AI makes mistakes about my brand: what to do with each kind

When a client tells us "AI makes mistakes about my brand", they usually mean one answer somebody screenshotted last week. Fair enough. But "mistake" covers several different problems, and they don't share a cause or a fix. A retired plan quoted as current is not the same as being mixed up with a company in another country. Both differ again from an engine crediting you with an integration you never built.

We've already written the checklist for working through a single wrong claim: what to check when AI says something wrong about your brand. This post sits one step before it. Before you trace anything, decide what kind of mistake you're looking at. The type tells you where to look and who has to act. It also sets how long you should expect to wait.

Two ways an engine knows about you

Some engines search the web while they answer. ChatGPT Search, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Microsoft Copilot work this way. When they get you wrong, it's often because they retrieved a page that says the wrong thing, or found the right page and blended it with another one.

Other models answer from what they learned in training. Promptwatch tracks Claude, DeepSeek, Mistral and Grok through their APIs, which shows what those models say from training data rather than from a live search. A model answering that way can't read the pricing page you fixed this morning. It repeats what it absorbed. Where it absorbed nothing, it may fill the gap with whatever sounds likely.

So one wrong claim can mean "a page is stale" in one engine and "the model's memory is stale" in another. Which engines repeat a mistake is your first clue to its type.

The five types

1. Outdated

The claim was true once. A plan you retired, an old price, a previous office address, a product you renamed.

The cause is usually retrieval. An archived blog post, a help article, a PDF nobody took down, or a directory listing from launch year is still the best match for the question. In models answering from training data, it can also be plain memory from before the change.

The fix is to find the page that still states the old fact. Update it, or add a dated note saying what changed, and redirect retired pages to their replacement. The citation side of that hunt has its own post: AI is citing wrong information about my brand. Expect live-search engines to correct first. API-tracked models update on a schedule nobody publishes.

2. Conflated

The engine merges you with a company that has a similar name, a former name, or the same acronym. The tell is detail that is true, just not about you: a founding city you've never been to, a product line you don't sell, a CEO who works elsewhere.

This is entity confusion. The engine hasn't settled which "Acme" the question means, so it pulls from both. Our explainer on entity resolution covers the mechanics.

The fix is to make yourself easy to tell apart. Your About page should say who you are, where you're based and what you sell, in plain sentences. If the confusion is common, say outright that you're not related to the other company. Use one consistent description across every profile you control. On the measurement side, set up brand aliases and Brand Book analysis instructions so the other company's mentions aren't counted as yours. Otherwise every later number is polluted.

3. Invented

A feature, integration, certification or plan that never existed. "Acme integrates with Salesforce." It doesn't, and never did.

This one is gap filling. When nothing on the web answers a specific question about you, a model may supply the answer that's typical for a product like yours. We see it most on integration and compliance questions, because competitors list theirs and you said nothing either way.

You can't correct a page that doesn't exist, so the fix is to write one. A plain list of what you integrate with. A short "what we don't do" section. A security page stating which certifications you hold and which you don't. Silence is what the model filled in. If the invented feature is actually on your roadmap, say that honestly rather than denying it.

4. Wrong audience or positioning

Every individual fact may be correct while the framing is off. "Best for enterprise" when most customers are small teams. "A budget option" when you're priced mid-market.

The cause is usually the source mix. Rival comparison pages and old listicles describe you in their own terms, and so does the review-site category you were filed under at launch. Engines summarize who you're for from whoever wrote about it most.

This fix is slower. Your own product and pricing pages need one quotable sentence about who you're for. Then work outward: correct your directory categories and profiles, and ask authors of stale roundups for an update. Our post on offsite mentions explains how we choose which pages are worth the email.

5. True but unflattering

"Support is slow to respond." "The mobile app is limited." And it's accurate.

The engine isn't making a mistake here. It's reading real reviews and threads correctly. We list it as a type because teams keep filing it under "AI got us wrong" and then spend weeks on content that can't fix it.

The fix is a product fix. Once the product changes, document the change on a dated page, so newer evidence exists to compete with older complaints. Until then, take the claim out of the content queue.

Sorting a mixed pile

In practice you never get one tidy type. You get twenty claims across six engines. We sort them like this:

1. Group the claims by theme: pricing, features, support, who it's for.
2. For each claim, note which engines say it: only live-search engines, only API-tracked models, or both.
3. Ask whether it was ever true. True and no longer: outdated. True about someone else: conflated. Never true: invented. True but framed wrong: positioning. True today: unflattering.

One pattern holds up well enough to use as a starting guess. An error that appears only in API-tracked models, with no page on the web stating it, is usually memory or gap filling. An error that live-search engines repeat in similar wording usually has a page behind it.

How we run this in Promptwatch

Doing the sort by hand works for one audit. Doing it every month across a client's competitors is where we use Promptwatch, and it's the platform we'd recommend for this job.

Brand Narrative, available on every plan, pulls the individual claims AI makes about you and your competitors out of tracked responses and groups the repeats, with how often each claim appears and which models say it. Claims are filed under themes such as Pricing, Features, Support and Reputation, each one scored positive, neutral, negative or mixed, and each theme gets a score from 0 to 100 you can compare with competitors. The "which models" column is what makes the live-search versus API split practical instead of theoretical.

Fact Check is on Business ($579 a month), Agency Scale ($799 a month) and Enterprise. It uses Agent Credits and shows an estimate before it runs. You mark a claim Accurate or Inaccurate, and it reads the pages crawled from your site for ones that state the claim or its opposite, returning quotes and a recommended next step. For an outdated claim it can turn up your own old page. For an invented claim it may find nothing, and that's useful too: no page on your site settles the question, so you know to publish one.

The newer piece, added in the November 4, 2026 changelog, is the accuracy score. It's the share of your checked mentions that are accurate, shown as one score and as a heatmap per model and per theme. A claim you marked yourself counts the way you marked it. Otherwise the latest completed fact check decides. Inconclusive and unchecked claims are left out. Brand Narrative now also lists the best-scoring theme first and puts the most negative claims at the top.

The heatmap is where classification gets faster. These readings are our own heuristics, not Promptwatch's. Low accuracy on Pricing across the live-search engines points at outdated pages. Low accuracy on Features in the API-tracked models only points at memory or invention, so publish the missing page and be patient. Low accuracy on one theme in every engine at once is worth checking for conflation.

Two cautions. The score covers only checked claims, so a high score on a handful of checks tells you little; report the number of checked claims next to it. And because inconclusive claims drop out, a theme full of vague, unverifiable statements can look healthier than it is.

On Essential, Professional, Kick-off or Growth there's no Fact Check. Brand Narrative still does the grouping, and you mark claims by hand against your own source of truth, which our fact-check walkthrough describes.

Next step

Take the last month of wrong answers about your brand and sort them into the five types before you fix anything. If you'd rather have the claims grouped and scored per model for you, start Promptwatch's 7-day free trial and open Brand Narrative first.