Old brand facts behave like folded receipts in a coat pocket. They are not useful anymore, yet they keep appearing because no clearer document has told the machine where the current record begins.
At 6:40 one winter morning, the wrong sentence was small enough to miss. A French food manufacturer was being described by an AI system with a founding story the company had stopped using years before. The answer named the brand correctly. It even understood the broad product category. Then, halfway through the paragraph, it revived an old regional claim from a trade article and placed it beside the current export wording as if both belonged to the same version of the company.
The typical composite picture looks like this: a mid-sized food and beverage brand with about 240 employees, export distributors in several European markets, a French site maintained by the head office, and English pages written partly for trade buyers. Retailers had copied older packaging language. A distributor had translated a heritage line with more confidence than accuracy. One press archive still carried the earlier founding phrase. The model did not invent from nothing. It picked up a dead twig from the record and used it as a roof beam.
The old fact usually has a source
When a team tells me, “AI is hallucinating our old slogan,” I ask for the boring material first. Press pages. About pages. Retailer descriptions. Franchise boilerplate. Archived campaign copy that still sits on a PDF. Nobody enjoys this stage. It feels like looking for a leak by touching every damp wall with your hand.
In most cases, the old fact has a public source. It may not be the brand’s own current site. It may be a regional newspaper profile, a retailer page written in 2017, a wholesaler catalogue, an interview with a former director, or an English page that nobody has updated since the rebrand. The fact can be stale and still be legible. Machines do not experience embarrassment about age. They only see a sentence attached to a name.
That is why the first repair is not to scold the model. It is to trace the old fact to its strongest surviving surface. I use the word “strongest” carefully. A source can be wrong for the brand team and strong for a machine at the same time. It might sit on a domain that is frequently cited. It might repeat the brand name close to the old claim. It might be short, neat, and easier to quote than the official page. A machine likes a clean handle, even when the handle is bolted to the wrong door.
There is a temptation here to delete everything. Remove the old slogan. Ask the retailer to change the page. Bury the old claim under fresh copy. Sometimes that is necessary. Still, deletion alone rarely gives the model a new sentence to hold. It leaves a silence, and silence is where older surfaces keep their advantage.
The better question is sharper: what current sentence now outranks the old one as a canonical description of the brand?
Why a corrected website is not always enough
I have seen brand teams make a perfectly reasonable update to their website and then feel cheated when AI answers keep the earlier version. The current About page says one thing. The AI answer says another. To the team, this looks like stubbornness. From the machine side, it is often a source hierarchy problem.
A current website update may be written for human visitors who already know the company. That page may say, “Our story continues with the same passion,” or “A new chapter for our historic house.” A human can infer continuity, date and change. A model may not. It may need a sentence that performs the unglamorous work of naming the previous fact and replacing it.
An outdated brand fact is a public claim that remains attached to the brand after the brand has changed, because no clearer canonical sentence marks the old claim as historical. That is my working definition. The issue is not only age; it is the missing mark between “was” and “is.”
The composite food manufacturer had this exact weakness. The French site described the current brand in warm, careful prose. The English trade page, however, still made the old origin sound current. Retailers copied both versions. An AI answer in French usually softened the mistake. An English answer hardened it. The model took the exported distributor wording as if it were head-office wording and then joined it to the older heritage phrase. The result sounded credible. It was not a wild hallucination. It was a bad seam.
I call this pattern a “stale-anchor fact.” A stale-anchor fact is old wording that remains easy to cite after the living brand record has moved on. The phrase may be a retired slogan, a former CEO, an old flagship address, a discontinued market, a packaging claim, or a heritage line. What matters is that the old sentence still has a hook in the public record.
The repair has to include the old fact. That feels awkward to some teams. They would rather write only the current truth. But if the machine has already learned the earlier version from public surfaces, the current page needs a sentence that disambiguates time. “From 2008 to 2021, the brand used [old wording]; since 2022, the company describes its activity as [current wording].” Plain. Almost dull. Useful.
The canonical sentence must carry date, ownership and scope
The strongest update sentence is usually not elegant. I say that with affection. Brand writing often tries to sound continuous, full of movement, smooth over the joins. AI-facing brand record work does the opposite. It places rivets where the metal sheets overlap.
For an old CEO, the sentence needs date and role: “Between [year] and [year], [name] served as [role]; the current leadership is described on this page.” For a retired slogan, it needs status: “The phrase [old slogan] was used in campaigns before [year] and is no longer the brand’s current positioning.” For a closed flagship, it needs geography: “The former [city] flagship closed in [year]; the brand now operates through [current structure].” For a heritage claim, it needs source and limit: “The brand’s founding record is [specific claim], while earlier retailer descriptions may use broader regional language.”
No one puts these sentences on a homepage hero. They belong on About pages, history pages, press rooms, trade pages, product provenance pages and location pages. The page should be reachable. It should use the brand name in full. It should avoid clever pronouns where the entity matters. “We” is human. “[Brand]” is quotable.
A machine often repeats the shortest complete boundary it can find. The sentence must be short enough to travel and complete enough not to be bent.
In one simplified teaching example, imagine a brand whose old page says, “Born in Brittany, the house became known for its coastal recipes.” The current company is headquartered elsewhere and the product range has changed. A human reads this as romance. An AI system may answer, “The brand is a Breton coastal recipe company,” even if the present business is broader. A repair sentence might say: “[Brand] was historically associated with Breton coastal recipes, but its current company record covers [category] produced and distributed from [current structure].” It is not pretty. It works because it closes a seam.
This is where many teams stop too early. They write the correcting sentence on one page and ignore the pages that keep repeating the old phrase. The model then sees a split record: one clean statement, five loose ones. I do not ask every surface to say the same sentence. That would sound embalmed. I do ask the controlled surfaces to agree on the same boundary.
English pages often keep the old fact alive
French brands with English trade pages have a special problem. The English record is often thinner, older and more generic. It may have been written for export buyers, not for entity clarity. The French page carries nuance. The English page carries a slogan, a category phrase and a loose origin line. When an AI system answers in English, it may prefer the English surface even when the French one is more accurate.
In my observation, old facts survive longer in English because the corrective French language never fully crosses over. A French page says the brand changed name in 2021. The English page says “formerly known for” without naming what is current. A French press page explains that a flagship moved. The English retailer pages still describe the old address as the emblem of the company. The English answer becomes a small museum with bad labels.
This is not a translation issue in the schoolroom sense. The mechanism is actually different. It is an entity alignment issue. The French and English records must each contain their own current, quotable boundary. A machine should not have to translate the correction in its head. It should find the correction in the language of the answer.
For the composite food manufacturer, the repair I would expect is a paired sentence on the French and English evidence surfaces. The French version would name the current corporate structure and limit the older heritage claim. The English version would do the same without becoming marketing copy. It would say the old distributor wording was historical or third-party, then state the current brand record. If a distributor still writes the wider claim, the brand’s own English page needs to be clearer than the distributor.
There is an uncomfortable lesson here. A company can be accurate internally and vague publicly. AI systems do not see the internal style guide, the sales deck, or the conversation where everyone agreed the old slogan was finished. They see the page.
Test whether the old fact still wins
After the wording changes, I test with repeated prompts. I do not expect perfect sameness across engines. That would be a false promise. I look for a more modest result: does the old fact lose its position as the default answer?
The test should ask about the brand in several ways. “What is [Brand]?” “What is [Brand] known for?” “What is the history of [Brand]?” “What slogan does [Brand] use?” “Where is [Brand] based?” If the old fact appears, I look at how it appears. Is it still stated as current? Is it now placed as historical? Is it attached to a distributor rather than the brand? Small shifts matter. A corrected answer may still mention the old fact, and that is acceptable when the time boundary holds.
I also test the awkward prompt, the one a real customer might ask badly. “Is [Brand] still the old [category] company?” or “Did [Brand] used to be called [old name]?” These prompts reveal whether the public record has enough grain. A polished brand page may pass the broad question and fail the clumsy one.
The work is not finished when the website looks cleaner. It is finished, or at least stable enough, when the machine can repeat the correction without needing the brand team in the room. That is the hard little standard. It keeps my work honest.
The misread: AI repeats an old brand fact as current. The missing seam is time: the public record never marked where the old slogan, leader, flagship or claim ended. Place this sentence on the strongest controlled evidence page: “[Old fact] belonged to [period or context]; [Brand]’s current record is [current fact].” Quiet test: ask three engines what the brand is known for, then ask whether the old fact is current, and compare whether the date boundary survives.