What to do when ChatGPT says something false about your client: a protocol for agencies
The email arrives on a Monday: "I just asked ChatGPT about my clinic and it says we close on Friday afternoons. We've been opening on Fridays for two years. Can you fix this NOW?".
First piece of bad news: there's no "correct ChatGPT" button. There's no complaints form to OpenAI for business data, no phone number to call, no guaranteed turnaround. Whoever sells your client that is selling them smoke.
Second piece of news, the good one: the error almost never originates in the AI. It originates in the sources it drinks from — and sources can be corrected. That's the work, and it has a method. Here's the full protocol, in the order you should run it.
Step 1: document with a screenshot before anything else
Before warning anyone, before touching anything: a dated screenshot. The exact question, the full answer, the false fact flagged. And repeat the query two or three times, ideally in different sessions: the AI's answers vary, and you need to know whether the error is systematic (shows up almost every time) or sporadic (showed up once).
The distinction matters enormously. A systematic error almost certainly has a concrete source behind it and deserves the whole protocol. A sporadic error might be a one-off hallucination by the model — it made the fact up that time — and may not appear again; you note it, watch it and don't mobilise the client.
Check the other AIs too: Gemini, Perplexity, Claude. If the same false fact appears in several, there's a contaminated source they're all reading — and Perplexity, which links its sources, often tells you which one directly.
Step 2: identify the likely source of the error
The AI doesn't invent that the clinic closes on Fridays (normally): it read it somewhere. The usual suspects, in order of frequency:
The client's own website. More often than they'll admit. An old, un-updated "opening hours" page, a 2021 landing nobody remembers, an indexed PDF with old prices. Search Google for site:clientwebsite.com + the false fact.
The Google listing (Google Business Profile). Primary source of hours, address and phone for almost every model. Wrong hours, the wrong category, or user-suggested edits that got approved on their own without anyone seeing.
Old directories. Yellow Pages, sector directories, local portals, that profile someone created in 2017 and abandoned. They carry authority in the models' eyes and nobody has touched them in years. Find them: the client's name in quotes + the false fact on Google.
Press releases and old articles. "The company, headquartered in Seville..." when they moved to Málaga three years ago. They can't be deleted, but they can be countered with more recent signals.
Reviews and forums. An old complaint on Google or a forum thread can slip in as a description of the business ("it's expensive", "takes weeks to deliver").
Document each source you find: URL, wrong fact, who can change it. That table is your work plan and, as a bonus, an excellent deliverable for the client.
Step 3: correct the sources, in order of impact
- Client's website: correct the fact, update the structured data (organisation/local business schema with hours, address, phone) and make sure the correct information is on indexable pages, not just in an image or a PDF.
- Google listing: correct it, check there are no pending user edits and keep the listing active (recent posts, photos): live listings carry more weight.
- Directories: claim ownership of the important profiles and correct them; request removal of the dead ones. No need to chase all fifty — prioritise the ones that show up on the first page of Google when you search for the client, since those are the ones the AI reads.
- Fresh content that states the correct fact: an updated page or post that explicitly says the true fact ("hours updated 2026: we open Fridays 9 am to 8 pm"). Models tend to favour the most recent and consistent information; give them a new, clear, dated source that contradicts the old one.
Step 4: monitor until the change is reflected
Here comes the part that calls for managing expectations: correcting the source today doesn't change ChatGPT's answer tomorrow. The models pick up changes when they pick them up — usually weeks, sometimes more, and unevenly across AIs (the ones that search the web in real time, like Perplexity, tend to reflect it sooner than the ones leaning more on their training memory).
So the follow-up can't be "I'll ask one day and see": you have to fire the affected queries periodically, across the 4 AIs, and record when the false fact disappears from each. It's the difference between telling the client "it should be fixed by now" and showing them "Perplexity was corrected on the 12th, on ChatGPT we're still watching." Doing this by hand for one case is manageable; doing it for the whole roster is the manual-screenshots problem all over again. In Surfeo for agencies this follow-up is automatic: each client's prompts are fired every week across the 4 AIs and the history shows you the exact moment the answer changed — the proof your work worked, with a date.
Step 5: manage the client in the meantime
The client wants the fix "NOW", and "NOW" doesn't exist. The script for the first conversation:
Key data
"We've verified and documented it. You can't correct ChatGPT directly — there's no such button, I'm telling you honestly — but we do know where it gets the error from: [the source]. We're already correcting it. The models take weeks to pick up changes, so we're going to watch the answer every week and let you know the moment it changes."
Three things make this conversation go well. First: you show diagnosis, not excuses — you know where the error comes from. Second: you put a deadline on what you control (the source correction, this week) and an honest range on what you don't (the reflection in the AI, weeks). Third: you turn the wait into visible service, with a two-line weekly mini-report: "sources corrected: 4 of 5; ChatGPT still shows the old fact; Perplexity now gets it right."
And a thought for the next sales meeting: this incident is the best possible demonstration of why continuous monitoring matters. The false fact had been out there who knows how long, answering every person who asked, and nobody knew. With 37.9% of the Spanish population already using generative AI (INE, last quarter of 2025), that's a lot of wrong answers. Catching it in the weekly measurement, before the client's brother-in-law does, is exactly the service you're selling — and it fits squarely into what you can actually commit to in a proposal.
Frequently asked questions
Can I complain directly to OpenAI, Google or Perplexity?
For business data, in practice there's no effective channel: there's no business-information correction process as such. There are generic feedback forms (the thumbs-down on the answer) that don't hurt to use, but expect nothing from them. The real route is the sources. A separate case is content affecting personal data or defamation, where there are legal and privacy routes — but that's already lawyer territory, not agency territory.
How long does the answer take to correct once the sources are fixed?
No guarantees: typically between two weeks and a couple of months, sooner in the AIs that consult the web live and later in the ones that depend on their training. That's why you monitor weekly instead of promising a date.
And if the false fact comes from a source I can't change, like a press article?
You can't delete it, but you can bury it: generate more recent, consistent, higher-authority signals that say the correct thing — an updated website, structured data, a live Google listing, new mentions. The models weigh freshness and consistency: if nine recent sources say X and one old one says Y, X ends up winning.
Should I tell the client if I spot the error before they do?
Yes, always, and with the screenshot. Spotting it first is one of the few times when delivering bad news makes you look good: "our monitoring has spotted this; we're already on it." It's the service working in front of their eyes. Hiding it and letting their brother-in-law find it is the scenario that costs you the renewal.
Do you know what the AIs are saying about your clients today? Take the free visibility test with any of their websites: better you discover the false fact this week than your client does next Monday.
