INSIGHTS / SOURCES & MEASUREMENT

AI Brand Accuracy: When a Mention Gets Your Business Wrong

Being mentioned is only half the check. An AI answer can name your company warmly and still describe the wrong service area, certification or scope. Compare each answer with an approved record of your facts, then fix the errors that could mislead a buying decision.

8 min read

TL;DR

  • Count a mention as good news only after checking its facts against an approved record of your business.
  • Independent news studies found frequent errors. Journalists flagged a significant issue in 45% of 2,709 assistant answers. These studies did not test company facts.
  • Correct first the errors that could change a buying decision, such as a wrong scope, service area or certification.
  • Fix your own pages, ask third parties for corrections and recheck. No markup or prompt can guarantee that a wrong claim disappears.

Why a visible brand still needs a fact check

Most AI visibility reports count whether you appear. Was the brand named? Was the site cited? Neither tells you whether the answer was right. A buyer who reads “they hold ISO 14001 certification” or “they serve the whole state” may act on it, even if your records say otherwise.

Keep two kinds of confidence apart. Apparent confidence is how sure an answer sounds: fluent wording, a citation, no hedging. Business accuracy is whether its claims match your approved record. The first is not evidence of the second.

A February 2025 test by the Tow Center for Digital Journalism showed the gap. ChatGPT Search misidentified 134 of 200 news excerpts, yet signaled doubt only 15 times and never declined to answer.

What the error studies found

The broader study is the EBU–BBC News Integrity in AI Assistants report, published in October 2025. Journalists from 22 public service broadcasters in 18 countries checked answers from free versions of ChatGPT, Copilot, Gemini and Perplexity.

Of 2,709 answers, 45% had at least one significant issue: a problem that could mislead the reader or badly weaken the answer. The chart breaks that down by type of problem.

Answers with a significant issue, by review criterion% of 2,709 answers; criteria overlap
Any significant issue45%
Sourcing31%
Accuracy, including quotes20%
050100

EBU–BBC, May–June 2025. Refusals and unevaluated answers are excluded. Criteria overlap; they do not add up to 45%. EBU–BBC: News Integrity in AI Assistants (PDF).

Two studies, two different tasks

The Tow Center used a different task. It gave eight AI search tools excerpts from 200 news articles and asked for each article’s headline, publisher, date and web address: 1,600 queries in all. The tools answered more than 60% incorrectly, from 37% for Perplexity to 94% for Grok 3.

The two studies differ in task, tools and error rules. Compare them before quoting either rate.

Two studies, two different tasks
Point of comparisonEBU–BBC News Integrity (2025)Tow Center citation test (2025)
TaskAnswer real audience news questionsIdentify the source of a pasted news excerpt
ToolsFree versions of four assistantsEight search tools, free and paid
Period and sampleMay 24 to June 10, 2025; 30 shared questions; 271 journalistsFebruary 2025; 1,600 queries, each run once
What counted as an errorA problem journalists judged could seriously misleadA wrong or missing article, publisher or web address
Headline result45% with a significant issue; 20% with a significant accuracy issueMore than 60% incorrect; 37% to 94% by tool
Limit for B2B useNews topics, not company factsA matching task, not a typical buyer question; 200 queries per tool

Do not average or rank the two rates. They agree on one point only: fluent answers with sources can still be wrong.

Why the news error rate is not your brand’s error rate

Neither study asked about companies. News changes fast and draws many competing sources. B2B facts such as service areas, certifications and minimum orders change slowly but may appear on only a few pages, some of them out of date. Errors about your business could be rarer, more common or simply different.

The tools have also changed. Between the EBU test and its report, ChatGPT’s default model changed. So treat these studies as a reason to test your own facts, not as a measure of your risk. Your own sample error rate comes from checking saved answers against your approved records.

Which errors deserve a correction?

Not every flawed answer needs action. Rank errors by what a buyer might do with them. This ranking is our editorial advice, not an industry standard.

Which errors deserve a correction?
Error typeExamplePriorityWhy
Wrong scope or capabilitySays you install systems you only maintainCorrect firstBrings the wrong leads and sets false hopes
Wrong service area or marketSays you serve regions you do not coverCorrect firstBuyers may shortlist or reject you on location
Wrong certification claimNames a certification you do not hold, or one that has lapsedCorrect firstBuyers often screen suppliers on these claims; involve the person who owns your certificates
Outdated offerDescribes a discontinued service as currentCorrect soonCreates requests you cannot fulfill
MisattributionCredits your finding to a competitor, or theirs to youCorrect soonWeakens trust in your own evidence
OmissionLeaves out a newer serviceTreat as a content gapNot an error; improve the pages that explain the service

Worked example — fictional: a fact incident log

A fictional company that services heating and cooling systems in office buildings checks 20 buyer questions in two assistants. Three answers contain errors worth logging. The table shows how each incident is recorded, from the wrong claim to the repeat check.

Worked example — fictional: a fact incident log
IncidentWhat the answer saidSource or possible contextApproved recordFix and ownerRepeat check
Service area“Serves the whole state”Cited: a trade group’s directory page from 2021Service map approved this quarter: three named metro areasOperations updates the service-area page; marketing asks the trade group to edit the listingSame questions two weeks after the listing changes; log the result even if unchanged
Certification“ISO 14001 certified”No citation. An old press release is possible context, not a proven source.Certificate register: ISO 9001 onlyQuality manager confirms; marketing adds a page naming each standard and its scope, plus a dated note on the press releaseMonthly for three months, then quarterly
Scope“Installs industrial refrigeration”Cited: a third-party comparison blogServices page: repair and upkeep only; no new installsMarketing adds a clear “what we do not do” line; the account lead sends the blog author the correction with evidenceAfter the author replies; if they decline, record the source as unresolved

Each entry keeps the saved answer, prompt, platform, date and cited link, so a later check compares like with like. An incident stays open until the team can read the corrected answer or source itself. An unchanged answer after a fix is a result to report, not a failure to hide.

What you can and cannot change

Your own pages are fully in your control. State each important fact clearly, in text, on the page a buyer would visit. Remove or date old versions. Structured data can repeat those facts. But our schema evidence review found no reliable proof that markup alone changes citations, and it cannot force an answer to drop an error.

Third-party sources are not yours to edit. Ask the publisher or directory owner for a correction, provide the evidence and accept that some will decline. Record those sources as unresolved rather than claiming a fix.

AI platforms offer feedback tools in some products, but a report does not guarantee a change. Correcting a claim in one conversation does not establish that the answer has been corrected for other users. For the full diagnostic steps behind a wrong claim, see how AI search cites sources.

Make accuracy part of routine monitoring

Keep a short approved fact record: services and exclusions, locations, certifications with dates and contact routes, each with a named owner. Check a fixed set of questions on a schedule. Label each answer correct, wrong or unverifiable. Count each distinct error once, and count the answers it affects separately. The counting rules are in our AI visibility measurement guide.

AI brand accuracy: common questions

Can we make an AI tool remove a wrong claim about our company?

Not directly, and not with certainty. You can correct the sources you control, ask third parties to correct theirs and use platform feedback where it exists. Then recheck the same questions and report what actually changed.

Is a positive but inaccurate mention still worth counting?

Count it as a mention, and log it as an error. A friendly answer that misstates your scope can bring unsuitable inquiries or cost you a bid. Report visibility and accuracy side by side, so one does not hide the other.

Do the news error rates apply to our business?

No. The studies tested news questions with specific tools at specific times in 2025. Use them as a reason to test, then measure your own error rate against your approved facts.

How often should we recheck our facts?

Check after any fix, after a real change to your business and on a regular schedule your team can keep. Use the same questions and platforms each time, so the results can be compared.

Check how answers describe your business

If buyers might act on what AI answers say about your services, coverage or credentials, start with a structured check. SemShift’s AI visibility audit reviews an agreed set of buyer questions. It examines how your brand is described and which sources support each answer. The findings go into a prioritized plan that separates confirmed errors from open questions.