W
Reviewed by Jacob Whitmore, Whito · Fact-checked for accuracy
Links marked (Ad) in this article are commission-based, meaning we earn a small fee if you purchase through them, at no extra cost to you. Full disclosure.

Last Updated on November 21, 2026

Paste what people said when you asked “Would you buy this?” The tool reads each answer and turns the words into a 1 to 5 purchase intent score. No sign-up, and nothing you paste leaves your browser.

No answers yet? Get practice answers from an AI chatbot

Fill these in, copy the prompt into ChatGPT, Claude or Gemini, then paste the replies above and tick the practice box. Use it to sharpen the offer before you ask real people, not instead of asking them.

Method checked against the published paper on 21 November 2026.

“Sounds great” is not a sale

Most small businesses test an idea by asking a few people what they think. People are kind. They say it sounds great. Then nobody buys.

The mistake is asking for a number or a yes. Ask “Would you buy this, and why?” and let people answer in their own words. The words carry the doubt that a polite “yes” hides: “maybe”, “if it were cheaper”, “I’d wait for reviews”.

This tool reads those words for you. We call it the say-it-first score: people say it in their own words first, then the words are scored.

How the score works

  1. Each answer is turned into numbers. A small language model running in your browser converts every answer into a list of numbers that captures its meaning.
  2. It is compared with 30 reference statements. Six sets of five, from “I would definitely not buy this” to “I would definitely buy this”. The closer an answer sits to a statement, the more weight that rating gets.
  3. The weights become a 1 to 5 spread. Each answer gets a share of every rating rather than one hard number. The tool averages the six sets, then averages every answer into the overall score.

The method comes from a 2025 study by PyMC Labs and Colgate-Palmolive, “LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings”. The authors tested it on 57 product surveys with 9,300 real people. Scored this way, AI-written answers matched real survey results far better than asking an AI for a number directly.

What we changed, and what that means for you

  • A smaller model. The study used OpenAI’s text-embedding-3-small. This tool uses bge-small-en-v1.5, a free model small enough to run on your device, so your answers stay private. We tried three small models on a labelled set of 14 answers, from a firm no to a firm yes. This one tracked the labels most closely and was the only one that reliably scored a plain “No, I would not buy this” as a no.
  • Sharper weighting. The study used a temperature of 1. We use 0.5, which pushes each answer’s weight towards its nearest ratings. With a small model, a temperature of 1 bunched every score near the middle.
  • Our own reference statements. The study’s statements were tuned to personal care products and are not published in full. Ours are written for everyday UK products and services. We have not tested them against a human panel.
  • Real answers first. The study scored AI-written answers. You can do that here, but the tool is most useful on what real customers told you.

So treat the score as a fast, consistent read of open answers. It is not a sales forecast.

How to get answers worth scoring

  • Ask at least 20 people who could actually buy it. Not friends, unless they are the customer.
  • Show the real price. Interest without a price tells you very little.
  • Ask one open question: “Would you buy this, and why or why not?” Do not suggest an answer.
  • Paste the answers as written. Do not tidy them up. The hedges are the useful part.
  • Change one thing at a time when you compare two offers: the price, the guarantee or the wording.

Reading the result

  • Above 3.8: strong. Most answers read like people who would pay. Now test with a real way to pay, such as a deposit or pre-order.
  • 3.3 to 3.8: warm but not sold. Read the “no” answers. The objection is usually price, trust or timing.
  • 2.7 to 3.3: on the fence. Fix the offer before you spend on ads.
  • Below 2.7: cold. The offer, the price or the audience is wrong.

Scores rarely reach 1 or 5. Every answer shares some weight across neighbouring ratings, so results cluster in the middle. Compare offers with each other rather than chasing a 5.

Limits

  • Small models sometimes miss negatives and sarcasm. Open “Every answer and its score” and check any that look wrong.
  • AI practice answers reflect what the model learned, not your local market. The study found they matched real people on age and income but not on gender or region.
  • A score is what people say. Paying is different. The only proof of demand is money changing hands.
  • The model downloads once from Hugging Face (about 34 MB) and runs on your device. Very old phones may be slow.

Questions

Does Whito see the answers I paste?

No. The model runs in your browser. The answers are never sent to Whito or anyone else.

Can I use AI answers instead of asking customers?

As a rehearsal, yes. They help you spot weak wording or an obvious price objection. They cannot tell you whether people in your area will pay.

How many answers do I need?

At least 20 per offer. Below that, one strong opinion can move the score a lot. The tool warns you when you have fewer.

The takeaway: do not ask people to rate your idea. Ask them why they would or would not buy it, then read the reasons. The score is a shortcut to those reasons, not a replacement for a first sale.

Whito logo
Whito
Whito is an independent UK research publisher. We publish research on how UK businesses are found, trusted and priced: what they cost to market, how their sectors work, and who actually gets recommended. Studies are free to read, every figure carries the date it was checked, and nobody pays to appear.