Consider a subscription-software company (a composite drawn from a pattern I've seen more than once, not any single client's story) whose sales team fielded about a dozen calls in a single month from prospects who all quoted the same monthly price. The number was wrong. It was the old price, retired nine months earlier when the company moved to a new tier structure. Nobody on the sales team had said it first. The prospects were reading it back to the reps, confident, because an AI assistant had told them that was the price.
The marketing team's first move was the obvious one: update the pricing page again, just in case. It was already correct. They checked the schema markup. Correct. They checked the About page, the press page, even old blog posts. All current. The wrong number kept surfacing anyway, in answers pulled from more than one AI assistant, weeks apart.
This is the pattern I want to walk through here, because it's more common than most brands realize once they start checking, and because the instinct to just fix the website isn't wrong. On its own, it's insufficient.
Why the First Fix Falls Short
In my work correcting AI-search errors, the first fix almost everyone reaches for is editing their own site. That makes sense. It's the property you control, and a decade of search engine optimization trained us to treat our own site as the primary lever. That reflex doesn't map cleanly onto how a lot of today's AI answers actually get produced.
An AI answer engine's response to a factual question is usually a blend of two things. Part of it comes from what the model absorbed during training (patterns learned from a snapshot of the web at some earlier point), and that belief persists inside the model until it's retrained or refreshed on whatever schedule the provider runs. Part of it, for retrieval-based tools like Perplexity, ChatGPT with search enabled, or Google's AI Overviews, comes from pages fetched at the moment you ask the question.
Here's the part that trips people up: even when the engine is fetching live, it isn't necessarily fetching your live page. In my experience it often favors third-party sources (a directory listing, an old press release, a comparison site, a cached company profile) that restate the outdated fact with more apparent authority than your own site carries on that specific claim. Your one corrected page gets outvoted. And if the wrong fact came primarily from training rather than retrieval, your page edit doesn't touch it at all until the underlying model changes.
That's the gap. Editing your site treats the symptom on the one page you control. It does nothing about a training-era belief, and it often doesn't out-rank the third-party pages doing the fetch-time damage.
Where the Wrong Answer Actually Lives
Once you accept that the fix isn't purely on-site, the next question is where the wrong fact is actually living. In my work I've found it almost always traces back to one of five kinds of source:
- Press archives, a launch announcement or press release from a prior pricing era, still indexed and still being cited
- Directories and listings, self-serve profiles on sites like G2 or Capterra, filled in once and never revisited
- Review and comparison sites, third-party "X vs Y" pages built by someone outside your company, often ranking well precisely because they're independent
- Knowledge panels and Wikidata, structured facts that feed multiple downstream tools at once, so one stale field can echo widely
- Old PDFs and press kits, spec sheets and media kits that were correct when published and have sat untouched since, frequently still linked from somewhere
None of these are on your website. All of them are indexed, and several are exactly the kind of source a retrieval-based AI assistant tends to treat as an authoritative third party rather than a self-interested first party. That distinction can work against you: the tool may weigh your own claims about your pricing less heavily, yet readily repeat someone else's stale claim about it.
I wrote before about the broader measurement problem this creates in The AI-Search Measurement Crisis: most brands can't even tell you how they're currently being described, let alone whether it's accurate. Correction work starts downstream of that same blind spot.
Tracing a Wrong Answer to Its Source
When I sit down with a brand to trace a specific wrong answer, the process is closer to detective work than SEO. Some AI assistants show their sources inline or in a footnote; when they do, that's the fastest path: you read the citation and go straight to it. When they don't, and plenty don't, I take the exact wrong phrase as it appeared in the answer and search it verbatim, in quotation marks, on a regular search engine. In my experience the exact phrasing usually surfaces the source page directly, because whoever wrote the underlying content didn't invent the wording; it got lifted.
From there I rank the candidate sources by two things: authority (how likely this page is the one multiple tools are drawing from, judging by its domain, its age, and how often it's cited elsewhere) and freshness (when it was last touched, and whether the platform shows any sign of when it was crawled). A five-year-old directory listing with steady backlinks can outweigh a company blog post from last quarter in what an AI tool reflects back.
The Correction Checklist
This is the sequence I use once a wrong fact has been identified and traced. It's slower than editing a page, and it's the part that actually moves the answer.
- Map current AI answers for your five to ten highest-stakes facts (pricing, product lineup, leadership, headquarters) across two or three surfaces (for example ChatGPT search, Perplexity, Google AI Overviews)
- For each wrong answer, check whether the tool shows a citation; if not, search the exact wrong phrase verbatim to locate the likely source
- Rank the candidate sources you find by authority and freshness, and start with the one most likely to be feeding the answer
- Correct at the source itself: send an editor correction request to a press outlet, update your self-serve panel on a directory, or send a formal correction notice where there's no self-serve option
- Update your own site's structured data, About page, and press page so they state the current fact clearly; this is necessary, but on its own not sufficient
- Claim your Google Business Profile and Knowledge Panel and request corrections, and check your Wikidata entry, where any of these exist, since these often feed multiple tools at once
- Re-test the same question two to four weeks later, not the same day; corrections at third-party sources and any model refresh both take time to propagate
- Log the correction date for each source and set a quarterly recheck, because these facts drift again
What "Fixed" Actually Looks Like
Fixed doesn't mean every AI assistant gives you the right answer the next time you ask. In my experience it means the wrong answer's frequency drops across the surfaces you're tracking over a several-week window, and when it does appear, it's inconsistent rather than confidently repeated by every tool you check. Full disappearance can lag behind the correction itself, because you're waiting on some combination of a crawl, an index update, and in some cases a model refresh you don't control and can't schedule.
I'd also flag a related trap on the other side of this problem, where the issue isn't a wrong fact but weak entity recognition: the AI describes what you do without confidently naming you. I covered that separately in Named or Just Described, and it's worth ruling out before you assume you're dealing with a stale fact rather than an entity problem.
The mental model I've settled on, after doing this more than once, is that your off-site presence, every directory, archive, and listing that mentions your company, is decaying inventory. It doesn't stay accurate because you were once accurate somewhere. It needs a maintenance cadence, the same way you'd maintain any asset that degrades on its own timeline whether or not you're watching it.
Related evidence: Google's Search Central documentation says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response, and can show a wider and more diverse set of supporting links than a classic web search. (Google Search Central guidance on AI features)
The research paper that introduced Generative Engine Optimization describes generative engines as synthesizing information from multiple sources, and notes that content creators have little to no control over when and how their content is displayed in those responses. (the arXiv paper introducing Generative Engine Optimization)
For more on this and related work, see BrianServes, the platform for deploying autonomous AI e-mployees and Webspot, the AI strategy firm in Beirut.
Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.
