AI makes mistakes about my brand: what to do, mistake by mistake
Five kinds of AI mistakes about a brand, each with its own fix path, plus a comparison of manual prompt checks, Brand Narrative and Fact Check for catching them.
"ChatGPT says we don't have an API." "Gemini thinks we're the company that got fined last year." "Perplexity calls us an enterprise tool, and our biggest segment is ten-person teams." If AI makes mistakes about your brand, the first urge is to treat them all as one problem called hallucination. They aren't one problem. An outdated price and a feature the model made up come from different places, and the fix for one does nothing for the other.
This post sorts the mistakes into five types, gives each its own fix path, and then compares three ways of catching them, from reading answers by hand to the claim and fact-checking tools in Promptwatch.
Quick verdict
Diagnose before you fix. Most wrong answers fall into one of five types: outdated, conflated with another company, invented, aimed at the wrong audience, or true but unflattering. Only the first three are factual errors in the strict sense, and only some of those start on your own site.
For catching them, manual prompt checks are fine for a handful of questions and fall apart at volume. Promptwatch's Brand Narrative, on every paid plan, is the practical middle ground: it lists the claims AI makes about you, grouped and scored, so you can sort them by type. Fact Check, on Business ($579 a month), Agency Scale ($799) and Enterprise, adds the part manual work is worst at, which is finding which of your own pages state the claim or the opposite.
Five types of mistake, five fix paths
| Mistake type | Example (hypothetical brand "Acme") | Where it usually starts | Fix path |
|---|---|---|---|
| Outdated | Quotes Acme's 2024 price | Old page on your site or a stale review | Update or retire the page, then confirm a re-crawl |
| Conflated | Mixes Acme Software with Acme Robotics | Similar names, thin "about" facts | State distinguishing facts plainly on your own pages |
| Invented | Says Acme integrates with a tool it doesn't | A gap your site never filled | Publish the real list, including what you don't support |
| Wrong audience | Calls Acme an enterprise suite | Old positioning, roundup categories | Rewrite positioning pages and fix third-party listings |
| True but unflattering | Says Acme's mobile app is limited | Real product feedback | Route to product, publish status honestly, don't "correct" it |
Outdated
This is the most common type and the easiest to fix, because there is usually a page behind it. An old pricing page that still resolves, an archived launch post, a comparison written before your last plan change. The engine isn't wrong about what that page says. The page is wrong about you now.
Find every page that still carries the old fact, on your site and off it. Update it, redirect it, or add a dated note that sends readers to the current version. Then wait for crawlers to fetch the new version before you judge anything. Our companion piece on AI citing wrong information about your brand walks through tracing the cited page and confirming the re-crawl.
Conflated
Here the model blends you with someone else: a company with a similar name, a former parent, a competitor in the same category. The giveaway is a claim that is true, just not about you. Your founder gets the other company's founder's history, or a lawsuit lands on the wrong brand.
The fix is mostly on your own pages. State the facts that tell you apart in plain text where a crawler can read them: what you sell, where you are based, when you started, who runs it. If the confusion is common, a short line saying you are not affiliated with the other company is worth more than it looks. In Promptwatch, brand aliases and the Brand Book's analysis instructions decide what counts as a mention of you, which keeps the other company's mentions out of your numbers while you work on it.
Invented
The model states a feature, integration, office or certification that has never existed. These are rarer than people assume and harder to trace, because often no page says it at all. The model filled a silence. Buyers ask "does Acme integrate with X," your integrations page doesn't mention X either way, and the answer gets guessed.
So the fix is to remove the silence. Publish the full list of what you support, and for the questions people actually ask, say clearly what you don't. If you run Fact Check on an invented claim and it finds no page on your site that states the claim or its opposite, that result is the diagnosis: your site has nothing to say on the subject.
Wrong audience or positioning
Nothing here is false in a way you could point to with a quote. The model just puts you in the wrong box. You get recommended to enterprises when you sell to small teams, or listed as a budget option after you moved upmarket. This one costs deals quietly, because the buyers who would fit never hear your name.
The sources tend to be old positioning copy and third-party roundups that filed you under a category years ago. Rewrite the pages that describe who you are for, make the comparison and use-case pages match, and ask the roundups that cite you to update the category. Brand Narrative is useful here because the claims show up under themes, and a claim like "best for large enterprises" repeated across several models is hard to miss once it's grouped.
True but unflattering
Some of what AI says about you is accurate and unwelcome. If your support queue really is slow, or the mobile app really is thin, the model is reporting what reviews and forums say. Treating that as a content problem wastes time and can look worse if you publish a page claiming otherwise.
Mark it as accurate, send it to whoever owns the product or the support team, and publish honest information when something changes: a changelog entry, a roadmap note, an updated help page. Then the claim has something new to pick up. In Fact Check terms, a claim you mark Accurate counts as accurate in the accuracy score, so an honest score separates "AI is wrong about us" from "AI is right and we don't like it." Our post on accuracy score, sentiment and theme scores covers how those numbers read together, so we won't repeat it here.
Three ways to catch the mistakes, compared
| Manual prompt checks | Brand Narrative | Fact Check | |
|---|---|---|---|
| What it catches | Whatever you happen to read | Individual claims from tracked responses, grouped by repeat and model | Whether pages crawled from your site state the claim or the opposite |
| Best for mistake types | Any, if you notice them | All five surface; sorted into themes | Outdated, invented, and contradictions on your own site |
| Judges truth? | You do | No, it scores tone per claim | It checks against your own pages, with quotes |
| Effort | High, grows with every prompt and engine | Low, you read a ranked list | Low per check; uses Agent Credits, estimate shown first |
| Weak spot | Small sample, answers vary run to run | Thin themes if your prompts never ask about them | Knows your site, not the wider web |
| Plan | None | Every paid Promptwatch plan, from Essential at $95/mo | Business $579/mo, Agency Scale $799/mo, Enterprise |
Source for the Promptwatch columns: Promptwatch llms.txt, ai-agent-instructions and changelog, read November 3 and 4, 2026.
Manual prompt checks
You open ChatGPT, Gemini and Perplexity, ask the questions a buyer would ask, and write down what's wrong. For a first look this is the right place to start, and it costs nothing. The trouble is sampling. The same prompt can give a different answer tomorrow, so one wrong answer may be rare or may be the norm, and you can't tell which from a single run. It also gets slow fast. Twenty prompts across four engines is eighty answers to read every time you check. Our fact-checking method is the manual version done properly.
Brand Narrative
Brand Narrative pulls the individual claims AI makes about you and your competitors out of tracked responses. It groups repeats, shows how often each claim appears and which models say it, and files claims under themes such as Pricing, Features, Support and Reputation. Each claim is scored positive, neutral, negative or mixed, each theme gets a score from 0 to 100, and you can compare themes side by side with competitors. The most negative claims sit at the top.
For this job, the grouping matters most. Sorting mistakes by type is quick when the claim "no API access" appears as one line with a count and a model list, rather than buried in forty answers. What it won't do is tell you whether a claim is true. A positive claim can be false. Our Brand Narrative versus sentiment comparison explains the difference in more detail.
Fact Check
Fact Check is the verification step. You mark a claim Accurate or Inaccurate, and it reads pages crawled from your site for ones that state the claim or the opposite, returning quotes and a suggested next step. Each check uses Agent Credits, and the estimate is shown before it runs. The accuracy score, added November 4, gives the share of your checked mentions that are accurate, as one number and as a heatmap per model and theme. Claims you marked count as marked. Otherwise the latest completed check decides, and inconclusive or unchecked claims are left out.
It is strongest on outdated claims, because it can find the old page you forgot about. It is also good for spotting contradictions, where one of your pages says the right thing and another says the wrong one. It can't judge a claim against the wider world. That still takes someone who knows the facts.
Who should pick which
- You have under twenty brand questions and one engine that matters: manual checks, on a fixed schedule, with a written list of what's true.
- You want to know which mistakes exist and how often: Brand Narrative. It is on every paid Promptwatch plan, so Essential at $95 a month covers it.
- Your mistakes are mostly outdated prices, plans or features: Fact Check, because it finds the pages on your own site that still carry the old version. That means Business, or Agency Scale for agencies.
- You are in a regulated category where a wrong claim is a compliance issue: Brand Narrative plus Fact Check, with the accuracy score as the number you report.
- Your problem is true but unflattering claims: no tool fixes this. Use Brand Narrative to show product and support teams what AI repeats, and how often.
For a wider look at tools that work on this problem, see our ranking of GEO tools for correcting what AI says about your brand.
Our recommendation
Sort first, fix second. Open the claims, tag each wrong one with its type, and send it down the matching path: outdated and invented claims to your own pages, conflation to your "about" facts, positioning to your messaging and the roundups, true but unflattering claims to the product team. Promptwatch is the one we'd evaluate first for this, because Brand Narrative gives you the sorted list on every plan and Fact Check on Business tells you which of your own pages are behind the factual errors. If wrong claims are already showing up in sales calls, start the 7-day free trial on Business and run Fact Check on the three most repeated negative claims first.