Find out what AI tells your customers about you. Then fix it and measure the fix.
A diagnosis on its own is worth very little. This measures the same thing twice, before and after, so the fix is proven rather than promised.
Credited in full against a pilot if you go ahead, so the audit costs nothing if it convinces you. If the second report is not better than the first, you do not pay for either.
Loading payment options...
Pay now and send your site and ten questions, or ask first if you would rather talk. Either way the reports come back within two working days.
What actually happens
- 1You send a link and ten facts
Your website, and ten things a customer might ask that your own pages already answer. Half an hour of your time, once.
- 2We ask an AI, as it is today
Same ten questions, no tools, an explicit instruction that saying 'I don't know' beats guessing. Whatever it says comes from training data, which is what your customers get now.
- 3We put your site on the agent web
Your pages become an endpoint the model can ask. Nothing on your site changes.
- 4We ask the identical questions again
Same model, same day, same questions. The only thing that changed is whether it could check.
- 5You get both reports
Question by question: what it got right, what it got wrong, and every claim it made up. Plus the difference, which is the part that matters.
What it looked like on a real product
We ran it on RunBoth, a small Python tool. Asked ten factual questions with no endpoint, Google's model answered none correctly, got two wrong, and stated six things that were false, including that the tool uploads your source code to a third party. It does not make a single network call.
| Before | After | |
|---|---|---|
| Answered correctly | 0 of 10 | 9 of 10 |
| Answered wrongly | 2 | 0 |
| False claims made | 6 | 0 |
| Facts from the site stated | 0 | 33 |
gemini-2.5-flash, 15 September 2026. The method and every question are published on what AI says about your business, and the harness is open source, so you can disagree with it by running it.
Honest limits, before you pay
- This measures what a model says when asked, not your search ranking or your placement in AI overviews. Nobody has measured those and we do not sell them.
- One model and one prompt style per run. It is a measurement, not a law. We will run a second model on request.
- If your business is well known, the before report may already be good. We will tell you that rather than manufacture a problem, and you will not be charged.
- The after report depends on your own pages being able to answer. Where they cannot, you get the list of questions that came back empty, which is a content plan.
Ask any AI what your company does, then ask it where it got that. If it cannot tell you, it made it up, and so will the one your customer uses. That test is free and takes a minute.
Book an audit