Key takeaways
- Answer engines volunteer comparisons, rankings, recommendations and caveats that the user did not ask for. That editorial layer is where most wrong brand claims appear.
- Independent audits find frequent errors: in the EBU and BBC study, 45% of AI assistant answers about news had at least one significant issue.
- Wrong claims usually have a page behind them: an expired landing page, an old comparison post or a stale listing. Engines repeat what they can read.
- Accuracy needs its own routine: list the facts that matter, check what engines say, trace each error to its source and fix it in order of control.
Visibility is only the first dimension
Ask an AI assistant whether a home-appliance rental service is worth it, and the answer rarely stops at yes or no. It compares the brand with two competitors you did not name, ranks them, explains why one is the better pick, mentions a fee and closes with a suggestion to consider something else. Each of those additions is a statement about a brand, and each can be right or wrong.
Visibility metrics count whether the brand appeared. They do not check what was said once it did. A brand that is mentioned often but described with an old price, a discontinued product or a competitor's weakness attached to its name has a visibility number that looks healthy and an AI presence that is costing it customers.
What engines add without being asked
The unrequested content in AI answers falls into a few recognisable types. Knowing them makes an audit faster, because each type fails in its own way.
| Type | What it looks like | How it goes wrong |
|---|---|---|
| Comparisons | Your brand set against competitors the user never named | Old specifications or prices on one side of the comparison |
| Rankings | Best, top or number one claims | Copied from affiliate round-ups with their own incentives |
| Rationale | Why the engine picked one brand over another | Reasons built on features or offers that no longer exist |
| Referrals | You might also consider another brand | Sends the buyer elsewhere on the strength of a stale review |
| Caveats | Fees, minimum terms and fine print | Terms from a previous contract or another market |
Why engines volunteer so much
- They are built to be complete. For a product question, complete means anticipating the follow-up. Ask about one brand and the engine assumes you will next ask how it compares, so it answers that too.
- Retrieval favours comparison pages. Round-ups and top-five posts are dense with the ranking language engines look for, so a question about one brand pulls in pages about five.
- Gaps are filled from memory. When retrieved pages disagree or say nothing about a detail, the model fills in from its training data, which is where ended promotions and old prices come from.
The engine does not only repeat what it reads; it completes the sentence for you. That is why a brand's own page is necessary but not sufficient: the answer is assembled from everything the engine can find.
What independent research says about accuracy
The largest accuracy audit so far is the EBU and BBC study News Integrity in AI Assistants, published in October 2025. Journalists from 22 public-service media organisations in 18 countries and 14 languages evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity. 45% of answers had at least one significant issue, 31% had serious sourcing problems and 20% contained major accuracy errors such as hallucinated or outdated details. The EBU described the failures as systemic and multilingual.
Earlier, Columbia University's Tow Center for Digital Journalism ran 1,600 source-identification queries through eight AI search tools in March 2025 and found they answered incorrectly more than 60% of the time, and that the tools rarely signalled uncertainty. Both studies looked at news, which is harder than questions about a single product. But they point in the same direction: answers are often wrong, and they rarely say so.
| Study | What was tested | Key finding |
|---|---|---|
| EBU and BBC, Oct 2025 | More than 3,000 news answers from four assistants, 14 languages | 45% with at least one significant issue; 31% with serious sourcing problems |
| Tow Center, Mar 2025 | 1,600 source-identification queries across eight AI search tools | Incorrect more than 60% of the time, rarely with a warning |
| Ahrefs, Jul 2025 | About 17 million citations from AI assistants | Cited pages were on average 25.7% fresher than organic search results |
The Ahrefs finding matters for the fix. Engines lean towards newer pages, so a dated, current statement of the facts has a real chance of replacing an old claim, provided someone publishes it.
Why this matters for Korean brands
Korean audiences are moving quickly. The Reuters Institute's Digital News Report 2026, published in Korea by the Korea Press Foundation, found that 14% of Koreans had used an AI chatbot for news in the past week, double the previous year and above the global average. People read the answer inside the chat, and few go on to check the page behind it.
The Korean web adds its own sources of stale claims: Naver Blog reviews written during a promotion, Naver Cafe threads that quote old terms, and marketplace listings that are rarely updated. All of them stay readable long after the facts change.
How a wrong claim survives: an illustrative example
Take an illustrative case. A rental brand ends a first-month-free promotion. It removes the banner from its homepage, but the landing page stays live at its old address with no end date, three comparison blogs written during the promotion still describe it in the present tense, and a community thread quotes the offer in its title. Nothing on the web says the offer has ended.
Months later, an engine asked which rental service to choose recommends the brand because the first month is free. Nobody asked about promotions; the engine volunteered it, and it was wrong. The cause was not a careless engine but a missing correction.
A routine for accuracy
Visibility can be checked monthly. Accuracy decays continuously, because every price change, promotion end and discontinued product creates a new false claim somewhere on the web. Run it as a routine, in this order.
- Build the claim set. Write down every fact an engine could state about you: prices, fees, minimum terms, delivery times, ingredients, certifications, launch and discontinuation dates. Give each a current value, an owner and a valid-from date.
- Audit the answers. Run the questions your buyers actually ask on ChatGPT, Gemini, Claude, Perplexity and Naver AI Briefing, including comparison and is-it-worth-it questions. Mark every claim about you as true, outdated, wrong or unverifiable.
- Trace each error to its source. Most wrong claims have a page behind them. Record the address and whether you control it.
- Fix in order of control. Owned pages first, marketplace and partner listings next, third-party content through outreach last. For claims you cannot get changed, publish a dated page that states the current fact plainly.
- Re-test and repeat. Ask the same questions again two weeks after each fix. A claim that comes back after a fix is a sourcing problem, not a content problem.
| Claim type | Where the error usually lives | Fix |
|---|---|---|
| Price and fees | Comparison blogs, old marketplace listings | A dated price page on your site, listing updates |
| Promotions and terms | Expired landing pages, community threads | End dates on old pages, redirects, a current-terms page |
| Product facts | Resellers, review sites, old press | Product pages with specifications as text |
| Discontinued lines | Cached retailer pages, old blog reviews | Discontinuation notices with a suggested replacement |
| Rankings and best-for claims | Affiliate round-ups | Outreach with current information, your own comparison page |
Write claims engines can repeat correctly
Generic copy gives an engine nothing to quote, so it fills the space with whatever it found elsewhere. Specific, dated statements give it something better to repeat. Compare, for example, a slogan such as the most trusted premium brand with a sentence such as standard delivery takes two business days from our Gimpo warehouse, as of August 2026. Only the second can be checked, quoted and kept up to date.
How flowgen helps
Brand Intelligence keeps the verified facts about your brand in one place and checks AI answers against them, so wrong and outdated claims show up as a list you can work through. GEO Insight shows which sources the engines cite for those answers, and Content Agent drafts the corrected pages from the same facts.
Sources
- EBU and BBC, News Integrity in AI Assistants, 21 October 2025
- Columbia Tow Center for Digital Journalism, AI search has a citation problem, March 2025, as reported by Nieman Lab
- Ahrefs, AI Assistants Prefer to Cite Fresher Content, 28 July 2025
- Reuters Institute, Digital News Report 2026, Korea findings published by the Korea Press Foundation, as reported by Newsis, 16 June 2026