Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 12 min

When AI gets your company wrong

When ChatGPT, Perplexity, or Google AI Overviews state something wrong about your company, the cause is usually a source, not a hallucination: an old pricing page, a third-party review that was never updated, a directory profile with the wrong category, or a namesake company mixed into yours. You cannot edit the answer. You can find the claim, trace it to the page it came from, correct that page, and measure whether the answer changes. As of November 2026, every major engine offers a feedback button and none offers businesses a way to correct answers directly.

In short

  • Most wrong AI claims trace to a page someone can fix. In Profound’s August 2026 data, inaccurate claims came mostly from competitor and earned-media pages, and 54% of brands had at least one inaccurate claim that cited their own content.
  • Pricing is the most error-prone topic. It was 12% of claims Profound evaluated and 24% of the inaccurate ones.
  • Audit with a fixed prompt set about your brand, run repeatedly, with every claim logged next to its source URL and an error type.
  • Fix by source: your own site first, then entity profiles, then the third-party pages engines actually cite. Feedback buttons are a last step, not a strategy.
  • Some errors come from training data or from pages you do not control, and those change slowly or not at all. Say so in the report.
  • Score accuracy as a number. A mention that states the wrong price is not a win, and share of answer should count it separately.

Why does AI state wrong facts about a company?

AI states wrong facts about a company for four main reasons, and each one points to a different fix.

  1. Training data is old. A model answering from what it learned in training knows your company as it was before its cutoff. If you changed pricing, renamed a product, or exited a market since, an answer without live search will repeat the old version.
  2. Retrieval found a stale page. When the engine searches, it can retrieve a review, comparison article, or forum thread that was accurate once and never updated. OpenAI’s ChatGPT search help notes that search results and citations can be incomplete, outdated, or incorrect. The citation is real. The fact is old.
  3. Sources conflict. If your site says one thing, your LinkedIn page another, and a directory a third, the model has to choose, and it may choose the wrong one or blend them.
  4. The entity is ambiguous. A company that shares a name with another business, a product, or a person can have facts from the namesake attached to it. This is the confusion that consistent entity data is meant to prevent.

Which of these applies depends on whether the engine searched. ChatGPT decides per question whether to search the web, and OpenAI does not publish the rule. Semrush’s clickstream analysis, as reported by MediaPost in early 2026, found ChatGPT used web search on about 34.5% of queries in February 2026. On the rest, the answer comes from the model alone. Perplexity and Google AI Overviews are search-grounded by design, so their errors usually point to a retrievable page. The full retrieval picture is in what query fan-out is.

Where do inaccurate AI claims come from?

They come mostly from other people’s pages, and a surprising share from the brand’s own. Profound published the largest public breakdown on August 17, 2026, from its FactCheck product: more than 158,000 claims, a snapshot as of August 14, 2026. Its findings:

  • Inaccuracy rates ranged from 4.3% to 6.8% depending on the engine.
  • Inaccurate claims came “mostly from competitor and earned media citations.”
  • 51% of brands had at least one inaccurate claim citing earned media, and 54% had at least one citing their own content.
  • Pricing and billing were 12% of claims evaluated and 24% of inaccurate ones. Among brands with at least ten pricing claims, pricing was the highest-error topic for 66%.
  • Errors were spread across a long tail. The top 10 domains accounted for 13% of earned-media inaccuracies and the top 100 for 40%. For a company with inaccurate earned media, the errors typically came from just four websites, often unique to that brand.

Read it with Profound’s own caveats. The data is observational and a single snapshot. Pricing may be over-represented because structured facts like prices are easier to check. Profound sells an accuracy product, so it has a reason to find errors. Even so, the practical point holds: the list of pages causing your errors is likely short, specific to you, and partly on your own domain.

How do you audit what AI says about your brand?

Run a fixed set of brand prompts on a fixed panel of engines, repeat each run, and log every factual claim beside the URL it cites. The method is the same one used in measuring AI search visibility, pointed at your own company.

  1. Write the prompt set. Twenty to forty prompts a buyer, analyst, or candidate would type. Cover what you do (“What is [company]?”), pricing (“How much does [product] cost?”), fit (“Is [company] good for mid-size manufacturers?”), comparison (“[company] vs [competitor]”), and facts (headquarters, founding year, leadership, integrations, compliance). Include the namesake test: ask about your name without context.
  2. Choose the engines your buyers use. Typically ChatGPT, Google AI Overviews and AI Mode, Perplexity, Copilot, and Gemini. Note the model or mode, because answers differ between them.
  3. Run each prompt more than once. The same prompt returns different answers run to run. Three runs per prompt per engine is a practical floor for a brand audit. A single screenshot proves nothing in either direction.
  4. Split each answer into claims. One row per factual statement: “starts at $49 per user per month,” “founded in 2015,” “integrates with SAP.”
  5. Log the claim, the cited URL, and the truth. If the answer cites a source for that sentence, record it. If not, mark it as uncited, which often means it came from the model’s training.
  6. Classify the error type. Use the table below, so the fix list sorts itself.
Error typeExampleLikely originFirst place to look
Outdated factLast year’s price or plan namesStale owned page, review profile, or training dataYour old pricing URLs, review-site profiles
Discontinued productRecommends a product you retiredOld product pages still indexed, old reviewsYour site’s redirects, comparison lists
Wrong positioningCalls you an “SMB tool” when you sell to enterpriseThird-party category labels, directory profilesReview-site category, LinkedIn, Crunchbase
Namesake confusionAttributes another company’s founder or HQ to youAmbiguous entity dataKnowledge panel, Wikipedia, Wikidata, sameAs
Wrong comparisonSays a competitor has a feature you have and they lackCompetitor content, comparison listsThe competitor’s page and the lists citing it
Fabricated detailAn integration or certification that never existedNo source, model inferenceUsually uncited; feedback and clearer owned facts
OmissionLeaves out your main product lineThin coverage, not an error in a pageThird-party corroboration

Omission is not inaccuracy, but log it in the same pass. It is the problem covered in why AI cites what others say about you.

How do you fix a wrong claim at the source?

Fix the page the claim came from, in order of how much you control it.

Start with your own site

Publish one canonical facts page in plain HTML: what the company does, who it serves, product names, current pricing model and starting prices, headquarters, founding year, leadership, and a “last updated” date. Link it from the footer. Then find every older page that contradicts it: retired pricing pages, old press releases with outdated product names, PDFs of last year’s rate card, and archived blog posts announcing features you removed. Update them, redirect them to the current page, or mark them clearly as historical.

Keep the facts in text. Prices in an image, a PDF, or a JavaScript widget are hard for engines to read, and most AI crawlers do not run JavaScript. If a distributor’s HTML table is easier to parse than your PDF, the distributor becomes the cited source for your price.

Make the entity consistent

Add Organization structured data with one stable @id and sameAs links to the profiles you control, as described in structured data and AI search. Schema does not force a correction. It helps machines tie the right profiles to the right company, which is the main defense against namesake confusion. It works only when the profiles it points to say the same thing as the page.

If Google shows a knowledge panel for your company, claim it. Google’s knowledge panel help describes the process: click “Claim this knowledge panel,” then verify through an official site or profile such as Search Console or YouTube. Once verified, you can suggest changes to specific facts through the panel’s feedback link, with a source for the correction. Local businesses use Google Business Profile instead.

Update the third-party pages engines cite

Work from your audit log, not from a list of every directory. Start with the domains that appeared as sources for wrong claims.

  • Review sites and directories you can edit. Update plan names, pricing, category, and company size on G2, Capterra, Crunchbase, LinkedIn, and similar profiles.
  • Articles and comparison lists you cannot edit. Email the editor with the specific error, the correct fact, and a public page that confirms it. Ask for a correction, not a rewrite. Editors fix factual errors more readily than they add vendors.
  • Wikipedia. Do not edit your company’s article yourself. Wikipedia’s conflict-of-interest guideline says COI editors are “strongly discouraged from editing affected articles directly.” Propose the correction on the article’s talk page, disclose your affiliation, and cite an independent source.
  • Competitor pages. If a competitor’s comparison page misstates your product, ask them to correct it, with evidence. Keep the request factual.

Use the feedback channels each engine offers

Feedback buttons exist on every major engine. None of them promises a correction for a business.

EngineWhat existsWhat it does not do
ChatGPTThumbs-down icon below any response; “Report” for content that may break OpenAI’s terms or the law. OpenAI’s privacy portal handles requests from individuals about personal data in responsesNo business correction channel. The privacy route covers people, such as a named executive, not company facts
Google AI OverviewsThumbs-down, then “Report a problem” with a category and details; also the three-dot menu on the overview. Google’s help page says “AI Overviews can and will make mistakes”No promised fix or reply. Knowledge panel feedback is a separate, verified route for entity facts
PerplexityFlag icon below an answer, or a support ticket with the query URL, a description of the error, and the expected result. Outdated information and misinformation are listed issue typesPerplexity’s publisher program shares ad revenue with participating news publishers (per TechCrunch, December 2024). It is not a correction channel

Use feedback after fixing the source, and include the source URL with the correct fact. Feedback alone, against a page that still says the wrong thing, is unlikely to stick, because the next retrieval finds the same page.

What can’t you fix directly?

You cannot edit an AI answer, remove a fact from a model’s training data, or make an engine re-crawl a page on your schedule. Four limits deserve a line in the report so nobody expects otherwise:

  • Training-data errors fade slowly. If an error appears without a citation, it may persist until a newer model ships, and you do not control when that happens.
  • Pages you cannot edit stay up. A stale article from a publisher that does not answer will keep being retrieved. The counter is a newer, more authoritative page that states the correct fact, not a takedown.
  • Answers vary. A fix can show in one engine and not another, or in two runs of three. Report rates, not single results.
  • Engines change. A model update can introduce or remove an error with no change on your side. That is why the audit repeats.

What you can control is how consistent the record is. When your site, your profiles, and the pages that cover you agree, an engine that retrieves any of them gets the right answer.

How do you track accuracy as a number?

Score it as the share of logged claims about your company that are correct, by engine, on the same prompt set every cycle. Accuracy is one of the four separate numbers in a GEO readout, next to mention rate, citation rate, and link rate, as set out in what gets cited. Keep them apart. A month in which mentions rose and accuracy fell is not a good month.

A workable scoring rule:

  • Accuracy rate = correct claims ÷ all factual claims checked, per engine.
  • Critical error count = wrong price, wrong product availability, wrong compliance status, or wrong company identity. Track these by name, not only as a rate.
  • Source-fixed rate = the share of wrong claims whose cited source has been corrected. This measures your team’s work, separately from whether the engine has caught up.

Run the full prompt set monthly and the critical prompts (pricing, identity) weekly, and keep the log, so you can show when each error first appeared, when the source was fixed, and when the answer changed. The platform keeps these scores in the same record as rankings and citations, because an accuracy problem is often a retrieval problem on the same query.

Worked example (illustrative)

This example is invented to show the method. The company, prompts, and counts are not client data.

A B2B scheduling software company retired its “Starter” plan in March and raised its entry price from $29 to $39 per user per month. In June, sales reports that prospects keep quoting $29. The team runs 30 brand prompts on four engines, three runs each, and checks 412 factual claims.

FindingClaims affectedCited sourceFixRe-check in August
Entry price stated as $2941The company’s own 2025 pricing page, still live at an old URL; a review-site profileRedirect the old URL; update the review profile9 remaining, mostly uncited
“Starter plan” recommended17Two comparison listsCorrection requests to both editors; one updated6 remaining
Described as “for freelancers”12A directory category labelCategory changed1 remaining
Founding year and HQ of a same-name agency5UncitedOrganization markup with sameAs; knowledge panel claimed; feedback filed4 remaining

Accuracy rose from 82% (337 of 412 correct) to 95% across the same prompt set two months later. The remaining pricing errors were mostly uncited, which points to training data, and the namesake confusion barely moved. The report says both plainly: the sources are fixed, and two error types are now waiting on the engines. That honesty is the difference between an accuracy program and a screenshot.

What are the common mistakes?

  • Treating every error as a hallucination. Most trace to a page. Find it before filing feedback.
  • Fixing the new page and leaving the old one live. The retired pricing URL is often the cited source. Redirect or update it.
  • Running each prompt once. Answers vary. One run can show an error that is rare or miss one that is common.
  • Editing your own Wikipedia article. Propose the change on the talk page and disclose who you are.
  • Expecting schema to correct an answer. Structured data clarifies the entity. It does not override a source that says something else.
  • Blending accuracy into visibility. Report accuracy as its own score, and name critical errors individually.
  • Promising a fix date. You control the sources, not the engines. Report what was fixed and what is still waiting.

Brand accuracy is the extraction and citation steps of the same system that SEO, AEO, and GEO share: the engine retrieves a page, lifts a claim, and repeats it. Fix the page and the claim follows, eventually. An accuracy audit on your own prompt set is part of a GEO diagnostic.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit