Platform · Fact Checking
Check that AI tells the truth about you
We extract the verifiable claims and compare them with what your official site states.
The brands we measure every day
The metrics
They do not just cite you: they state facts about you
We extract the verifiable claims from the answers, compare them with your official sources and say which model is wrong, on which kind of claim, and where the error comes from.
-
The error rate is a risk metric
Across hundreds of answers analysed each period, 6% factual errors means one answer in sixteen carries a figure your own site contradicts.
599 answers analysed last 30 days6,3% errors- Verified 93,7%
- False 6,3%
-
Which model is wrong, and about what
Errors are not evenly spread: they concentrate on one assistant and one kind of claim. That is the figure that says where to act first.
Errors by model and claim type false claims · by type- Company
- Product
- Service
- Model A 16
- Model B 8
- Model C 8
- Model D 4
- AI Overview 1
-
Every error is a row, with prompt and model
The claim type is the key to priority: a product feature that does not exist weighs on the choice, an outdated company figure weighs on credibility.
Factual errors in detail prompt · type · model- “which product for use X” · product Model A 1 error
- “is there a version Y” · product Model C 1 error
- “who makes Z” · company Model B 1 error
-
The verdict comes with the authoritative source
Every error carries the model’s claim, the contradiction and the official URLs that disprove it: without the source it is an opinion, with the source it is a document you can take into review.
Error · source of truth verdict excerpt- Non-existent product attribute official site Wrong
- Outdated list price pricing page Outdated
- Stated coverage is partial network page Imprecise
-
Priority is spread weighted by decision impact
An error repeated by several assistants on a claim that weighs on the choice comes before an isolated inaccuracy. In regulated sectors the observed figure stays separate from the recommendation.
Claim × assistant errors foundProduct Price Coverage Policy Model A 62 30 18 10 Model B 12 44 8 0 Model C 24 16 34 6 Model D 8 12 10 0 -
The improvement can be demonstrated
New errors raise an alert, resolved ones stay in the history: the effect of a fix reads on the same prompt set, model by model.
Factual error rate last 12 months24% da 86%- Oct
- Jan
- Apr
- Sep
A wrong figure stated by an assistant weighs like an official statement: nobody in the company wrote it, everybody reads it.
Frequently asked questions
Taken verbatim from the monitored prompts. FAQPage schema.
Where does the source of truth come from?
From the official URLs you provide at setup, plus any documents you upload. Every verdict reports the sources it is built on.
Can it be used in pharma?
Yes, and it is the most requested use case, with a dedicated review workflow and traceability of the checks.
Can I report a correction to the model?
There is no reliable direct channel. You act on the sources the model cites.
How often does it run?
On the same cadence as the monitoring: new divergences raise an alert, resolved ones stay in the history.
Try it on your own data
Start the 7-day free trial yourself (card required, no call needed): see your real data before choosing a plan.
