How to Fix AI Getting Your Brand Wrong (Old Prices, Wrong Features)

AI gets brands wrong by two routes: stale memory from training, and weak pages pulled in at answer time. Neither checks your current price list.

• Stale memory. Every model has a knowledge cutoff. Anthropic's list (October 2026) runs from August 2023 for Claude Opus 3 to June 2026 for Claude Sonnet 5.5, so a model in use today can predate your last pricing change.
• Weak retrieval. When an engine searches, Anthropic says answer quality depends on the sources it reads. An old review or comparison page that still ranks becomes the answer.
• Confident guessing. A September 2025 arXiv paper argues that training and evaluation reward guessing over admitting uncertainty. In a Columbia Tow Center test of eight AI search tools identifying news excerpts, ChatGPT misidentified 134 of 200 articles but signalled low confidence only 15 times. That was a news task, not brand facts.

Check both, because the fixes differ: no mention is a visibility problem, a wrong mention is a source problem.

In AEOeye's audit records, absence was the bigger problem: 157 of 892 answers named the audited brand, and only 1 of those 157 mentions was negative (144 positive, 12 neutral). The snapshot covers 127 self-selected audits of 27 brands from June 26 to September 7, 2026. Read it as a pattern in our data, not a market statistic or engine comparison: over half the answers are Claude's (aggregates, method).

Our free audit shows ChatGPT's real answers to three unbranded buyer questions, so it shows whether you are named, not whether your price is right. If AI never names you, start with why AI doesn't mention your brand. If it names you wrongly, keep reading.

Ask the same branded questions in every engine, log each wrong claim with the sources the answer shows, and match each claim to a page. A claim that matches a shown source is a retrieval error. A claim with no matching source is usually stale memory or a guess.

1. Write five branded questions: price, features, audience, alternatives, and whether a retired product still exists.
2. Ask each in ChatGPT, Gemini, Perplexity, Claude and Google AI, with web search on where offered.
3. Log the claim, engine, date and every linked source.
4. Open each source and find the wrong claim on it.
5. Mark claims with no source as memory errors. Those are the slow ones.

Start with your own site. Old pricing pages, archived PDFs and retired product posts are the one source you can fix today. Google says the Removals tool takes a page out of its results within a day, but the effect lasts about six months, so update or remove the page itself.

On a page you control, in plain text, and then on the few third-party profiles that describe you. Consistency matters more than any single listing, because conflicting sources make engines guess (entity SEO explains why).

1. Your canonical page. Put the current fact in the first sentence of your pricing, product or about page, with a visible date.
2. Make it readable. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, and Perplexity says PerplexityBot surfaces and links websites. In AEOeye's crawler access study (October 2026), 60 of 375 homepages, including 24 of 88 e-commerce sites, blocked or challenged our self-identified crawler. It is a convenience sample with one fetch each (data), but check your firewall logs.
3. Google's own surfaces. Google says knowledge panels are generated automatically from public information and are not created or deleted by hand, but verified owners can suggest edits and get a review within a few days. Local businesses can edit their verified Business Profile.
4. Wikidata and Wikipedia. Wikidata is not open to every brand. An item needs a sitelink, or a clearly identifiable entity with serious and publicly available references, or a structural need. Wikipedia strongly discourages conflicted editors from editing directly and lets them propose changes on talk pages.
5. Directories. Claim and correct G2, Crunchbase, LinkedIn and app-store listings.

State the current fact in the first sentence, date it, name the old value once, and keep the wording identical everywhere. Retrieval quotes passages, so each fact must make sense when lifted out of the page.

• Front-load. "As of October 2026, Acme Pro costs $49 per month, billed monthly."
• Name the old value in one dated line. "Before March 2026, Pro cost $29." Our view, untested: a whole page about the wrong claim risks becoming the passage that gets quoted.
• Mirror the question. Add a short FAQ asking "How much does Acme cost?" and answer in the first sentence.
• Add markup second. Google says Organization markup helps it disambiguate your organization, and schema.org's Offer type defines price, priceCurrency and priceValidUntil. But Google also says AI Overviews need no special optimizations beyond being indexed and snippet-eligible.

So fix visible text first; schema is cheap insurance, not the lever (structured data guide).

Every major product has a thumbs-down or report control. Use it once per wrong answer, but treat it as a signal: no help page we read promises a fix.

• ChatGPT: thumbs down under the message. OpenAI says reported domains and content may be reviewed by its Model Quality team, but that form covers Terms of Use and legal violations, so a stale price may not qualify. We read an archived copy of that page, so check current labels in the app.
• Google AI Overviews: thumbs down, then Report a problem. Google says the report includes your query and its search results.
• Gemini: Bad response or Send feedback, both requiring sign-in; a Report legal issue option sits under More.
• Perplexity: three-dot menu, then Report.
• Claude: thumbs down on the response.

Nobody publishes a correction timeline for brand facts, so any "2 to 8 weeks" claim is a guess, including the one on this page's previous version. What vendors do document is narrower: Google says crawling takes a few days to a few weeks with no guarantee, and OpenAI and Perplexity say robots.txt changes reach their systems within about a day.

Retrieval-backed answers can change once your page is recrawled. Memory-backed answers change only when a new model ships (how ChatGPT's two clocks work).

Measure it yourself: ask the same five questions weekly for six weeks and log which engines moved and which cited your page. Request a recrawl of the corrected page in Search Console.

Ask the owner to correct it, send a dated source, and if that fails, out-publish it. You cannot remove someone else's page: Google's Removals tool covers pages hosted on your own site.

• Press and blogs: email the author with the correct fact and a link to your canonical page.
• Wikipedia: if you are connected to the brand, disclose it and propose the change on the talk page.
• Review sites: claim the profile and edit it.
• Defamatory or unlawful claims: ChatGPT and Gemini both have legal-report routes, but take advice from a lawyer first.
• Nothing works: publish a fresher, consistent page for the same question and earn links to it. Stale pages stay quotable until something stronger replaces them.