Can synthetic respondents measure willingness to pay?
By Lewsearch
Synthetic respondents can generate willingness-to-pay estimates, but a plausible number is not evidence that customers will pay it. Use synthetic pricing work to explore an offer and prepare a human study. Before using it to set prices, require validation of the price-response pattern and the resulting decision for the relevant product and audience. A concept-screening score alone does not establish demand, elasticity or an optimal price.
Start with the pricing decision
A team considering a price increase needs something different from a team improving how it explains value. Write down which job you are buying:
- Explore value perceptions: identify possible objections, confusing terms and benefits worth investigating with customers. Generated explanations are hypotheses to check.
- Estimate stated willingness to pay: measure what people say they would pay under a specified research design. Keep that outcome separate from purchases.
- Predict demand or choose a launch price: estimate buying behavior at different prices, with the alternatives and market conditions that matter to the decision.
These are different deliverables. A synthetic answer to “Would you buy this for $20?” does not become observed demand when it is multiplied by a market-size estimate.
Qualtrics' Gabor-Granger documentation describes a pricing design that asks willingness-to-pay questions at different prices and calculates demand and revenue curves. That describes how the survey analysis works. It does not establish that replacing its human inputs with generated answers preserves the result.
What current research establishes
In a September 2026 study, Sebastian Oetzel and Tobias Maiberger compared generated willingness to pay with purchase-linked human valuations for four inexpensive consumer products. They found differences in the distributions and pricing implications, even where an average looked close. The study used 387 campus participants and GPT-family models. Three synthetic prompts disclosed numerical price bounds that humans did not see; removing those bounds changed results inconsistently across products. It is a bounded evaluation, not a verdict on every provider or product. Read the study and its limitations.
The April 2026 revision of James Brand, Ayelet Israeli and Donald Ngwe's working paper reports both useful similarities and important WTP errors. Fine-tuning on relevant human survey data improved alignment within the same product category and population; those gains did not extend similarly to new categories or segment differences. It is a working paper, not a Lewsearch evaluation. Read the current paper.
The buying implication is specific: ask for evidence about the pricing task you need, rather than borrowing confidence from an unrelated benchmark.
Agree on a pricing-specific validation plan
The following is our proposed buyer checklist, not a certification standard.
- Specify the complete offer. Record the product, pack size or service period, currency, fees, delivery terms, competitors and purchase occasion. A monthly subscription and a one-time purchase are not interchangeable stimuli.
- Match the reference to the claim. Human survey comparisons can test agreement with stated preference. A claim about actual purchases needs suitable behavioral evidence or a properly designed purchase-linked study. Neither guarantees that future market conditions will stay the same.
- Keep evaluation outcomes separate. Agree which inputs may be used for calibration, then reserve an appropriate human comparison before tuning or selecting runs. Do not reveal the target price results and score the tuned output as an untouched test. The blind-validation guide covers the separation procedure.
- Compare across the relevant prices. Inspect the pattern and magnitude of acceptance, including important audience segments, rather than only the average maximum price. If the intended decision concerns a price increase, evaluate that change explicitly.
- Score the decision. Lock the candidate prices, analysis and tolerances before seeing the comparison. Ask whether the method changes the recommended price, whether that recommendation survives uncertainty, and what loss an incorrect choice could cause.
- Keep costs and constraints visible. Revenue, contribution and profit are different objectives. A revenue calculation that omits unit costs, capacity, channel effects or customer acquisition does not establish the most profitable price.
Record the exact instrument and starting-price logic. Pew Research Center's question-writing guidance explains why wording and preceding context matter in human surveys. That is a reason to preserve and inspect the design, not evidence that a synthetic method reproduces those human effects.
Illustrative example: a believable curve can recommend the wrong price
All figures below are invented. This is arithmetic, not a study, a customer result or measured Lewsearch performance. The two candidate prices are $10 and $20 for one unit of the same hypothetical product. Each price has an equally sized, separate group of 100 eligible prospects in each hypothetical source. There are no fees, repeat purchases or unit costs in this simplified example.
- Synthetic stated intent: 60 of 100 at $10 and 40 of 100 at $20. Multiplying price by those counts gives $600 and $800. That calculation favors $20.
- Hypothetical observed purchases: 80 of 100 at $10 and 25 of 100 at $20. The corresponding revenue is $800 and $500. That comparison favors $10.
Both patterns slope downward as price rises. The synthetic pattern looks sensible, but its recommendation differs from the hypothetical purchase evidence. Its $800 calculation is an intent-based scenario, not $800 of observed revenue.
The example does not show how frequently this happens, establish a correction factor or identify a profitable price. It shows why “the curve looks reasonable” is an incomplete acceptance test. A real evaluation also needs an appropriate design and uncertainty assessment.
What to request before commissioning the work
Ask the provider to name the intended output: exploratory reactions, stated WTP, an estimated price-response curve, or a behavioral forecast. Request the full price grid, instrument, audience definition, human reference and evaluation results, together with the model or service version, calibration inputs and exclusions. Keep human, synthetic and blended outputs distinguishable.
If the available evidence supports only exploratory work, use the result to improve the offer or human questionnaire. Do not choose a launch price from an unvalidated synthetic curve. A large generated file does not add observed customer purchases.
For a Lewsearch project, discuss the intended decision and agree whether pricing-specific validation is included before commissioning it. The concept-testing overview describes a stated-preference screen; it is not a claim that Lewsearch has validated a demand or price-optimization product.
FAQ
Is synthetic purchase intent the same as willingness to pay?
No. Intent at one stated price and a maximum acceptable price are different measurements. Neither is an observed purchase. Specify the measurement and evidence your decision requires.
Can I run a Van Westendorp or Gabor-Granger questionnaire with synthetic respondents?
Generating answers and running the calculation does not validate the resulting price recommendation. Evaluate the complete design against an appropriate human reference, including the prices and audience you intend to use.
Does a downward-sloping synthetic curve prove price sensitivity is accurate?
No. It can have the expected shape and still misstate the size of the response or favor a different price. Check the resulting decision, not only the shape.