When AI Gets Your Business Wrong: The Correction Playbook

Wrong beats missing: why false claims are the expensive gap

AI answers arrive with the confident, unhedged tone of a colleague who has already checked. That tone is doing a lot of work it has not earned. When the Tow Center for Digital Journalism tested eight AI search tools on identifying the source of news excerpts, the tools collectively returned incorrect answers to more than 60% of queries, with Perplexity wrong on 37% and Grok 3 wrong on 94% — and they rarely signalled uncertainty when they were wrong.

Provenance of news excerpts is not the same task as summarising a vendor, but the failure mode transfers exactly: a specific, checkable fact is asserted fluently and without a hedge. For a buyer in a comparison prompt, that is indistinguishable from research.

The practical consequence is that a false claim compounds. Assistants cite each other's ecosystem — one stale directory listing feeds a summary, that summary gets scraped, and the error acquires a second apparent source. Correcting early is much cheaper than correcting after the claim has been laundered through three intermediaries. Treat false claims as the top of your AI visibility queue, above thin mentions and missing comparisons.

• Step 1 — Pin down the claim before you try to argue with it
• paragraphs [],

Step 1 — Pin down the claim before you argue with it

Vague reports (“ChatGPT says our pricing is wrong”) cannot be traced or verified. Log the claim like a bug: the exact prompt text, the assistant and model, whether web search or browsing was active, the date, and the full answer including any citations. Screenshot it, and copy the text so you can search it later.

Two conditions matter and are usually skipped. First, run the prompt in a fresh session with memory and personalisation off — otherwise you may be reading back your own earlier conversation rather than the model's default view of you. Second, test the phrasings buyers actually use, not the ones you would: “how much does X cost”, “is X still available”, “X vs Y”, “is X the same company as Y”. False claims cluster in buying-intent prompts because those prompts demand specifics.

Then classify the error, because the fix differs. A retrieval error means the assistant cited a live page and the page is wrong. A parametric error means the claim came from training data with no citation. A conflation error means two entities were merged. Most pricing complaints are retrieval errors. Most same-name complaints are conflation.

Step 2 — Trace the citation, not the model

Ask directly: “Which sources support that?” Then open every URL and check whether it actually says what the assistant claimed. This single step resolves the majority of cases, and it splits them into three very different repairs.

If the source is real and the information on it is genuinely stale, you have a content problem you can fix today. If the source is real and correct but the assistant misread it — price buried in an image, a footnote read as the headline number, a legacy tier read as current — you have a clarity and structure problem. If there is no source at all, or the cited URL does not contain the claim, you are looking at model memory or a fabrication, and no single page edit will resolve it.

It helps to know that the retrieval layer is a live crawl you can influence. OpenAI documents separate crawlers for training and for search answers in its bots documentation, Anthropic publishes how Claude fetches web content, and Perplexity documents its crawlers and user agents. What those crawlers can read on your page today is the material for tomorrow's answer.

Step 3 — Fix it at the source, in order of leverage

Start with your own site, because it is the only source you control outright. Give every contested fact one canonical page that states it explicitly, in HTML text, with a date. Contradictions inside your own domain are a common cause of AI errors: an old campaign landing page, a PDF datasheet, a help-centre article and a changelog entry all quoting different numbers. The model has no way to know which one you consider current, so it picks — often badly.

Then make the fact machine-readable. Product and offer markup lets you state price, currency and availability unambiguously; Google documents the required and recommended fields in its merchant listing structured data guide, and schema.org defines availability values including Discontinued. Prices that exist only inside a screenshot or a client-side widget are invisible to most crawlers.

Next, third-party datasets. Wikidata and Wikipedia disproportionately shape how entities are described; if you edit there, follow Wikipedia's conflict of interest guidance and use an edit request on the talk page rather than editing the article yourself. Update your Google Business Profile, correct industry directories and review-site vendor pages, and ask resellers and partners quoting old prices to refresh their listings.

Finally, invite a recrawl. Update sitemap lastmod values and ping the changed URLs through IndexNow, which Bing supports — relevant because Copilot's web answers lean on Bing's index.

The three corrections that go wrong most often

Pricing, sunset products and name collisions account for most of the damaging false claims we see, and each has a specific failure pattern.

Pricing errors persist because old numbers live in more places than current ones — press coverage, reseller pages, review-site comparison tables, your own archived pages. Publishing a new price does not delete the old ones. Fix it by making one pricing page unmistakably canonical and dated, expressing the figures in text with offer markup, and chasing the top three third-party pages that still quote the old number.

Discontinued products fail in the opposite direction: teams delete the page. A 404 removes your ability to state the fact, so assistants keep answering from cached third-party descriptions and stale reviews. Keep the URL live, label the status and the end-of-sale date in plain text, mark availability as discontinued in your markup, and name the successor product on the same page. You want the assistant to say “discontinued in 2024, replaced by X” — that only happens if a crawlable page says it.

Name collisions are an entity problem, not a content problem. Two companies sharing a name get merged unless the graph separates them. Use one consistent legal name plus a category descriptor everywhere, cross-link your verified profiles as sameAs identifiers, keep your Wikidata item accurate, and — where the confusion is costing you deals — publish a short, factual disambiguation line stating that you are not affiliated with the other entity.

Step 4 — Verify the answer actually moved

A fix you have not verified is a hypothesis. Re-run the exact logged prompts, in fresh sessions with memory off, across every assistant where you found the claim, and record the date each answer changes. Test the prompt variants too — assistants frequently correct the direct question while the comparison prompt still carries the old number.

Expect asymmetric timelines. Answers grounded in live retrieval can change within days of a recrawl. Claims baked into model weights may survive until a model update, which is why broad corroboration across independent sources beats a single perfect page. If the wrong claim only appears when web access is off, you are waiting on training data, not on your CMS.

Measure the share of correct answers over repeated runs rather than one screenshot, because outputs vary between sessions. This is the loop Growli automates for our customers — the same prompts re-run across ChatGPT, Gemini, Claude, Perplexity, Copilot and Grok, with an alert when a corrected claim reappears; how it works walks through the mechanics.

When the answer refuses to move

Use in-product feedback on the specific response — every major assistant has a thumbs-down and report path, and it is the only channel that reaches the platform with the offending output attached. Keep your bug log; a report with the exact prompt, date and correct fact is far more actionable than a general complaint.

Where the false claim concerns an identifiable person — a founder credited with something they did not do — data-protection law may apply directly. In the EU, the right to rectification covers inaccurate personal data, and providers publish privacy request routes for exactly this. Reserve legal escalation for claims that are defamatory or commercially damaging, not for tone.

One tactic to avoid: blocking AI crawlers as a response to a wrong answer. It removes your pages from the retrieval layer without removing anything already in the model, so the assistant answers from third-party sources you control even less. Being the clearest, freshest, most machine-readable source about yourself is the durable correction.

FAQ