Across 158,000 fact-checked AI claims, a brand's errors typically traced back to just four sources, and more than half of brands had a wrong claim that cited their own site. This is a fixable problem, if you look in the right place.
AI brand monitoring usually starts with a scare: somebody asks ChatGPT about your company and it says something wrong. The natural reaction is to assume the model is hallucinating and there is nothing to be done. The data says otherwise. Most inaccurate claims about a brand are not invented. They are lifted, faithfully, from a small number of pages that are themselves wrong, stale or ambiguous. Find those pages and you have found most of the problem.
Profound published the clearest evidence for this in August 2026, from 158,000 claims that AI engines made about brands and that were then fact-checked against the truth. Three findings from it change how monitoring should work.
The instinct is to look for the big offenders: the review site or the forum that keeps getting your facts wrong. Profound's data says that list is nearly useless. The ten most-cited earned-media domains accounted for only 13 percent of earned-media inaccuracies. The top hundred covered 40 percent. The rest was scattered across a long tail of pages that differ from one brand to the next.
Zoom in on a single brand, though, and the picture flips. For brands with inaccurate claims, the errors "typically come from just 4 websites," and those four are usually specific to that brand. So there is no industry-wide list to check. There is a short, personal list, and finding it means running your own prompts and reading your own citations.
This is the finding that should reorder the priorities. Fifty-four percent of brands had at least one inaccurate claim where the cited source was the brand's own content. That is slightly more than the 51 percent with an inaccurate claim citing earned media.
An engine citing your page and still getting the fact wrong usually means one of three things: the page is out of date, the page is ambiguous enough that a reasonable reader could get it wrong too, or two of your pages disagree. All three are within your control. None of them require outreach, PR or arguing with a forum moderator.
Pricing and billing claims were 12 percent of everything evaluated and 24 percent of everything inaccurate. Within earned media, pricing accounted for 35 percent of all errors. Among brands with at least ten pricing claims evaluated, pricing was the single most error-prone topic for 66 percent of them.
It is not hard to see why. Prices change, tiers get renamed, a promotion runs for a quarter and the page announcing it never comes down. A third-party review written in 2024 quotes the 2024 price. Your own pricing page uses "starting at" without saying what you start with. The engine takes the most confident-looking number it can find and reports it as fact.
ChatGPT and Claude overlapped on only 12.5 percent of the claims they returned about the same brands, and on just 11 percent of the inaccurate ones. Claude's inaccuracy rate was more than 1.3 times ChatGPT's, with the overall range across models sitting between 4.3 and 6.8 percent. Checking one engine tells you almost nothing about what another is saying. An AI brand monitoring routine that only watches ChatGPT is watching roughly one-eighth of the problem.
A related Profound study of 50,000 prompts across seven industries found one wearable manufacturer with about 11 percent of the claims made about it being false or inaccurate in a single week. Nearly half of every AI response in that study was editorial the person never asked for: comparisons, rankings, fee caveats, recommendations. The engines volunteer claims about you. They do not wait to be asked.
Not brand queries. The comparison, pricing, "is it worth it," "does it work with," and "how do I cancel" questions that come up in sales calls and support tickets. Profound's monitoring research found that the mix of prompts you track affects your numbers by an order of magnitude more than how often you run them, so this list is the foundation of everything after it.
Given a 12.5 percent overlap between ChatGPT and Claude, one engine is a sample of one. ChatGPT, Gemini, Perplexity and Claude cover the main behaviors; add Copilot or AI Overviews if your buyers live there. Capture the answer and every cited source, not just the mention.
This is the tedious part and the part that matters. For each false claim, note which cited page it came from. You will end up with a short list of pages, and if Profound's data holds for you, it will be about four. Sort by how often each page appears.
Because 54 percent of brands are their own worst source, and pricing is a quarter of all errors, the highest-return fix is usually a pricing page that states the number, the tier, the billing period and the date it was last updated, in plain text a machine can lift. Then reconcile any two pages of yours that disagree. Then retire the stale promotion.
Between June and July 2025, 40 to 60 percent of the domains cited for identical prompts changed. Running the same prompt ten times a day instead of once improved precision by only about 10 percent. So a monthly rerun of the same twenty questions is enough to see whether your fixes landed and whether a new source has appeared. Anything more frequent mostly measures the engine's own drift.
The 158,000-claim study is Profound's fact-checked dataset, which covers brands using their platform and the topics those brands chose to evaluate. It is large but it is not random, and "typically four websites" is a typical figure, not a promise about yours. The engine inaccuracy rates are for one window in 2026 and the models have shipped several updates since. Treat the shape of the findings as durable and the exact percentages as dated.
What is not in doubt is the direction of the fix. The old model of brand monitoring was watching for mentions. The useful model is tracing claims to sources and repairing the sources, which mostly means repairing your own pages. Running buyer prompts across the engines on a schedule and scoring the pages they cite is how that becomes routine rather than a one-off panic; it is what Prism Optimizer is built for, and it is also something a determined person with a spreadsheet can do once a month.
Either way, start with the pricing page. The data has been telling everyone the same thing.
Find your four.Run your buyer prompts across ChatGPT, Gemini, Perplexity and Claude, get every cited source, and see which of your own pages need fixing first.
Request a demo