AI is citing wrong information about my brand: how we trace and fix the source page
When AI is citing wrong information about your brand, the cited page is usually the problem. How we find it, the three cases we see, and the fix for each.
A wrong answer with a citation next to it is good news, in a way. It means the engine read something, and you can go and look at what it read. Most of the "AI got us wrong" tickets we handle where a source is listed end up being about that page, not about the model.
This post is the source side of the work. If you haven't worked out what kind of mistake you're dealing with yet, start with AI makes mistakes about my brand: what to do, which sorts mistakes into types. Here we assume the answer cites something, and the job is to work out which page is doing the damage.
Step one: open the cited page and read it
This sounds obvious. It gets skipped constantly. Someone sees a wrong price in an answer, assumes the model hallucinated it, and starts writing a blog post. Then we open the cited URL and the wrong price is right there, in a table on a page the client forgot existed.
So for each wrong claim, we record the prompt, the engine, the date, and every cited URL. Then we open each URL and look for the claim, word for word if possible. The outcome falls into one of three cases.
Case one: your own page is stale
The citation points at the client's own domain, and the page says the wrong thing. A pricing page from before the last change. A launch post that described a plan that no longer exists. A help article written for the old interface. Two pages that disagree, where the engine happened to pick the old one.
The fix is the boring one: update the page, or redirect it if it should no longer exist. Put the current fact in plain text near the top. If there's a date on the page, update it honestly. If the page carries structured data (Product, Offer, FAQ), make it match the visible text, because a schema block that still says the old price is one more place for the old fact to live. Check the sitemap lists the page with a current lastmod, and that internal links point at the page that should own the fact rather than the stale one.
Then look for siblings. A stale fact rarely lives on just one page. We search the whole site for the old plan name or number before closing the ticket.
Case two: a third-party page is wrong
The citation points somewhere else. In our client work, three kinds of page come up again and again.
Directory and review listings. Software directories, review profiles, marketplace listings, local listings. These are often the oldest description of the company on the web, written once at signup and never touched again. If the client can claim and edit the profile, this is a quick fix: update the description, category, pricing fields, and screenshots in one pass. Our post on whether review sites help a software brand get recommended goes into why these listings carry so much weight.
Competitor comparison pages. A rival's "us vs them" page with a feature table that was accurate eighteen months ago, or never was. You can't edit it. You can send a polite correction request with evidence (a link to your current page that states the fact), and some publishers do update. The part you control is your own comparison page that states the facts fairly, including where the competitor is stronger. If it reads like a sales sheet, it's just one more biased page in the pile.
Old forum and Reddit threads. A three-year-old thread where someone complained about a missing feature that shipped since. Nobody is going to edit that thread. What works is a clearly disclosed reply from the company's own account with the current information and a link, plus making sure your own site answers the same question directly. Don't use fake accounts. It gets found out, and then that becomes the thread.
Case three: the right page, read wrong
The cited page is correct, and the answer still gets it wrong. This is the most frustrating case, and it usually comes down to how the page presents the fact.
A few patterns we've seen. A price shown per user per month, read as a flat monthly price, because the unit was in a small grey label under the number. A comparison table where the engine attributed a row to the wrong column. A page that's correct but only in an image or an accordion that loads with JavaScript. And blending: the engine took your correct page and someone else's wrong page and merged them into one sentence, citing only yours.
Fix path: make the fact hard to misread. Write it as a full sentence in body text, with the unit spelled out. Don't rely on tables or images alone for anything buyers ask about. Keep the same wording across every page that mentions it. If blending is the cause, the other source is really a case two problem, so go and find it. Query fan-outs help here, because they show the sub-queries the engine ran to build the answer, and one of them usually explains which other page got pulled in. Our notes on structured data for AI search cover the markup side.
Confirming the fix was read
Changing a page doesn't mean the engine has seen the change. Before re-testing answers, we check that the AI crawlers have actually fetched the corrected version. If a bot is blocked or hitting errors, the engine keeps working from the old copy, and rewriting the page again won't help.
How we run this on Promptwatch
Promptwatch is the platform we run client programs on, and it covers each step above.
Finding the page: citation analytics has All Citations, URL Positions, and Citation Trends tabs, with a monitor filter. Citation search flags competitor domains in the results, which makes a rival's comparison page stand out fast. Each citation is classified into 20 source types, including a Directory type for listing and review sites, so you can filter straight to the directory problem. Offsite mentions show third-party pages that name the brand without linking, which is how we find other pages repeating the same wrong claim. The longer version of that is in how we chase offsite mentions.
Watching a specific URL: Page Tracker follows any page, yours or a competitor's, so we put the bad third-party page and the corrected client page on it and watch both.
Case one: Fact Check reads the pages Promptwatch has crawled from the client's site for ones that state a claim or its opposite, with quotes, which is how we find the sibling pages. It's on Business ($579 a month), Agency Scale ($799 a month), and Enterprise, and uses Agent Credits with an estimate shown first. Content Refresh then proposes diff-style updates for the existing page. Our fact-check process post walks through that end to end.
Confirming the fix: Agent Analytics shows IP-verified AI crawler logs and the crawl-to-citation path, with a citation rate per page. That tells an access problem (never fetched) from a content problem (fetched, not cited). Crawler logs start on Professional. Since November 4, each page in the sitemap view also has tabs worth opening after a fix. The Content tab shows the page as crawled, with how much changed between the last two audits, so you can see the site audit picked up the new version before you go looking for AI crawls. The Agent Analytics tab shows crawler visits and citation visits for that one URL. The AI Visibility tab shows the URL's citation rate and position over time, with the prompts and responses that cite it. Those three tabs answer "did they read it, and did it start getting cited" for a single page.
What we tell clients
A corrected page is one input. Engines blend sources, and a third-party page you can't edit may keep showing up for a while. We report how often the wrong claim appears before and after, and which cited pages changed. That's a more honest measure than one screenshot.
If you want to try the workflow yourself, Promptwatch has a 7-day free trial on monthly billing for its self-serve paid plans, including Business, where Fact Check is included.