ChatGPT is wrong about your business. Now what?

When ChatGPT tells people your business closed, quotes prices you retired, or credits you with someone else's story, you can fix it — but not by arguing with the chatbot. Engines get businesses wrong for three reasons: stale training memory, wrong third-party sources, and identity confusion with a similarly named company. Each cause has its own correction path, and the paths run on different clocks: source-side fixes can land in weeks, while the model's internal memory waits for a training cycle nobody schedules for you. This page walks through the diagnosis, the repair for each cause, and the timelines as they actually are.

Step zero: write it down, then breathe

A wrong AI answer about your own business reads like an emergency — it's your name in a stranger's mouth, saying things you never said. Before doing anything, know that you have company. This is a structural failure mode of how answer engines work, not a judgment aimed at you:

Then document the error before you touch anything. Record the date, the exact prompt, the engine, whether web search was on or off, and a screenshot. Two reasons. First, this becomes your before/after receipt — the only honest way to prove a correction landed is a dated wrong answer next to a dated right one. Second, AI answers drift between runs: the 2025 RankOS AI Visibility Benchmark — a visibility platform's measurement of its own runs, so directional — found only about 30% of brands named in one run were named again on the next run of the same query. Re-ask the same prompt a few times over a day or two — the five-prompt check is built for exactly this — so you know whether you're fixing a repeating story or chasing a one-off.

Last, triage. An outdated tagline is a footnote; “permanently closed,” invented prices, or a blended identity loses customers silently and goes to the top of the list. And if the engine doesn't mention you at all, that's a different problem with a different page — why your business doesn't show up in ChatGPT covers absence; this one covers wrongness.

The three causes, and the one-minute tell

Every answer an engine gives comes from some mix of two pipelines — what the model memorized in training, and what it retrieves from live web search (the two pipelines are their own page). Wrong answers trace back to one of three causes, and each leaves a distinct fingerprint:

• Stale training memory. The error matches your past: the old address, the discontinued service, the pre-rebrand name. The tell — it shows up with web search off, and often corrects itself when search is on.
• Wrong sources.The error matches something published somewhere: a directory that never updated, a 2023 article, an abandoned profile. The tell — it shows up with web search on, and the answer's citations carry the same wrong fact.
• Identity confusion.The “facts” were never yours at all — they belong to a similarly named business. The tell — details you can't place, in either mode.

The causes stack. A thinly documented business with a stale snapshot anda common name can hit all three at once — which is fine, because the paths below don't conflict. Run each one that applies.

Cause one: the model's memory of you is stale

Training doesn't store your business as a record it can edit; it compresses everything the public web said about you — up to a cutoff date — into a statistical impression. If you rebranded, moved, or killed a product line after that cutoff, the model keeps reconstructing the old version of you with total confidence.

The path.Be clear about what's possible: there is no way to edit a model's memory, no correction form for its weights, and anyone selling a service that “removes” wrong facts from a trained model is selling something imaginary. What you can do is outflank the memory:

1. Publish a current, unambiguous facts page — name, category, locations, hours, what you sell today — with Organization or LocalBusiness schema that says the same thing the visible page says.
2. Make the current story consistent everywhere the next training run might read: your site, your profiles, the directories that carry you.
3. Get the corrected pages indexed, so live retrieval — the pipeline that usually answers buying-intent questions — reads the new story and overrides the old one.

The honest timeline. The masking move — winning the retrieval layer so search-on answers tell the current story — is a weeks-scale project. The memory itself corrects only when the vendor ships a model trained on a newer snapshot of the web. No engine publishes that schedule, no one can move it up for you, and the copy you publish now is your application to be read correctly by the next one.

Cause two: the sources it reads are wrong

This is the most fixable cause, because it leaves a paper trail. When a search-on answer states a wrong fact, expand its citations and read them — one of them almost always carries the error. And the error is rarely on your own site: citation studies — including one the monitoring vendor Octolens published from its own dataset — put roughly nine in ten cited sources on third-party pages rather than the brand's own. Your story is being told by a stale directory listing, an old press mention, or a profile you forgot existed.

The path. Work outward from the citations:

1. Fix what you control today. Your own pages, your Google Business Profile, your social profiles, listings where you hold the login. Same name, same facts, everywhere.
2. Request edits where a third party controls the page. Directories and review platforms usually have an update path; publications sometimes correct, sometimes don't. Their clock, not yours — log the request and move on.
3. Outweigh what won't edit.A dead blog post from 2022 can't be deleted by wishing. It can be beaten: publish fresher, more specific, better-structured pages on the same facts, and the stale source loses the retrieval contest over time.
4. Get the corrections crawled and re-indexed. A corrected page nothing has crawled is, to a retrieval step, a page that was never written. Two parts: make sure the search crawlers can reach it — OAI-SearchBot for ChatGPT Search, Claude-SearchBot for Claude, PerplexityBot for Perplexity, none of which are the training crawlers people usually block — and then push the URLs into the indexes.

Submitting corrected URLs through IndexNow gets them in front of Bing within minutes — Bing then decides what to index and when. The same submission is the whole story for Microsoft Copilot, which answers from Bing's own answer layer, so a correction that lands there is fixed on at least one surface with no inference required. The Bing and Copilot guide covers the mechanics.

The honest timeline.Days to weeks for everything you control; third-party edits range from days to never. The “never” cases are why the outweigh step exists — you don't need every wrong page gone, you need the pages engines actually retrieve to be right.

Cause three: it thinks you're someone else

A statistical impression of two similarly named companies can merge into one confident, wrong description — your name attached to their locations, their reviews, their lawsuit. The blend punishes the less-documented party: whichever business the web describes more thoroughly tends to win the merged identity.

The path.You can't make the other business disappear, so you make yourself unblendable:

• Use the full, distinguishing name everywhere — “Crestline HVAC of Tampa, Florida,” never bare “Crestline” — on the pages engines quote: title tags, headings, schema, profiles.
• Keep name, address, and phone identical across every listing — the local AI search stack runs on that consistency — and add Organization schema with sameAslinks tying your site to your real profiles, so the engines' picture of “you” has hard edges.
• Make geography and category explicit in the copy itself. An engine can't confuse two companies it can cleanly tell apart; it merges the ones described in interchangeable generalities.

The honest timeline.Split. Retrieval-side disambiguation is a weeks-scale project like cause two. The blend inside the model's memory is the stubbornest problem on this page — it lasts until a retraining pass reads your now-unambiguous record. What makes the wait survivable is that buying-intent questions usually trigger live search, so the layer you can fix is the layer most buyers actually hit.

What fixes in weeks — and what honestly waits

Everything above sorts into two piles, and knowing which pile you're working in is the difference between a plan and a frustration.

Moves in weeks: your own pages and schema, profile and directory corrections that accept edits, search crawlers let back in, submitted pages getting indexed, and the search-on answers that read all of it. This pile is fast enough to watch:

Waits, and nobody can honestly promise otherwise: the model's trained memory, identity blends baked into that memory, third-party pages that refuse to edit, and the slow accrual of enough corroboration that engines repeat your story unprompted — practitioners put consistent recommendation at a 6–12 month build in most categories (Cited.so, 2026). That longer build is its own playbook. If a vendor tells you they can edit what a model already believes, on a date, they can't. The realistic promise is narrower: make the live layer right in weeks, and let every future snapshot of the web read a record that finally says the same true thing everywhere.

Monitoring is the tripwire, not the trophy

The correction isn't done when you publish the fix. It's done when the answers change — and stays done only as long as they stay changed. Both halves are empirical questions, and both need the same instrument: the documented prompts from step zero, re-asked on a schedule, with dates attached.

One clean re-check proves little on its own — the same drift that made you confirm the error in step zero applies to the fix. What works is a cadence: the same buyer questions, across more than one engine, week after week, with the answers kept. When the wrong story reappears — a resurfaced directory, a new bad source — the cadence catches it while it's one answer old, and your dated before/after record turns “I think it's fixed” into a receipt you can show anyone.

The obvious conflict of interest, addressed

Yes, the company writing this correction protocol sells the machine that runs it: AEO Action's product is this page's loop as software — daily scans that re-ask your buyer questions across up to six engines and keep the dated history, an autonomous department that prepares the ground-truth artifacts — schema, facts and FAQ pages — for your one-click approval before anything publishes, IndexNow submission of what you approve (Bing decides what to index), and fate-tracking that registers each approved landing page against its target question and re-checks it on every following scan for up to 120 days: shipped on date A, cited on date B. The services pagelays out the whole scope and what it costs. The correction protocol above is true whether or not you automate it — the software's job is the cadence humans quit.