Asking about price: how to measure willingness to pay in a survey
Ask people what they would pay and you get a number that is too high and easy to steer. Why price questions mislead, how Van Westendorp, Gabor-Granger and choice tasks work, and how to read the results without fooling yourself.
By SurveyLane · The team building SurveyLane
You are about to launch a paid plan, so you add one question to your customer survey: "How much would you pay per month for this?" The median comes back at €40. You launch at €39 and conversion is half what you planned. The wording was fine. The format was the problem: an open question about money pushes answers up. A survey can still tell you a lot about price if the method fits the decision and you read the output as a range.
Why "how much would you pay?" gives you the wrong number
When nothing is at stake, saying yes is free. The gap between stated and real spending is called hypothetical bias. A meta-analysis by Murphy, Allen, Stevens and Weatherhead in Environmental and Resource Economics (2005) pooled 28 studies that compared hypothetical and real payments collected with the same method. The median hypothetical answer was 1.35 times the real one, with a long right tail: in some studies the stated amount was several times what people paid.
Most of those studies valued environmental and public goods. The direction still carries over. A respondent who isn't entering a card number is being agreeable, or signalling that they like you.
The open box makes it worse. Most people have never thought about the question, so some type a competitor's price and others a round number, and the median averages guesses made for different reasons.
Price answers are built on the spot
People don't carry a fixed willingness to pay around. They construct one when asked, from whatever is nearby. In a 2003 paper in the Quarterly Journal of Economics, Dan Ariely, George Loewenstein and Drazen Prelec had participants write down the last two digits of their social security number and say whether they would pay that many dollars for an everyday product, such as a bottle of wine. Then they named their maximum price. Higher digits produced higher prices. Each person's valuations of related products still lined up neatly, which is why anchored data looks trustworthy.
Survey anchors are quieter: a price in an earlier question, examples like "€10, €20 or €50", a competitor named three screens back. Treat everything before a price question as part of it. The post on question order explains the general mechanism; for price it is strong enough to plan around.
Start with the decision the data has to support
Write down what you will do with the answer before you pick a method. "Is €39 in a sane range?" is a sanity check, and a perception method is enough. "Which of three prices earns the most?" is a demand question, so you have to ask whether people would buy. "Does API access belong in Pro or Business?" is a trade-off, and price is one attribute among several.
The usual mistake is to ask a perception question and use it as a demand forecast. The post on measuring priorities makes a similar point about features.
The Van Westendorp price sensitivity meter
The best-known perception method was presented by Dutch economist Peter van Westendorp at the ESOMAR congress in 1976. It asks four open questions about one product:
- At what price would this be so expensive that you would not consider buying it?
- At what price would it be getting expensive, but still worth considering?
- At what price would it be a bargain, a great buy for the money?
- At what price would it be so cheap that you would doubt the quality?
Plot the cumulative share of answers to each question against price. Where "too cheap" crosses "too expensive" is the optimal price point; where "bargain" crosses "getting expensive" is the indifference price point. The outer crossings, of "too cheap" against the inverse of "bargain" and "too expensive" against the inverse of "getting expensive", usually mark the acceptable range.
Put a short product description, billing unit included, right above the questions. Drop anyone whose "too cheap" price sits above their "too expensive" one. And use numeric fields with the currency shown, so you aren't parsing "about 30 bucks".
Where Van Westendorp misleads
The method measures perception. Someone can call €39 a bargain and never subscribe, because they don't need the product.
In 1993, Newton, Miller and Smith added two purchase-likelihood questions, asked at each respondent's own "bargain" and "getting expensive" prices. That lets you weight the curves by likelihood to buy and estimate reach or revenue. If you use Van Westendorp for more than a sanity check, add those two questions.
The product is also judged in isolation, while real buyers compare. Current customers anchor on what they pay now. So treat the optimal price point as the middle of a reasonable zone. It is too soft to print on your pricing page.
Gabor-Granger: asking about buying at specific prices
The Gabor-Granger method, developed in the 1960s by economists André Gabor and Clive Granger, asks about buying directly. Show a price, ask whether the respondent would buy at it, then show a higher or lower price depending on the answer until you find the highest yes.
Across respondents you get a demand curve. Multiply each price by its share of yeses for a revenue curve, which usually peaks below the price that maximises buyers.
You build it with conditional logic: each answer branches up or down. The first price is an anchor, so randomise where the ladder starts. People also learn that no brings a lower price, and it becomes a negotiation. Four to six prices is plenty.
In a 1966 paper in Economica, Gabor and Granger also showed that price works as a signal of quality: below some point, cheaper makes buyers suspect a worse product. In a ladder that can appear as a dip in demand at the lowest prices. It is the same effect Van Westendorp's "too cheap" question tries to catch.
Choice-based questions get closer to behaviour
Real buying is a choice: your plan, a competitor's plan, or nothing at all. Questions that copy that structure tend to get more realistic answers. The Murphy meta-analysis found that choice-based elicitation mattered for reducing hypothetical bias, although too few studies used each format to rank them.
The simple version is one choice task: two or three packages side by side with prices and features, plus "none of these". Vary prices across respondents and watch choices shift. The full version is choice-based conjoint, where each respondent sees a series of such tasks built from an experimental design and a model estimates how much each attribute, price included, drives choice. It needs more respondents and specialist analysis. For packaging decisions it is worth that.
Always keep the "none" option. Without it everyone is forced to buy, and your simulated market share is fiction.
When purchase intent can be trusted
"How likely are you to buy this at €39?" on a five-point scale is cheap and common. In a 2007 paper in the International Journal of Forecasting, Vicki Morwitz, Joel Steckel and Alok Gupta found that stated intentions predict sales better for existing products than new ones, for durables than non-durables, over short horizons than long ones, for a specific brand than a category, and for trial than total sales. Intent asked in a comparative format also predicted better.
For a software team that means asking about a specific plan and about the next three months. Existing users weighing an upgrade give you more to go on than strangers imagining a product. And never count "probably will buy" as a sale.
Wording a price question so it measures what you think
Compare these two:
"Would you pay €39 for SurveyLane Pro?"
"SurveyLane Pro costs €39 per user per month, billed annually, excluding VAT. It includes unlimited surveys, conditional logic and AI analysis of open answers. If this plan were available today, would you subscribe?"
The first leaves people to guess the billing period, the unit, the tax and what they actually get.
Avoid "fair" and "reasonable" too. Those words measure expectations. And don't put a price question straight after a list of benefits, because that reads as a sales pitch.
You can also tackle hypothetical bias directly. In the American Economic Review in 1999, Ronald Cummings and Laura Taylor tested a "cheap talk" script: a short explanation that people tend to state higher amounts in surveys than they pay, with a request to answer as if the money were real. In their experiments it brought hypothetical answers close to real ones; later studies found it less reliable with experienced buyers. It costs one paragraph, so pretest it on your audience.
Who you ask matters as much as how
Current customers already decided you were worth the old price. Free users are often there because it is free. And the prospects who never signed up are the hardest to reach, though their answers may matter most. Survey only newsletter openers and you measure your keenest fans; the post on sample quality explains why a big response count doesn't fix that.
Plan the analysis by segment first. Company size and plan often explain more variation than anything else, and each segment's curve needs enough respondents of its own. The post on how many respondents you need shows the calculation.
Reading the results, and where AI fits
A price survey produces a range with uncertainty. An optimal price point of €37.50 from 150 respondents implies precision you don't have; a handful of answers can move a crossing by several euros. Report the range, the base and how much it shifts by segment. Look at distributions too. A median hides a split between people happy at €20 and people happy at €80, and that split may mean two plans. The post on reporting results covers showing spread honestly.
Add a short open follow-up such as "What made you choose that amount?" The reasons tell you whether resistance comes from budget, missing features or a competitor. AI-assisted analysis is useful here, because it can group hundreds of reasons by segment. The post on open answers shows how to check the themes are real, and the MCP post how to query your own data. It can't fix the number, though. A neat summary of inflated prices is still inflated.
Then test with money on the table: a pricing page experiment, or a limited offer to some new sign-ups. The survey tells you which two or three prices deserve that test.
Frequently asked questions
How many respondents do I need for a Van Westendorp analysis?
There is no fixed minimum, but crossing points are unstable in small samples because a few answers can move a curve. What matters more than the total is having enough answers in each group you will compare. A practical test is to split your data into random halves and see how far the crossing points move; if they jump by several euros, collect more responses before you trust the range.
Should I ask existing customers or prospects about price?
Both, analysed separately. Existing customers are anchored on their current price, so they tell you about upgrades and increases. Prospects tell you about acquisition, with more hypothetical bias. Mix them and the curve describes neither group.
Is Van Westendorp or Gabor-Granger better?
They answer different questions. Van Westendorp shows how prices are perceived and helps early, when you have no idea where the price should sit. Gabor-Granger estimates how many would buy at specific prices, which helps when you are choosing between a few candidates. For which features go in which plan, use a choice-based format such as conjoint.
Can I add a price question to a customer satisfaction survey?
You can, but it affects the answers around it: later questions turn more critical of value, and earlier satisfaction questions can push the price answer up. Put it near the end, keep it away from your trend questions, and check for a shift before comparing that wave with earlier ones.