AI says something wrong about your brand: what to check
A field-notes checklist for when an AI answer claims your product is too expensive, missing a feature or aimed at the wrong buyer, and you need to know whether the claim is real, repeated or your own fault.
Someone on the team pastes a screenshot into chat. An AI answer says your product is "too expensive for small teams", or that "its mobile app lacks features", or that you only serve enterprise. Nobody agrees on whether it's true. Half the room wants to email the vendor, the other half wants to ignore it.
Both reactions skip the diagnosis. A single answer is an anecdote. What matters is whether the claim repeats, which engines say it, and whether something you published is feeding it. Here is the order we work in.
Step 1: separate the claim from the answer
An AI answer is a bundle of statements. "Acme is a solid CRM, but it's pricey for small teams and the mobile app is limited" holds three claims: category fit, price positioning and app quality. Treat each one on its own.
Write each claim as a short sentence you could mark true or false. Vague ones ("has a mixed reputation") go in a separate pile, because you can't verify them, only influence them.
This sounds fussy. It isn't. If you treat the whole answer as "negative", you'll fix nothing, because nobody can act on "negative". They can act on "the engine says we have no free tier".
Step 2: check whether it repeats
Ask the same question ten times, spread over a few days, with search turned on. Vary the wording a little: a brand question, a comparison, a "best for small teams" question. Log which claims appear and how often.
Three outcomes are common:
- The claim shows up once and never again. Noise. Move on.
- It shows up in one engine only. That engine is probably leaning on a specific source.
- It shows up everywhere, with similar wording. Something upstream is feeding it, and that is worth tracing.
Wording that matches across engines is the tell. Models don't coincidentally pick the same phrase. They are reading the same pages.
Step 3: look at your own site first
This is the step people skip, because it's uncomfortable. Search your own site for the claim and for its opposite.
Old pricing pages are the usual culprit. So are a retired plan name in a help article, a blog post from two years ago that says "we don't support X yet", a comparison page written for a campaign nobody remembers, and a PDF that was never taken down. If your site states the thing the AI says, the AI isn't wrong about what you published. It's wrong about what's true today.
If your site says the opposite, or says nothing, the source is likely elsewhere.
Step 4: trace the outside sources
Open the citations behind the answers that repeat the claim. Look for review sites, forum threads, listicles and competitor comparison pages. A comparison page written by a rival will happily call you expensive. A Reddit thread from 2024 can still be the best match for a question.
You can't edit those pages, but you can respond: publish a clear, dated page that states the current facts, correct what you control (your profiles, your listings), and where a third-party page is plainly out of date, ask the owner to update it. We're careful with that last one. It works sometimes, and a polite note costs nothing.
If you want the general version of this routine, our post on when AI engines get your facts wrong goes through it for factual errors like founding dates and locations. The fact-check walkthrough covers the manual prompt panel. And if the problem is that you aren't named at all, use the ChatGPT diagnostic checklist instead.
Step 5: decide which claims deserve effort
Not every claim matters. Rank them by two questions. Does it show up in the prompts your buyers actually use? And is it the kind of claim that ends a shortlist? "Expensive for small teams" ends a shortlist for a small-team buyer. "Founded in 2019" doesn't.
Then pick the true ones and the false ones separately. A true claim you dislike (you really don't have a mobile app) is a product conversation, not a content fix. A false claim is a content and sources problem. Mixing them up is how teams spend a quarter rewriting pages to hide something real.
Where the manual routine breaks
The routine above works for one brand and a handful of claims. It falls apart when you need it weekly, across several engines, against competitors. You end up with a spreadsheet of pasted answers and an argument about whether this week's wording counts as the same claim as last week's.
That is the point where we use Promptwatch, and it's the one recommendation we'd make for this problem. Two pages under Brand perception, launched October 30 per the Promptwatch changelog, map directly onto the steps above.
Brand Narrative does steps 1 and 2 for you. It pulls the individual claims out of every tracked response, groups repeats, and shows how often each claim appears and which models say it. Each claim is filed under a theme such as Pricing, Features, Support or Reputation and scored positive, neutral, negative or mixed. Each theme rolls up to a score from 0 to 100 that you can compare against competitors, and it shows sentiment per competitor, so you can see where an engine talks about a rival more kindly than about you. The changelog lists it as available on every plan.
Fact Check does step 3. You mark a claim Accurate or Inaccurate, and it reads the pages crawled from your site to find ones that state the claim or the opposite. Each check returns the pages, the quotes and a recommended next step. It's available on Business, Agency Scale and Enterprise, and each check uses agent credits, with the estimate shown before it runs. On the plans below that, you'd do step 3 by hand, which is fine for a small site and tedious for a large one.
What it doesn't do is step 4. Tracing outside sources and deciding whether to approach a publisher is still judgment, though the citation reports in the same platform show you which pages the answers lean on.
A note on expectations
Fixing a wrong claim is slow. Engines re-read pages on their own schedule, and a correction can take weeks to show up in answers, if it does. We wouldn't promise a client a date. We'd promise a clean diagnosis: here is the claim, here is how often it appears, here is the page on your site that contradicts or supports it, and here is the next action.
That framing also keeps the internal debate short. Instead of "AI hates us", you get "the pricing claim appears in a share of tracked answers, and our own 2024 pricing article states it". People can act on that sentence.