167 answers counted as mentions said something false about the brand
A mention rate asks one question of each answer: was the brand named? It doesn't ask what the answer said once it got there. We wanted to know how often those two things come apart.
The sample: answers from Proofsource workspaces between October 4 and 16, 2026, on ChatGPT, Perplexity, Claude and Google AI Overviews. For every answer that named a tracked brand, a checker compared each claim about the brand with facts the brand had confirmed or that we'd read from its own site. 2,137 answers, covering 80 brands in 63 workspaces, held at least one claim that could be checked against a fact on file.
167 of them, 7.8%, held at least one claim that contradicted the brand's own facts (95% interval 6.8% to 9.0%). Claim by claim, 200 of 4,761 checkable claims were wrong, 4.2%. Spread across brands, 36 of the 80 had at least one wrong answer in under two weeks.
Every one of those 167 answers counted toward a mention rate. On the dashboard they looked exactly like the answers that got everything right.
Most wrong claims were numbers: counts, prices and what the product is
We sorted the 200 wrong claims by what they were about, using a simple text rule, so treat the split as rough.
- Counts and other figures, 92: integrations, locations, customers served, a placement rate one point off.
- What the product is or does, 50: an estimating tool described as project management software, a security firm described as a blockchain project.
- Prices and plans, 48: a starting price quoted below the real one, credits per plan misstated.
- Dates, 10 by the rule, though only 6 were really founding years, most of them off by one.
Claude got a fact wrong in 12% of answers, Perplexity in 5%
By engine, the share of checked answers with at least one wrong claim was 12.3% on Claude (58 of 470), 8.7% on Google AI Overviews (45 of 516), 6.0% on ChatGPT (37 of 612) and 5.0% on Perplexity (27 of 539).
Don't read that as a league table. Each engine answered a different mix of questions in our sample, and an outside study with a different method ranked them differently (more on that below). What holds across all four is that none was clean.
Answers that cited the brand's own site were wrong half as often
On questions that named the brand, such as "what does this company do" or "how much does it cost", the answer cited the brand's own site 1,353 times. 7.1% of those got a fact wrong (96). When the brand's site wasn't among the citations, 15.3% did (47 of 307, interval 11.7% to 19.8%).
On buyer questions that didn't name the brand, there was no clear gap: 4.8% against 5.7%, on much smaller counts. This is a correlation, not a test. But it points at the cheapest fix there is: when an engine describes you, it gets you right more often if it is reading you.
176 wrong facts, 162 of them seen only once
Group the wrong claims by brand, fact and value, and the 200 collapse to 176 distinct wrong facts. 162 of those showed up in a single answer. Only 14 repeated, and only 8 appeared on more than one engine.
That changes how to act. A wrong price that one engine stated once is a draw you may never see again. One that keeps coming back, or appears on two engines, is coming from somewhere, and that somewhere is usually a page. We also found 65 wrong facts about 17 brands on pages the engines cited: third-party pages that already carried the error.
What to open behind your own mention rate
A tool that shows only the rate can't tell you any of this. Ask for the answers.
- Read the full answer text for every question where you were named, starting with the ones that name you most.
- Check each claim about you against your pricing, product and about pages. Prices, counts and founding dates go stale first.
- Check that the answer is about you. A company with a similar name can be counted as your mention.
- Open the cited pages. If your own site isn't among them, the engine is describing you from someone else's page.
- Count repeats before acting. Fix the claim that recurs across days or engines first, and find the page it comes from.
How Proofsource handles it
Proofsource keeps the full text and cited pages of every answer. For each answer that names your brand, it checks the claims about you against the facts you confirmed and the facts read from your own site, and keeps a wrong verdict only when the quoted claim and the correct value both appear word for word. The same wrong value seen again adds to one alert's count rather than raising a new one, and answers about a different company with a similar name are flagged instead of counted as yours.
Where this could be wrong
The checker is a language model reading against facts on file. We hand-read 25 of the 200 wrong claims: 21 were clear contradictions, and 4 were weak (a missing detail read as an error, a range, a fact list that disagreed with itself, and a claim about a different product). It also judges only claims that match a stored fact. Another 5,330 claims had no matching fact and weren't judged, so this misses errors about anything a brand never wrote down.
A brand's own facts can be out of date too, which would turn a correct answer into a false "wrong". The 80 brands are mostly small companies that signed up for Proofsource, not a random sample, and twelve days is a short window, which also limits how often a wrong fact had the chance to repeat.
An outside comparison: Searchable asked employees at 1,704 brands to grade 32,556 facts AI engines stated about them, and they flagged 9.2% as false, with pricing the worst category at 13.5%. Their graders could flag anything they knew to be wrong, including outdated details, which is broader than our fact-matched check and likely why their rate is higher. They also found Perplexity the least accurate of the three engines they compared, the opposite of our ordering.
