A prospect mentions, halfway through a call, that ChatGPT told them your company shut down. Or that you were acquired by a competitor. Or that your flagship product was discontinued in 2023. None of it is true — and you have no idea how many other buyers got the same answer and never called.
This is a revenue problem, not a curiosity. Roughly half of B2B software buyers now start their research in AI chatbots, and 71% rely on them somewhere during the buying process (G2 via PRNewswire). A false answer at that stage is a lost deal you never see.
Most advice on this problem ends with "buy our monitoring dashboard." Monitoring measures the damage; it does not repair it. This is the repair manual: document the error, trace where the model got it, then work a five-layer fix — the Correction Stack — in the order that actually moves answers. Six weeks covers one full cycle: fix, propagate, re-test. The done-for-you version lives in our AI visibility service; everything below you can run yourself.
Key takeaways
- AI gets companies wrong for three reasons: a stale training snapshot, bad retrieval sources, or entity confusion with a similarly named business. Each needs a different fix.
- Triage before acting. Harmless drift, commercial damage, defamation-grade — only the last two justify a six-week project.
- Evidence first: a 10-prompt log across five platforms, with screenshots and dates, is the foundation for every later step, escalations included.
- Work the Correction Stack in order: your own pages → structured entity data → Wikipedia edit requests → third-party sources → platform feedback buttons. Most teams start at the feedback button and fail.
- Source-level corrections typically reach AI answers in roughly four to eight weeks (Sight AI), so re-test on a two-week cadence and judge results at week six.
Why AI gets your company wrong
Three mechanisms produce false answers about companies, and the fix is different for each.
The training snapshot. A model's built-in knowledge is frozen at its training cutoff. If you rebranded, relocated, changed CEOs, or simply survived a rough year after that snapshot, the model remembers the old version of you. Answers given without web access come from this memory, and no support ticket edits it.
Retrieval gone stale. When an assistant searches the live web — ChatGPT with search, Perplexity, Gemini and Google AI Overviews, Copilot — it summarizes whatever it retrieves. If the most retrievable documents about you are an abandoned Crunchbase profile, a 2021 news article, and a directory listing from a previous address, the answer is faithful to bad sources.
Entity confusion. Models merge similarly named businesses into one biography. A large share of "the AI says my company closed" cases trace to a namesake that really did close, get acquired, or get sued. You inherit its history.
Which mechanism you face determines everything downstream — the deeper mechanics are in how AI decides which brands to cite.
Triage: not every wrong answer deserves a project
Before committing six weeks, grade the error.
| Tier | Looks like | Response |
|---|---|---|
| Harmless drift | Slightly old headcount, missing newest product, fuzzy tagline | Log it. Fix at layers 0–1 in passing. No project. |
| Commercial damage | "Company closed" or "was acquired," wrong pricing, absent from category answers, confused with a competitor | Run this playbook now. |
| Defamation-grade | Invented lawsuits, fraud claims, sanctions, fabricated safety incidents | Run the playbook and brief a lawyer in parallel. Preserve everything. |
Be honest at this gate. Most errors are drift, and drift does not justify the effort below. The dangerous middle tier is commercial damage: statements false enough to change a buying decision but not false enough for a courtroom.
Week 1 — reproduce and document: the 10-prompt evidence log
AI answers are not reproducible by default; the same question yields different answers across sessions. So before fixing anything, freeze the evidence.
Build ten prompts:
- 4 brand basics — "What is [company]?", "Is [company] still in business?", "Who owns [company]?", "Where is [company] based?"
- 3 buying prompts — "[company] pricing," "[company] vs [competitor]," "best [category] providers"
- 3 adverse prompts — "[company] problems," "[company] lawsuit," "is [company] legitimate?"
Run all ten on every platform that matters to your buyers — ChatGPT, Gemini, Perplexity, Copilot, Grok — in clean or logged-out sessions, with web search toggled on and off where possible. For each answer, log: date, platform and model version, whether browsing was on, the full answer text, a screenshot, and the false claim verbatim.
This step is non-negotiable for two reasons. First, every escalation channel you may use later — a Wikipedia talk page, a vendor privacy form, a lawyer — will ask exactly what was said, where, and when. Second, this frozen prompt set becomes your re-test baseline. Without it you cannot tell a real fix from lucky sampling.
Week 1–2 — trace the source the model leans on
Modern assistants increasingly show their work. Use that.
- If the answer carries citations — ChatGPT in search mode, Perplexity, Copilot, AI Overviews — open every cited link. One of them usually contains your false claim or something close to it. That page, not the chatbot, is your fix target.
- A useful follow-up prompt: "Which sources say [false claim] about [company]?" Treat the reply as leads, not truth — open and verify each one.
- If there is no citation and the answer is wrong even with browsing off, you are dealing with training-data memory. There may be nothing on the live web to correct. The play is different: publish dense, retrievable, correct material so that search-grounded answers override memory now, and the next model snapshot learns the right facts later.
End the trace with a one-line verdict per false claim: claim → mechanism (memory, retrieval, or entity confusion) → source URL if one exists. That table drives the next section.
The Correction Stack: five layers in fix order
The Correction Stack is ordered by control: start where you have full control and same-day speed, end where you have neither. Most teams invert it — they hammer the "report" button and stop. Platform report tools exist, but they are unreliable as a primary correction channel (Mention Network): a feedback queue with no SLA, no status page, and no obligation to act.
- Layer 0 — your own pages. An about page that states founding year, ownership, status, and locations in plain sentences. A pricing page that shows the real price. An FAQ that answers the exact questions buyers ask, plus llms.txt for crawlers. Make the true fact the easiest sentence on the internet to retrieve. Cost: hours.
- Layer 1 — structured entity data. Wikidata statements with references, schema.org Organization markup (foundingDate, address, sameAs), and consistent facts across LinkedIn, Crunchbase, and business registries. This is the layer that untangles entity confusion — covered in depth in our Wikidata and knowledge graph guide.
- Layer 2 — Wikipedia accuracy fixes. The highest-weight source most models lean on, and the easiest one to damage by editing it yourself. Edit requests only — next section.
- Layer 3 — third-party authority sources. The stale article that seeded the claim: request an update or a correction note from the publisher. Refresh the major profiles. If the falsehood came from coverage of a namesake, one piece of accurate new coverage gives retrieval something better to quote.
- Layer 4 — platform feedback. Thumbs-down, report forms, privacy and legal channels. Submit once, attach your evidence log, expect nothing on a schedule.
The permanent version of this architecture — owned pages, entity layer, Wikipedia, monitoring as standing infrastructure rather than a one-off operation — is the AI Reputation Stack.
Per-platform levers: what each feedback channel can actually do
There is no form to update ChatGPT's information about your company. What exists per platform:
| Platform | Feedback mechanism | What it can realistically change | Realistic latency |
|---|---|---|---|
| ChatGPT | Thumbs-down / report on a message; OpenAI privacy portal for personal-data requests | Feedback is a training signal, not a ticket. Privacy requests cover individuals, not brands. Brand facts move when cited sources change | Source-driven: weeks. Baked-in memory: next model update. Feedback alone: no SLA |
| Gemini / AI Overviews | Per-answer "feedback" link; Google's legal removal tools; suggest-an-edit on a claimed Knowledge Panel | Overviews follow Google's index and Knowledge Graph, so source and entity fixes propagate on recrawl | Days to weeks after recrawl |
| Perplexity | Per-answer feedback; support email | The most retrieval-driven of the five — answers track the cited pages closely | Often days after a cited page changes |
| Copilot | Per-answer feedback; Microsoft's content concern form | Grounded in Bing's index; refreshing sources and requesting recrawl via Bing Webmaster Tools moves answers | Days to weeks |
| Grok | In-app response feedback | Leans on X activity plus the web; corrections travel through visible X presence and web sources | Opaque; no published process |
These latencies are operational observations, not vendor commitments — no platform publishes a correction SLA. Plan the six weeks around layers 0–3; the buttons are a free lottery ticket you buy once.
The Wikipedia and Wikidata lever, done compliantly
If a Wikipedia article repeats the error, fixing it is the highest-leverage move in the stack — and the easiest to get wrong.
What we will not do, and you should not either: edit your own article directly, logged out, or through an undisclosed account. Conflict-of-interest edits get reverted, tagged, and archived, and the cleanup becomes part of your public record.
The compliant path:
- Register an account and disclose your affiliation on your user page. If you are paid for the work, Wikipedia's terms of use require saying so.
- Post an edit request on the article's talk page: quote the incorrect sentence, state the correct fact, and attach an independent reliable source — not your own website.
- Keep it factual and minimal. Uncontroversial corrections — founding year, headquarters, current operating status, a leadership change — are routinely accepted. Requests that upgrade adjectives are not.
- For false claims about living people — a founder accused of something that never happened — Wikipedia's biography policies require fast removal of poorly sourced claims, which makes well-documented requests move quicker.
Wikidata is friendlier: affiliated editors may correct statements directly, provided each statement carries a reference and stays neutral. Audit your entity for a wrongly attached dissolution date — the classic root of "AI says the company closed" — merged or duplicate records confusing you with a namesake, and outdated officers or websites. Models and Google's Knowledge Graph read Wikidata as ground truth — the quietest high-leverage fix in the stack.
Latency honesty: talk-page requests take days to weeks depending on page traffic; the effect on AI answers follows the usual propagation window after that.
What not to do
- Do not spam the feedback buttons. Ten thumbs-downs from your team is noise to an anti-abuse filter, not a louder signal. Report an incorrect AI answer once, with evidence, and move on.
- Do not publish a rebuttal that repeats the false claim. A post titled "No, [Company] has not shut down" hands retrieval systems the exact string you want unlearned. State the true fact in headlines and structured data; rebut directly only when a falsehood is already spreading on its own.
- Do not lawyer-first over ordinary errors. A cease-and-desist about a wrong founding year has no lever to pull — there is no editor to compel and no single document to retract. Legal review is right when the output is defamation-grade: fabricated crimes, fraud, or regulatory violations, especially recurring across platforms. Then your evidence log becomes the exhibit file, and counsel can work the vendors' legal channels and the source publication carrying the claim.
- Do not declare victory after one clean answer. Sampling variance produces correct answers by accident. The re-test protocol exists for this.
The re-test protocol: weeks 2, 4, and 6
Re-run the same ten prompts, on the same platforms, in clean sessions, every two weeks — three runs inside the six-week window. Classify every false claim per run:
- Fixed — correct in two consecutive runs, with browsing both on and off.
- Partial — correct when the model searches the web, still wrong from memory.
- Unstable — alternates between right and wrong across runs.
Calibrate expectations against the propagation data: source-level corrections typically take about four to eight weeks to surface in AI answers (Sight AI). A quiet week 2 is normal. A partial at week 4 usually means your strongest sources updated but the entity layer still disagrees — re-audit layer 1. Still unstable at week 6 means the trace missed a source: go back to the citation step with fresh answers.
After week six, drop to a monthly cadence with the same frozen prompts. When manual logging stops scaling, or the board wants a single number, that is the moment for tooling — see our review of AI brand monitoring tools.
When it will not fix — and what monitoring buys you meanwhile
Honesty about the edges:
- No-source hallucinations. Some wrong answers cite nothing and match no document on the live web — a hallucination about your company that the model assembled itself. There is nothing to correct, because nothing was written. The only mitigation is density: enough consistent, structured, retrievable truth that search-grounded answers outweigh memory, until a newer model snapshot learns the right facts.
- Model-version lag. Your fix can land in the current search-grounded mode while older or offline modes stay wrong until the vendor ships an update. No amount of source work shortens a release cycle.
- Heavyweight namesakes. If you share a name with a much larger or more covered entity, confusion can resurface with every new model, even after a clean fix. Distinct entity data reduces the frequency; nothing eliminates it.
In those cases, monitoring is not a consolation prize. It catches regressions early, builds a dated record showing the falsehood persists on the vendor's side — which matters if things ever turn legal — and proves to leadership that the fixable part got fixed.
If you would rather start with the baseline done for you — what ChatGPT, Gemini, Perplexity, Copilot, and Grok each say about your company today, which sources they lean on, and which layer of the stack needs work first — that audit is how our AI Visibility service begins. We do not promise to change a model's mind on a schedule. We fix the sources it reads, then measure whether it listened.