AI makes mistakes about my brand: what to do, by type of mistake
AI makes mistakes about my brand: what to do depends on the kind of mistake. Our intake process, five mistake types, and the fix path we use for each.
The ticket usually reads something like "ChatGPT says we don't integrate with Salesforce. We do." Sometimes it's a price, sometimes a competitor's name, sometimes a line about bad support that stings because it's partly true. The client wants it gone by Friday.
We don't start by rewriting anything. A wrong answer about a brand can come from at least five different places, and each one has a different fix and a different owner. Rewriting the homepage when the problem is a directory listing from 2023 burns a week and changes nothing. So the first job is to work out which kind of mistake you're looking at.
How we take the ticket
We ask for four things before anyone touches the site: the exact prompt, the engine, the date, and the full response text. A screenshot is fine if that's all there is, but we want the cited sources too, because they're often the fastest route to the cause.
Then we ask the client for the truth in one sentence. "We integrate with Salesforce through a native app, on every plan" is a target. "That's wrong" isn't. If nobody can write that sentence, the site probably can't either, and that is often the actual problem.
Last, we check whether it's one answer or a pattern. We run the prompt again, plus a handful of variations, across the engines the client cares about. One odd answer on one model is a note in the log. The same claim showing up across several engines and prompt phrasings is a work item. Our claim-by-claim audit process covers how we build that prompt set.
The five kinds of mistake
Outdated: the answer was true once
A retired plan, a price from two pricing changes ago, a product that was folded into another one, an old office address. These are the most common mistakes we see and the easiest to fix, because the model isn't inventing anything. It is reading a page that still says the old thing.
The source is often the client's own site: an old pricing page that was unlinked but never removed, a launch blog post, a help article written before the change, or a press release. Off the site, it's usually directory listings and review profiles nobody updated.
Fix path: find every page that states the old fact, update or redirect it, and state the current fact plainly near the top of the page that should own it. If the site uses Product or Offer structured data, make sure it matches the visible price. Then go through the listings the client controls (review sites, marketplaces, partner directories) and update them in one pass. This is ordinary SEO hygiene, and it pays twice.
Conflated: the model mixed you up with someone else
Two companies with similar names, a brand that shares a name with a consumer product, a client that was acquired and now shares facts with its parent. The answer stitches attributes from both together.
Fix path: give the model a clearer entity to hold on to. A plain about page that says what the company is, where it's based, and what it does not make. Organization structured data with the legal name and sameAs links to the official profiles. Consistent naming across every listing. If the confusion is with a former name, say so on the site ("formerly known as") rather than hoping the old name fades.
On the measurement side, conflation also pollutes tracking. If the tool counts the other company's mentions as yours, every number is wrong. We set the brand's aliases and the analysis instructions in the Brand Book before trusting any score.
Invented: a feature or integration that doesn't exist
This one surprises clients. The answer says they offer a mobile app, an API, or an integration they've never built. Models fill gaps. When a buyer asks "does X work with Y" and no page answers it, the model guesses from what similar products do.
Fix path: answer the question on your own site, including when the answer is no. An integrations page that lists what exists, an FAQ entry that says "we don't have a native mobile app; the web app works on mobile browsers," a features table that's actually text and not an image. Then check whether a comparison page or roundup somewhere claims the feature. If it does, you're in third-party territory, which is the subject of our post on AI citing wrong information about your brand.
Wrong audience or positioning
Nothing in the answer is technically false, but the framing is off. The model calls a mid-market product "an enterprise tool," recommends a B2B platform for freelancers, or files the client in the wrong category entirely. This costs deals quietly, because buyers who fit never see the brand in the right shortlist.
Fix path: this is slower content work. Pages that say who the product is for and who it isn't for. Pricing that makes the entry point visible. Category listings on directories that match how buyers search. Comparison content against the competitors the client actually wants to be shortlisted with. Expect this to move over months, not weeks.
True but unflattering
"Users report slow support response times." "The reporting module is limited." Sometimes the model is right. We say that to the client directly, because the fix isn't content. It's a product or support issue, and trying to bury it with blog posts tends to fail and looks bad when it does.
What content can do is report the change once it's real. If support hours were extended or the reporting module was rebuilt, publish that clearly, with a date, and make sure the help docs and listings reflect it. Until then, we route the claim to whoever owns the product and leave it out of the SEO backlog.
Check your own site before you blame the internet
For the first four types, the client's own site is the place we look first. In our audits it is also where we most often find the source: a stale article, two pages that contradict each other, a fact that only exists inside a PDF. Fixing your own page is a day of work. Getting a third party to change theirs can take a month.
How we do this on Promptwatch
Promptwatch is the platform we run client programs on, and three parts of it map onto the process above.
Brand Narrative, on every plan, does the collection step. It pulls the individual claims AI makes about the brand and its competitors out of tracked responses, groups the repeats so you see how often each appears and which models say it, and files them 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. That's the frequency question answered without copying sentences into a spreadsheet.
Fact Check does the "check your own site" step. You mark a claim Accurate or Inaccurate, and it reads the pages Promptwatch has crawled from the client's site for ones that state the claim or the opposite, with quotes and a suggested next step. For an outdated price, it's the quickest way we know to find the forgotten page. Fact Check is on Business ($579 a month), Agency Scale ($799 a month), and Enterprise. It uses Agent Credits, and you see the estimate before it runs. On Essential, Professional, Kick-off, or Growth, you'd do this step by hand.
Since November 4 there's also an accuracy score: the share of the brand's checked mentions that are accurate, as one number and as a heatmap per model and theme. A claim you've marked counts as you marked it; otherwise the latest completed fact check decides. Inconclusive and unchecked claims are left out, so the score only reflects what has actually been checked. We use the heatmap to show a client which model and which theme to work on next. Details are in the Promptwatch changelog.
Re-measure, and set expectations
After a fix, we wait for the page to be recrawled, then re-run the same prompts and compare how often the claim appears. We report the change in frequency. We don't promise the sentence disappears, because models blend sources and one corrected page is one input. Our post on how often to re-check AI answers for accuracy covers the cadence we use after a fix and after any pricing or product change.
If you have one of these tickets open right now, send the prompt, the engine, and the response to hello@1001seomedia.com. We'll tell you which of the five types it is and where we'd start.