Someone on your sales team forwards a screenshot. ChatGPT has told a prospect your product does something it stopped doing last year, or priced you at a tier you retired, or filed you in the wrong category entirely.
The usual advice is to go and correct it. That advice works, and it’s incomplete, because the answer drifts back.
Why AI gets your product wrong in the first place
Two things feed every answer, and both can be wrong about you, as we have covered in where ChatGPT gets its information and how ChatGPT is trained. The model learned a version of your company at training time, and it pulls in whatever it can find live when somebody asks. A stale review page and a two-year-old funding announcement will each outrank your own product page if that’s what the system reaches first.
Those two sources fail in different ways and respond to different work.
| Where the error lives | What it looks like | What moves it |
|---|---|---|
| Something it fetched | Wrong detail, with a citation attached to it | Fixing the page it cited |
| Something it learned | Wrong detail, stated confidently with no source | Being described correctly, widely, over months |
| Mistaken identity | Your details blended with a similarly named company | Clearer, more consistent signals about who you are |
The third one catches companies with generic names, and it’s the most frustrating to sit with, because nothing you publish about your product addresses it. The entity is the asset there. Until the system can tell you apart from the other firm with your name, everything else you fix gets attributed to a coin flip.
How to tell which kind of error you’re looking at
Ask the same question twice, once in a way that triggers a live search and once broadly enough that the model answers from memory. If the wrong fact appears only in the broad answer, it’s baked in. If it comes with a citation attached, follow the citation, because that page is your actual problem.
The broad version is something like asking what your company does, with no other context. The narrow version names you alongside a current detail worth checking, which pushes the system into looking rather than recalling.
Run both on more than one assistant, including Perplexity, which fails differently. They don’t share sources and they don’t fail the same way, so a wrong detail that shows up everywhere is telling you something different from one that shows up in a single place.
Write down the citation for every wrong answer that has one. That list of URLs is your work queue, and it’s usually shorter than people expect. We were surprised the first time we did this properly at how often four or five bad answers traced back to the same two pages.
Fixing the source is the easy half
Correcting what the model found is genuine work and it does move answers. Update the page it cited, get the stale directory entry changed, publish the current version somewhere crawlable and in a format AI pulls from. Errors that came from retrieval respond to this within weeks, sometimes days, and the improvement is real.
So teams do it. They work the queue, they watch the answers flip, and they close the ticket. We tried stopping there ourselves on an early engagement and it didn’t work, which is how we learned to build the next part into the plan from the start.
That’s the right first move, and we’d do the same. Where it goes wrong is what happens next, which is nothing. The queue is empty, the answers look correct, and accuracy quietly comes off the list of things anybody is responsible for.
And then it comes back
The answer drifts back because nothing structural changed. A new comparison article gets written without you in it. The review you corrected scrolls under six newer ones. The model retrains on a web that still describes you the way you were. Correcting an answer is maintenance, and maintenance you do once isn’t maintenance.
Think about what you actually changed when you fixed those pages. You edited a handful of documents in a corpus of thousands that mention your category, most of which you don’t control and can’t see. The system re-reads that corpus constantly. Your correction is one vote against a lot of older, better-linked, more frequently cited votes.
Then the slower problem underneath it. Whatever the model absorbed at training time is still in there, and no page edit reaches it. The only route in is being described accurately across the web, consistently, which is work that happens on other people’s sites, for long enough that the next training run picks up the current version of you rather than the old one. Build the entity first and the answers follow; work the answers alone and you’re weeding a garden you never planted.
None of this argues against fixing the sources. It argues against treating the fix as a finish line. The companies whose AI answers stay right are the ones who noticed the drift the third time and started watching for it, rather than the ones who ran a single big correction project.
A 30-minute call where we map what AI says about you and build a tailored roadmap.
Treat accuracy as a number you watch
Score it the way you score anything else you manage. Take a fixed set of factual questions about your category and your company, built the way we build a prompt list for finding visibility gaps, run them on a schedule, and record how many answers get you right. One number, tracked over time, sitting next to sentiment and share of voice.
That’s the whole method, and the discipline is in the fixed part. The set of questions has to stay the same between runs or the number means nothing, which is the mistake we see most often when teams try this themselves.
A few things worth deciding up front. What counts as correct, since “close enough on category, wrong on pricing” needs a rule rather than a judgment call each time. Which assistants you track, because they diverge. And how often, with monthly being enough for most companies and weekly being overkill for all but the largest.
Accuracy earns its place on that scoreboard because it’s the metric a buyer feels directly. Share of voice puts you in the conversation; accuracy decides whether being there helps you. An answer that names you and then describes the wrong product is an influence channel working against you, and it’ll do that quietly for months if nobody’s counting.
We’ve set out how this connects to revenue in how to measure AI search pipeline for B2B SaaS, and where these wrong answers start in why your B2B SaaS doesn’t show up in ChatGPT.
Frequently asked questions
Three questions come up whenever we walk a team through this: whether you can go straight to the AI companies and have it corrected, how long a fix takes to show, and whether any of it matters while buyers still end up on your website anyway. Short answers below.
Can you contact OpenAI or Perplexity to correct it?
There are feedback routes, and they’re worth using for anything defamatory or seriously misleading. For ordinary staleness, don’t build your plan around them. The reliable lever is changing what the systems read.
How long before a corrected answer actually changes?
Retrieval errors, days to a few weeks once the cited page is fixed. Errors sitting in training data, months at best, and tied to a retrain you don’t control. That gap is the single most useful thing to know before you promise anyone a timeline.
Does this matter if buyers still visit our website?
The website visit happens after the shortlist. If the answer that built the shortlist described you wrongly, you either aren’t on it or you’re on it for the wrong reasons, and the site never gets the chance to correct the record.
