as a trusted instrument.
The experiment isn't an experiment
Almost nobody is talking about this one. In May 2026, a team from Google DeepMind and Carnegie Mellon11 showed that an experiment run on AI respondents looks like a controlled experiment, where only the price changes, but behaves like an observational study, where the people change too.
The model learned from data where people's traits shaped both what they saw and how they reacted. So when you change the treatment, such as the price, the synthetic respondent drifts into a different latent persona. They showed this across several model families, open and proprietary.
In pricing terms: the persona shown €19 and the persona shown €29 are not the same person. The gap you measure mixes a real price response with a hidden swap of people, and the output looks exactly like a clean result.
The fix has the same shape as before. Adjusting for generic demographics helps inconsistently and can make the bias worse. Adjusting for treatment-relevant traits, like budget constraint and category familiarity, actually helps.
- Budget
- Comfortable
- Knows the category
- Well
- Price sensitivity
- Low
- Budget
- Tighter
- Knows the category
- Barely
- Price sensitivity
- High
Put Van Westendorp into this environment and the problems compound.
The survey already asks people to invent prices in isolation. Hand that job to a synthetic panel and you stack four more problems on top:
- 01Answers sit too close together.
- 02Results get worse for subgroups.
- 03Numbers shift when the model changes.
- 04Personas drift when the price changes.
All to invent numbers for customers the model doesn't really understand.
Three years of evidence, on one timeline
- Apr 2023Brand, Israeli & Ngwe · Harvard Business SchoolWay forwardLLM choices in conjoint-style tasks can recover willingness to pay.
- 2023Aher, Arriaga & Kalai · ICMLWarningCoin "hyper-accuracy distortion": AI answers cluster too tightly.
- 2024Bisbee et al.WarningSame narrowing confirmed in survey responses.
- Oct 2025Brand, Israeli & Ngwe · revisionWarningRe-run on GPT-4o: willingness to pay roughly a third of the GPT-3.5 figure.
- Oct 2025Maier et al. · PyMC LabsWay forwardSemantic Similarity Rating restores realistic spread on purchase intent.
- Nov 2025Peng et al. · Columbia Business SchoolWarning19 studies, 2,000+ people versus their twins: 94% of twin responses under-dispersed.
- Feb 2026Taday Morocho et al. · IIT-CNR, Pisa, FlorenceWarningPersona prompting often degrades survey alignment, worst for underrepresented groups.
- Mar 2026Xuan, Hwang & LeeWay forwardTwins built from real review histories hit 87.73% on choice tasks.
- May 2026Lin et al. · Google DeepMind & Carnegie MellonWarningSimulated experiments behave like observational studies.
- May 2026Wang et al. · ConsumerSimBench, Shanghai Jiao TongWarningBest of 13 models reconstructs 47.8% of real consumer reactions.
- Jun 2026Tigre & Souto · Recast, DellWarning + fix~2× compression and 32% sign flips, plus a 50 to 300 person correction that removes most of the bias.
- 2026Tan et al. · ACL 2026Warning57.4% subgroup accuracy, skewed toward some groups. Fine-tuning can widen the gap.
- Sep 2026Wang, Hunt, Tang & Joseph · University at BuffaloWarningTwins reproduce the average effect while carrying almost no information about individuals.
Shift to methods that ask for choices.
Three survey designs share one idea: show concrete options at specific prices and watch what gets picked. Discrete choice shows several rounds of options at different prices and asks people to pick one each time, or none. Brand-price trade-off (BPTO) shows a set of brands at given prices; each time someone picks a brand, its price goes up and they choose again. Monadic testing shows each person a single price and asks whether they would buy, then compares groups who saw different prices. Nobody, human or synthetic, has to invent a number.
At what price would this seem too expensive?
Given these three, which would you pick?
The output changes too
A choice-based study doesn't hand you one optimal price. It gives you a substitution pattern. At what relative price does demand move to the alternative? How much price difference does a feature improvement justify? Where does a price-sensitive segment defect?
These are richer, more useful findings than "the optimal price is €27".
Rank order isn't automatically safe
It is tempting to say choice methods survive this narrowing because we care about ranking and crossovers, not exact magnitudes. The evidence doesn't support that.
When 32% of estimated effects point the wrong way, as Tigre and Souto found, the ranking itself is wrong. And in September 2026 a team at the University at Buffalo14 tested digital twins on twelve behavioural tasks. Twins often reproduced the average human effect while knowing essentially nothing about individuals. On some tasks they predicted the opposite direction.
So the honest claim is a narrow one. Choice methods hold up better than asking for a price because they depend less on the extremes. They are not immune. Treat the point where an AI panel switches to a rival as an early signal of direction, and test it with people.
What works: behaviour beats biography.
Read the evidence side by side and one pattern stands out. Accuracy improves when you feed the AI what people actually did, and barely moves when you describe who they are in more detail. It is the difference between a persona and a twin built from real behaviour.
How choice-based synthetic pricing works
Brand, Israeli and Ngwe don't ask the model for a price. They show it two product configurations at different attribute levels and prices, have it choose, and repeat across hundreds of trials. Willingness to pay for each feature is then worked out from the pattern of choices, just as it is in a human conjoint study (a survey where people repeatedly trade features off against price).
Asking the model to explain its reasoning before choosing (chain of thought) narrows, but doesn't close, the gap with humans. Their estimates lined up with separate human studies, including Fong et al.'s (2023) conjoint on household products. Fine-tuning on prior human surveys improved alignment for existing features and new features in the same category.
Xuan, Hwang and Lee9 went further. They built twins of active Reddit users from their full review histories, stored them in individual vector databases (a searchable memory for each person), and let each twin reason over its own past preferences using retrieval-augmented generation (RAG). On conjoint tasks they predicted real preferences with 87.73% accuracy, the highest figure in this literature.
Personas, and twins built from what people say about themselves
Twins built from what people actually did or chose
These studies use different tests and metrics, so the figures aren't directly comparable. The direction they point in is consistent.
The question to ask at the start of any synthetic pricing project is not "how good are our personas?" but "what real behavioural data can we attach to them?"
What we give up when we switch
Choice-based methods aren't a free upgrade. Here is the trade, side by side.
| Van Westendorp | Choice-based | |
|---|---|---|
| What it asks | Invent a price threshold | Pick between real options |
| Design effort | Low: four questions | High: competitive set, attributes, levels, balanced tasks, no impossible combinations |
| Finds prices outside the tested range | In theory, thanks to open answers | No. Only what you showed |
| Works in a brand-new category | Exploratory only, with humans | No. Needs real alternatives to choose between |
| Holds up on synthetic panels | Poorly | Better, not immune |
| Exposed to persona drift | Less: no price conditions to compare | More: every price condition is a chance to drift |
| What you get | One "optimal" price | Substitution patterns and crossover points |
So which method, when?
The shift we're making is narrower than "choice-based replaces Van Westendorp".
Choice-based becomes the default for synthetic pricing work in known categories. Van Westendorp moves to a smaller, exploratory role with humans only. Neither is treated as a verdict.
One more caveat from Brand et al.: fine-tuning does not improve alignment for new product categories. Step outside a category with prior human data and synthetic estimates become unreliable again.
Segments: the question clients and teams ask first.
Younger or older. Higher or lower income. Loyal or switching. Can synthetic choice methods tell us which segment drops out at which price? On synthetic data alone, not reliably.
Brand et al.3 are explicit that their approach does "not find similar improvements in the alignment... for differences between customer segments". Synthetic conjoint smooths across segments the same way it narrows answers overall. The neat segment-by-segment breakdown in the output is often an artefact of the method.
Tigre and Souto's10 example is stark. On the American National Election Studies (ANES) data, their synthetic panel showed a strong population-level correlation, while political Independents came in at about 0.02. No relationship at all. Nothing in the output told the two apart.
The Columbia study found the same asymmetry on another axis. Twins did better for higher-education, higher-income and politically moderate people, quietly amplifying the WEIRD skew (Western, Educated, Industrialised, Rich, Democratic) that market research has spent two decades trying to correct.
A healthy aggregate correlation tells you nothing about whether any individual segment in the study is real.
REBORRN's rule of thumb: test every segment separately.
The fix now has a method, and a price tag
Tigre and Souto split the error into two sources, give three diagnostics with published decision thresholds, and correct both at once with a statistical method (a doubly-robust estimator) fed by a small sample of real people.
The AI's idea of "the population" differs from your real customers or your target population.
The AI responds differently from real people, even when the people described are a match.
Why this matters commercially
"Anchor it in real data" now has a quotable cost: a few hundred human responses instead of a full second study. That makes mixing AI and human research something you can plan and budget, rather than a vague safety net.
It gives you a defensible "no". The diagnostics let you tell a team that a particular segment cut is not usable, with a threshold behind it instead of an opinion.
Run choice tasks across many price levels and personas matched to your real segment mix.
Shortlist the price range. Turn segment patterns into hypotheses, not findings.
Correct covariate and concept shift. Check each segment cut separately.
Know where you're sure, where you aren't, and why.
Respect the tails. The median lies.
Most synthetic pricing reports lead with the average willingness to pay or the median acceptable price, as if the median respondent were the market. That is wrong twice over.
First, real markets segment. There is a price-insensitive tail that will pay a premium and a price-sensitive group that leaves at the slightest increase. Where those two meet your cost structure is where pricing strategy lives. A single "optimal price" hides all of it.
Second, it is the part synthetic panels get least wrong. The median is what LLMs reproduce least badly. The extremes are what synthetic panels lose most. Leading with the median means leading with the part of the output that tells you least about the real market.
How to read the chart
- 1The black curve is what real people would pay. The blue curve is a synthetic panel asked the same question.
- 2The grey zones are the tails: walk-aways below €15 and premium payers above €35.
- 3Drag the slider under the chart. The median stays at €25, but the tails in the table below it empty out.
| Real market | Synthetic panel | |
|---|---|---|
| Median priceIdentical. Looks fine on a slide. | €25 | €25 |
| Would pay more than €35Your premium buyers | 4.8% | 0.04% |
| Won't pay more than €15Your first walk-aways | 4.8% | 0.04% |
How to recover the extremes
Oversample the edges
Run more personas than feels statistically necessary, weighted toward the unusually price-sensitive, the unusually loyal and the segments that might respond to a premium position.
Ask why
Have each persona explain its threshold. The reasoning often shows whether it has a real reference frame or is producing a plausible-sounding number.
Report the spread
Show the whole spread as well as the centre. It is harder to put on one slide than a clean number, and much closer to how people actually make price decisions.
What if the narrowing comes from how we ask?
In October 2025, Maier and colleagues at PyMC Labs published Semantic Similarity Rating (SSR). Their claim: narrow synthetic distributions aren't built into language models. Demand a number on a scale and the model regresses to the centre. So don't demand a number.
The persona describes how it feels about buying, in its own words.
Convert the answer into numbers that capture its meaning (an embedding).
Measure similarity to a set of anchor statements, one per scale point.
If this generalises, much of the pessimism here becomes a critique of prompt design rather than of the tool. It is the first serious argument that the narrowing can be fixed, and we think the direction is promising.
Why it stays a research direction for now:
- It was tested on purchase intent, using agree-to-disagree rating scales (Likert scales), not on willingness to pay.
- It doesn't fix demographics. Gender, region and ethnicity patterns are still unreliable.
- Results depend on anchor design and embedding choice, a version of the model-dependence problem one level down.
What this means in practice.
Four things we’re changing in how we run synthetic pricing work.
Treat synthetic as a triage layer, not a verdict
Use it to surface hypotheses, narrow the testing range and pick which questions deserve a human round. Never use it to set the launch price. The Buffalo team frames the real bar well: not "does it look human?" but "does running it measurably reduce our uncertainty about a specific human outcome?". On their twelve tasks that was rarely confirmed, so the point stays open.
Ask people to choose between options
Discrete choice, BPTO and monadic testing translate to synthetic work. Van Westendorp doesn't. When a team still wants the Van Westendorp chart, because they're used to it or need a specific number, run it with humans and use synthetic personas only for a choice-based supplement.
Report the full range of answers
"The optimal price is X" is the comfortable answer. The useful one: here's the shape of the demand curve, here's where we're confident, here's where we aren't, and here's the question we'd take to real people before committing. It is harder to fit on one slide, and it is the honest version.
Run diagnostics before you trust the output, and budget for them
Before a synthetic pricing result goes into a deck, it should clear these checks. Tick them off as you go.
0 of 4 checks cleared
The obvious bet, and why we don't fully take it.
The field moves fast enough that anything written here has a short half-life. The obvious bet is fine-tuning on real prior pricing data to close the gap between synthetic and human estimates.
It isn't always right. Tan et al. found cases where fine-tuning improved accuracy on previously unseen subgroups but widened the gap between subgroups on distance-aware measures: it raised the average by helping the groups the model was already good at more than the groups it was bad at.
| Are synthetic personas useful for pricing? | Yes, as a new instrument in the toolkit |
|---|---|
| A full replacement for human research? | No, not at the moment |
| Equally suited to every method? | No. Choice-based methods survive best |
The median is rarely the answer. The tails are where the strategy is.
The firms that earn client trust on pricing research over the next two years will be the ones willing to be specific about what synthetic methods can and can't do.
