On 7 September 2026, Carolyn Shelby, Principal SEO at Yoast and founder of CSHEL Search Strategies, published in Search Engine Journal an analysis of the factual errors generative AI makes about companies. When ChatGPT, Perplexity or Google's AI Overviews name a former executive, a discontinued product or an outdated price, the cause is most often an accumulation of conflicting versions of the same fact, published at different times and never linked to one another, even though the correct information exists somewhere online.
The facts
- Columbia University's Tow Center for Digital Journalism ran 1,600 queries on eight AI search engines, using excerpts from 200 articles by 20 news publishers: more than 60% of the answers were incorrect, from a 37% error rate for Perplexity to 94% for Grok 3, and the paid versions gave wrong answers with more confidence than the free ones (Columbia Journalism Review, 6 March 2025).
- A test by the vendor Searchable, reported by Search Engine Journal on 20 July 2026, found at least one false piece of information in the answers of ChatGPT, Gemini and Perplexity for 64% of UK high street retailers, most often a wrong postcode; among 165 small London businesses, 93% had at least one basic fact wrong or missing. The test comes from a provider of AI visibility tools and its full methodology has not been published (Search Engine Journal).
- Carolyn Shelby describes an unnamed company for which an AI asked about its current CEO could name any one of four former executives, each documented by pages still online. The person who runs the company today holds the title of SVP and General Manager within a group, no page links the old title to that role, and some introductory texts still describe them as CEO (Search Engine Journal, 7 September 2026).
Why does an AI still name former executives or discontinued products?
A generative AI builds its answer from the pages it retrieves for a given question, then merges them into a single text that looks final. When the question carries an outdated assumption, for instance "who is the CEO of this company" when the role no longer exists under that name, the retrieval system favours the pages that use exactly that vocabulary: old biographies, press releases, conference profiles and acquisition announcements, all still online. If the pages describing the current structure use another term without ever mentioning the old one, they never enter the set of retrieved sources. An AI cannot cite a page it has not found. Classic search let several competing pages sit side by side in the results and the user decided; the generative engine decides in their place.
The fix proposed by Carolyn Shelby is a sentence of bridge content that names the old role, explains why it no longer applies and gives the current equivalent. A sentence such as "Jane Smith is SVP and General Manager" leaves the old question unanswered; the complete version states that, following the acquisition, the company no longer has a standalone CEO and that Jane Smith now leads it in that role within the group. The same principle applies to a renamed product, a discontinued subscription plan, a merger, an expired certification or a changed service area. She then recommends tracing the chain of evidence behind the wrong answer rather than adding one more page: update, consolidate or redirect owned content, date and annotate what must stay online, and ask the directories and partner profiles that are meant to stay current to update their entries, without demanding the rewrite of a news article that was accurate when it was published.
This inventory, which she calls a brand claim audit, differs from the usual content audit built on URLs, traffic and rankings. For each important fact, it records the question a customer is likely to ask, the outdated assumption it may contain, the old terminology and its current equivalent, the canonical source of the approved fact, the places where old versions remain, the sources cited by AI engines in inaccurate answers, the action required and the team responsible for follow-up. The scope goes beyond HTML pages: media kits, downloadable sales materials, help-centre content, schema.org values, product feeds, app-store listings, speaker biographies, job listings and old subdomains. One last point of method: AI visibility tools measure whether a brand is mentioned, and that presence is falsely reassuring when the answer is wrong. A citation that names a former executive, repeats an old price or attributes a discontinued feature to the current product counts as a failure.
AIxH's view
This mechanism first affects companies that have changed their name, leadership, product range or prices in recent years, a frequent case on the Luxembourg market, where acquisitions, mergers of firms and rebranded outlets leave traces in directories, the press and old press releases. Our SEO agency in Luxembourg includes this inventory of conflicting versions in its audits, because Google reads the same pages as the AI engines; our GEO agency in Luxembourg then replays the questions your customers put to ChatGPT, Gemini and Perplexity to check what they answer, and writes the bridge content that links each old version to the current reality. The test takes five minutes: ask two or three AI engines who runs your company, what it sells and at what price. An outdated answer becomes visible immediately. A free GSO audit is then enough to establish the full picture, from a simple mention to a recommendation.