Answer checking
What does ChatGPT say about your business, and is it true?
AI models answer questions about your business whether you like it or not. We ask several of them in fixed wording and compare every answer with your own facts. You get the sentences that are wrong, quoted.
Five questions, one model, about a minute. No account needed.
Our own run, 17 August 2026, four clinics in one city
- The number on file
- 0561 93725805
- Model without web search
- 0561 7393430
On the second run the same model produced a third number: 0561 870830. Neither belongs to that clinic.
in 2 of 2 runs
Run a check
Name and city are enough. A phone number and a website make the check sharper, because then there is something to compare against.
Six kinds of finding, six colours
Every finding has its own colour, and it stays the same everywhere: in the report, on the knowledge pages, in the email. What has no colour is not a difference.
- Wrong phone number in AI answersWhere does a phone number come from that we never had?
- Wrong website in AI answersWhy does the answer send people to an address we do not own?
- Wrong price in AI answersWhy does the answer quote a price we never charged?
- Invented service in AI answersWhy does the answer offer something we have never done?
- Wrong city in AI answersWhy does the answer place our business somewhere else?
- The model does not know your businessWhat does it mean when the answer knows nothing about us?
The damage never reports itself
“It’s not just whether you get named, it’s whether what it says about you is correct. The failure mode that quietly costs money is being recommended and then handed a phone number that’s been disconnected for two years. The customer never calls and you never find out.”
From a thread on AI recommendations for local businesses, 6 July 2026. Source
- Clinics given a wrong number or website
- 3 / 4
- Wrong numbers that changed on the second run
- 2 / 3
In our run on 17 August 2026 the model without web search gave a wrong phone number or website for three of four clinics. On the second run, two of those wrong numbers were different again. The model with live web search was right both times.
Two things follow from that: the problem is common, and a single query is not an answer. What counts is the frequency across several runs.
You could just ask the model yourself
To find out what a model says once, type the question yourself. It costs nothing and it is the right first step. From the second time on it gets tedious, and that is where the difference sits.
| Case | Asking yourself | Visibility tools | modelsays |
|---|---|---|---|
| What gets measured | one answer | whether your brand appears | whether the facts are right |
| Answers that vary | go unnoticed | partly covered | counted across runs |
| Work per month | by hand, every time | setup, then automatic | facts once, then automatic |
| Price per month | nothing | $29 to $495 | €0 or €12 |
What it costs
One check a month costs nothing. Watching costs twelve euros.
- Sample
- One check per month. Five questions, one model. Findings with the quote they came from.
- €0 / month
- Watch
- Weekly check. Three model families, one with live web search. Three runs per model, so the frequency means something. Email as soon as a new finding shows up.
- €12 / month
What this tool cannot do
- We reach models through their APIs. The ChatGPT app and the answer boxes in search engines can answer differently from the same model through its API.
- We change nothing inside a model. There is no button that deletes a wrong answer. What we deliver is the finding and a pointer to the likely source.
- Answers vary. A finding from one run out of three is a hint, not a fixed statement. That is why every finding carries the number of runs it came from.
- We check what you tell us. If your own facts hold an old phone number, the check will report the correct one as an error.