What Should You Do When an AI Engine Describes Your Product Wrong?

What should you do when an AI engine describes your product wrong?

Fix the source before you complain to the engine. A wrong answer is almost always a retrieval problem, not an invention, which means somewhere on the open web there is a page that says the wrong thing more clearly than your page says the right thing. That page is the real target.

The instinct is to look for a complaint form. There is one, and it is worth using, but it is the second step rather than the first. Reporting is a request. Publishing a better source is a fix.

Here is the order I work through when a client shows me an AI answer that misstates what they sell.

Why does this happen in the first place?

Because the engine is assembling an answer from whatever it can retrieve, and your own page is only one candidate among many. Old pricing on a review site, a stale directory listing, a comparison article written by a competitor, or your own outdated page can all outrank the current truth.

The second cause is worse and more common than people expect. Your site may simply not say the thing clearly. If your homepage describes an outcome and never names the mechanism, an engine asked what the product does has to reconstruct it from third parties, and third parties are usually a year behind.

Google is explicit that this goes wrong. Its own help documentation says AI responses may include mistakes and that AI Overviews can and will make mistakes, because the technology may provide inaccurate or offensive information. Treat that as a design constraint rather than a scandal.

Should you report it, and does reporting work?

Report it, but understand what you are buying. Google documents a feedback path on AI Overviews: use the thumbs down if the overview is unhelpful, inaccurate, biased, or otherwise problematic, then choose Report a problem, pick the category that best describes your issue, add details, and submit.

The same options sit under the three dot menu at the top right of each overview, so you do not have to hunt for them.

What Google promises in return is modest and honest. The documentation says your feedback helps improve AI Overviews, and that feedback and human reviews are used to evaluate and improve quality. It does not promise that your specific answer will change, and it does not promise a reply.

So I file the report, screenshot the answer with the date, and move on to the work that actually changes the outcome. If you are relying on the report to fix it, you are relying on something nobody committed to.

What can you actually control on your own site?

The claim itself, stated plainly, in one place, in a sentence an engine can lift. Most companies have the correct information distributed across a homepage, a pricing page, and a feature page, in three different phrasings. That is three weak signals instead of one strong one.

So I write a single canonical statement and repeat it verbatim wherever it belongs. What the product is, who it is for, what it costs or how pricing works, and what it explicitly does not do. That last one matters more than people expect, because a clear negative statement is the fastest way to kill a wrong positive claim circulating elsewhere.

Then I make it easy to retrieve. A heading phrased as the question a person would ask, an answer in the first forty to sixty words underneath it, and no dependence on an image or a table to carry the meaning. This is the same structure that earns citations generally.

Which crawler settings actually matter here?

Only the ones tied to the surface you care about, and they are separate levers. OpenAI's own crawler documentation says OAI-SearchBot is used to surface websites in search results in ChatGPT's search features, and that disallowing it prevents your site from appearing in ChatGPT search answers, though it may still show as navigational links.

GPTBot is a different lever entirely. OpenAI describes it as the crawler used to collect content that may be used in training its generative AI foundation models, and says disallowing it indicates that a site's content should not be used in training. You can decline training while remaining eligible for search, because the settings operate independently.

Two more names are worth knowing so you do not block the wrong one. ChatGPT-User is described as being used for certain user actions in ChatGPT and Custom GPTs, and the documentation states it is not used for automatic web crawling or to determine appearance in ChatGPT Search. OAI-AdsBot is described as being used to validate the safety of web pages submitted as ads, with data not used for training. If you have inherited a robots file, check which of these you are actually blocking before you conclude the engine is ignoring you.

How do you find the page that is feeding the wrong answer?

Ask the engine to show its work. Most answer surfaces cite something, so start by reading the cited sources rather than the answer, and look for the sentence that most closely resembles the mistake. Nine times out of ten it is sitting there in plain text on a page you forgot existed.

Typical culprits are a directory profile you filled in three years ago, a marketplace listing with old pricing, a partner page describing an earlier version of your product, and an old blog post of your own that you never updated. The last one stings, and it is the easiest to fix.

Ask the same question across several engines before you decide what the problem is. If one engine is wrong and three are right, the issue is retrieval on that surface. If all four are wrong in the same way, you have a source problem in the world. I have a longer piece on why the same question gets different answers across AI engines, and it is the diagnostic step most people skip.

What do you do about sources you do not own?

Ask, in writing, with the correct information attached. Directory operators, review platforms, and publications update listings far more often than founders assume, because an out of date entry is their problem too. A polite email with the current facts and a link to your canonical page works more often than it fails.

When the source is a competitor comparison page, you are not going to get it changed, so the response is to publish your own version of the comparison and be scrupulously fair in it. A page that states the competitor's genuine strengths is more likely to be trusted as a source than one that does not, by humans and by retrieval systems alike.

Where you cannot change the source, change the weight of evidence around it. More pages saying the correct thing clearly, on domains that get crawled, is a slower fix than a report form and a far more reliable one.

How long should a correction take, and how do you know it worked?

Nobody publishes that number, so do not trust anyone who quotes you one. The honest answer is that it varies by engine, by query, and by how often your page is recrawled, and the only way to know is to measure your own case rather than accept a general claim.

So set up a check. Write down the five questions where the wrong answer appears, run them on a fixed schedule across the engines you care about, and record the answer and the cited sources each time. A spreadsheet is enough. What you want is a dated before and after, not a feeling that things seem better.

Be aware that answers move on their own even when nothing changed on your side, which makes a single retest meaningless. That instability is the subject of why AI answers vary between sessions, and it is why I insist on a schedule rather than a spot check.

When is a wrong answer a bigger problem than a marketing one?

When it states something false and damaging about your company as fact, rather than describing your product imprecisely. Pricing that is out of date is an annoyance. An answer asserting that you were involved in something you were not is a different category, and that is a conversation for your lawyer, not your SEO checklist.

There is a commercial angle here that is easy to miss. Gartner's survey found that 69 percent of B2B buyers turn to sales reps to validate AI-generated insights. That cuts both ways. It means a wrong answer often reaches your sales team before it reaches you, and your sales team is your earliest detection system.

Which is why I ask clients to forward anything strange a prospect repeats back to them. The pattern in those forwarded messages usually reveals the wrong source long before any monitoring tool would.

What should you do next?

Pick the three questions a buyer is most likely to ask about your product, run them through the engines your buyers use, and write down exactly what comes back along with the sources cited. That is an hour of work and it will tell you whether you have a problem worth solving.

If something is wrong, fix the clearest source first, publish one canonical statement on your own site, file the report, and then re-run the same three questions in two weeks. That sequence respects what you control and what you do not.

I spend a lot of my week on exactly this kind of diagnosis, and the pattern rarely changes: the engine is not out to get you, it is reading something old that you forgot to update. If you want help working out which source is feeding a wrong answer about your business, reach out and let's chat about it. You can also start with how to measure AI visibility for your brand if you have no baseline yet.