Report the specific answer through the tool's own feedback control, then go fix the page the model actually pulled the fact from. The feedback button tells the platform something looked wrong. It does not update what the model says next time, and it is not a channel for sending a correction to a human who edits your company's entry anywhere. The only thing that reliably changes what an AI assistant says about you is changing the source material it is reading.
What actually happened when Google's own Gemini got a fact wrong?
In February 2025, Google ran a Super Bowl ad for Gemini built around a Wisconsin cheese shop. The ad showed the owner asking Gemini to write a product description, and the draft it produced stated Gouda "is one of the most popular cheeses in the world, accounting for 50 to 60 percent of the world's cheese consumption," according to Fortune's reporting. That number is wrong. Industry data cited in the same Fortune piece puts mozzarella and cheddar well ahead of Gouda, which does not place in the top ten.
A travel blogger flagged the error publicly before the game aired. Google's response, reported by Fortune, came from Jerry Dischler, the company's president of cloud applications, who posted on social media: "Hey Nate, not a hallucination" and "Gemini is grounded in the Web, and users can always check the results and references." Google's position was that the model had faithfully repeated a number that other websites had already published, not invented one. Google quietly edited the ad to remove the statistic before it aired, and a company spokesperson later told Fortune that "the AI didn't hallucinate, it drew on a website that was wrong."
That distinction matters more than it sounds like it should. If the problem is the model inventing something from nothing, the fix is a better model. If the problem is the model accurately summarizing a bad source, the fix is a different source, which is a problem PR and content teams can actually do something about.
Why does an AI assistant state a wrong fact so confidently?
Because the systems behind ChatGPT, Gemini and Perplexity are built to synthesize what is already published, not to independently verify it. A wrong number repeated across enough low-quality pages starts to look, to a retrieval system, like a number worth repeating. The model has no separate process for fact-checking that output against a primary source before it answers you. It is reading the web the way a very fast, very confident intern would, and an intern who read the same wrong statistic on five different sites will repeat it just as confidently as if it were correct.
This is also why a company's own thin, inconsistent, or outdated web presence is a direct liability. If the only pages mentioning your company are a few years old, written by someone else, or contradict each other on basic facts like your founding year or headquarters, that is the material an assistant has to work with when someone asks about you.
What is the actual reporting mechanism on each platform?
Each of the three major assistants has a narrow, built-in way to flag a bad answer, and none of them is a correction request.
- Gemini. Google's own support page for Gemini Apps describes clicking the thumbs-down ("Bad response") icon under a reply, then optionally choosing a reason and adding written detail before submitting, according to Google's help center. The same page notes that the conversation, including your prompt and Gemini's response, is submitted along with the feedback.
- ChatGPT. OpenAI's own help center describes a report flow reached from the message itself: select the report option on a conversation and follow the on-screen prompts, per OpenAI's published article on reporting content. ChatGPT's interface also carries a standing disclaimer that it can produce inaccurate information about people, places or facts.
- Perplexity. Perplexity's own support page lists the flag icon below an answer, or a support ticket through its help channel, with misinformation and outdated information given as example issues worth reporting, per Perplexity's published support page.
All three ask what is wrong with a single answer. None of them offers a path to tell the company "here is the correct founding date for this business, please update your records," because there is no single record to update. There is only a model deciding, at the moment someone asks, which pages to trust.
Does reporting a wrong answer actually fix it?
Sometimes, eventually, for that specific question, and only if enough signal accumulates. Feedback like a thumbs-down is one input into how a platform tunes its systems over time. It is not a guarantee that the next person who asks the same question gets a different answer, and it does nothing to the underlying pages the model is drawing from. Google was explicit about this distinction in the Gouda case: the ad was fixed by changing the product copy on the cheese shop's own website, not by anyone filing feedback inside Gemini.
If your goal is "stop the model from saying this," reporting the single bad answer is a reasonable first step and takes thirty seconds. If your goal is "make sure this stops being said consistently," reporting is not the mechanism that gets you there.
What should you actually do instead of just clicking thumbs down?
Treat it as a sourcing problem, not a complaints process.
- Ask the assistant where it got the claim. ChatGPT, Gemini and Perplexity will often name or link the pages they drew from, especially Perplexity and Gemini, which show citations by default. Ask directly: "what source did you use for that."
- Go fix or flag that source. If it is your own website, correct it immediately and make sure the corrected version is the one search engines and crawlers can actually see, not buried behind a login or a slow-loading script. If it is a third-party page, an old press mention, a directory listing, or a forum post, request a correction from whoever controls it.
- Publish the correct version somewhere durable. A dated, factual page on your own site, a quote in a piece of actual press coverage, or an updated entry on a reference source the assistants already trust gives the next retrieval pass something accurate to find. Earned coverage does a disproportionate share of this work, because independent reporting tends to outrank a company's own marketing copy in how these systems weigh trust.
- Report the specific bad answer anyway. It costs nothing and is one more data point, even though it is not the fix.
- Recheck in a few weeks, not a few hours. These systems do not recrawl and resynthesize instantly. AI browsers and chat interfaces pull from indexes that update on their own schedule, not on yours.
When is this a PR problem rather than a technical one?
As soon as the wrong fact is something a journalist, investor, or customer could reasonably repeat without checking it. A minor factual slip buried in a long answer is low stakes. A wrong founding date, a wrong claim about who leads the company, a wrong number tied to funding or revenue, or an outdated product description that a prospective customer might act on is a different situation entirely, closer to a crisis communication problem than a search problem. In those cases, the fastest real fix is usually the same one that works for any factual error in the press: a short, specific, dated correction published somewhere credible, not a long explanation of why the AI got it wrong.
It is also worth checking occasionally rather than waiting for a customer to mention it. Ask ChatGPT, Gemini and Perplexity the same handful of basic questions about your company every few months: what you do, when you were founded, who runs it, what you are known for. A pattern of wrong answers is a sourcing gap worth fixing before a bigger audience notices it.
Legal recourse is not a reliable backup plan here either. A German court recently ruled that Google can be held directly liable for false claims its AI Overviews generate, while a separate German court reached the opposite conclusion weeks later. The law is unsettled even in the one country where it has been tested, so fixing the source material remains the only lever you can count on regardless of jurisdiction.
The direct answer, for anyone who read this far
Report the single wrong answer through whichever platform's built-in feedback control, since that is the only official channel any of them offers, but do not treat it as the fix. Find the specific page the assistant actually drew the claim from, correct or flag that page, and get an accurate, durable version of the fact published somewhere credible so the next retrieval pass has something better to find. If a wrong fact about your company is sticking around across ChatGPT, Gemini or Perplexity and you want help tracking down where it is coming from and fixing it at the source, tell us what the assistants are getting wrong.
