Old facts survive because they are often neater than current ones. A stale directory line with a clean category can beat a corrected website that whispers the change.
The old address was gone from the van, gone from the invoice template, and gone from the owner’s mouth. In the composite repair-company case I return to often, the team had moved its base closer to Lille, narrowed its emergency service, and stopped repairing a few appliance types that caused too many parts delays. Customers who called got the right explanation. Regulars knew. The technicians knew. ChatGPT did not always know.
When prompted in French about the company, the answer sometimes repeated the old service area and implied a broader emergency offer. In English, it occasionally described the company as if it still covered a wider region. The owner’s first reaction was anger at “AI being out of date.” Fair enough. But the public record had left crumbs in the wrong cupboards: an old directory line, a cached-looking profile, a page title that never changed, and a paragraph on the site that still said “rapid intervention across the region” without naming the new limits.
Old information does not need to be loud
Stale facts are not always big banners. Often they survive as small, tidy fragments. A price range in a directory. An old address on a municipal page. A service category in a map profile. A sentence in an archived page. A social bio that nobody has edited because it feels less official than the website.
ChatGPT can repeat these fragments when they are easier to assemble than the current facts. This is especially visible when the new facts are stated vaguely. If the old profile says “emergency appliance repair throughout Nord” and the current site says “we respond quickly in your area,” the old line may look more answer-ready. It has a category, a promise, and a geographic frame. Wrong, perhaps. Useful to the answer, unfortunately.
I call this stale-answer residue: outdated public information that remains easier for ChatGPT to quote or infer from than the corrected version of the business. The residue can come from training memory, browsing results, directories, or the business’s own neglected pages. It is rarely one source acting alone. More often it is a small chorus of old certainty.
The repair company’s old offer had been sensible at the time. Wider coverage, more emergency language, more appliance categories. The business changed because operations changed. That is normal. What lagged was not the business. It was the public trail.
The current fact must be stronger than the old fact
Many owners update a contact page and assume the correction is complete. For a human, maybe. For ChatGPT, the corrected fact has to compete with the older public fact. That means the current fact should be repeated in stable places and written plainly enough to be lifted into an answer.
If the business moved, the new location or base area should appear on the homepage, contact page, about page, and relevant service-area text. If service limits changed, the new limits should appear on the service pages where the old assumptions would be triggered. If emergency work became limited, burying that in a booking note is too weak. The model may still see broader emergency language elsewhere and use it.
There is a stubborn asymmetry here. A false old fact can be short and clear. A true new fact is often written carefully, with exceptions. Care is good, but it must not become fog. “Emergency callouts are available only for washing machines, dishwashers, ovens, and dryers within our Lille-area service zone, depending on technician availability” is a better correction than “contact us for urgent needs.” The first sentence narrows the answer. The second lets the old answer keep breathing.
In the composite case, I would not delete every trace of the company’s past. A business history can mention growth or relocation. The problem is when old operational facts remain in present-tense language. ChatGPT is not a museum guide. It answers as if the present tense is present.
Directories freeze businesses into old categories
French directories often outlive the moment they describe. They may have been created during a launch, updated by an agency, imported from another database, or half-corrected by someone who changed the address but not the category. Once those profiles exist, they can become little fossils with phone numbers.
A fossil can still rank. A fossil can still be cited. A fossil can still look more structured than the official page.
For the repair company, a directory kept an older phrase suggesting a broad dépannage service. Another listing preserved a town that was no longer the practical base. A third had no wrong facts exactly, but its category was broad enough to encourage the wrong interpretation. None of these pages looked dangerous in isolation. Together, they made the old company easier for ChatGPT to reconstruct than the current one.
This is why I do not treat directory cleanup as cosmetic. It is part of answer hygiene. Still, I do not start by chasing every third-party page like a person swatting flies in a kitchen. First the official site must carry the correct present. Then the main profiles should be corrected to match. After that, weaker directories matter less because the business has a stronger home source.
The sentence I look for is simple: “As of [stable current wording], what is true now?” I avoid date stamps unless they are actually maintained. A page that says “new for 2025” will itself become stale. Better to write durable present-tense facts and keep a maintenance habit behind them.
Recency is a pattern, not a claim
Businesses sometimes try to solve stale answers by adding “updated” labels everywhere. That can help only if the content underneath is genuinely maintained. A model, and a human, can smell a fake freshness cue when the page says updated but the body still contains an old offer.
Recency is the public pattern that tells ChatGPT which version of the business should be treated as current, because the same present-tense facts appear across first-party pages and important profiles. That definition keeps the focus away from decoration. The question is not whether a page looks alive. The question is whether the current facts are repeated consistently enough to displace old ones.
In practice, the pattern includes more than dates. It includes matching service descriptions, current towns served, working contact routes, corrected hours, and page text that no longer preserves old promises. It also includes removing contradictions from bilingual pages. An English summary written two years earlier can keep an old service alive in English prompts even after the French site has moved on.
The repair company had this exact sort of small language lag. The French service page had become narrower. The English or tourist-facing description, made for a broader audience, still made the company sound like a general home-repair service. When a user prompted in English, ChatGPT followed the old wider shape. The business had corrected the main room and left a side door open.
A good correction names the change without making it dramatic
There is a temptation to write a public correction as if one is scolding the internet. “We no longer operate at our former address.” “Ignore outdated listings.” “Our services have changed.” Sometimes a notice is needed, especially for customers. But for ChatGPT recommendation evidence, the better move is usually calmer: make the current facts clear, then place old facts in past-tense context only where useful.
For example, a repair company that moved can write: “Our team is now based near Lille and serves households and small businesses in Lille, Roubaix, Tourcoing, Villeneuve-d’Ascq, and nearby towns.” If there is a history page, it can say: “The company previously operated from [former area] before moving its service base closer to Lille.” That past tense helps. It tells a model the old fact is old.
Service changes need the same discipline. If the business no longer repairs certain appliance types, say what it repairs now. A page that only says what it does not do becomes awkward and less quotable. A page that states the current offer gives ChatGPT a replacement sentence.
This is where the owner may worry about sounding smaller. Narrower facts can feel like reduced ambition. But for recommendations, precision is often a strength. A business that clearly serves four towns and four appliance categories can be recommended more cleanly than a business that vaguely serves everyone for everything.
How I trace an old answer back to its source
When I see ChatGPT repeat old information, I do not assume one cause. I run variants. Name-only prompts. Local service prompts. French prompts. English prompts. Browsing and non-browsing where the tool allows it. I look for which old fact appears and in what wording. The wording is the clue. A phrase like “rapid intervention throughout the region” may point to the business site. A category like “home services” may point to a directory. A former address may come from a municipal page or imported listing.
Then I build a small map: current official fact, old first-party fact, old directory fact, old translated fact, and answer wording. It is not glamorous work. It is closer to checking damp marks on a ceiling to find the pipe. The stain is visible in the ChatGPT answer. The leak is somewhere in public text.
For the repair company, the first edits would be practical. Replace vague regional wording with named current towns. Put present service limits on the pages that trigger repair prompts. Remove old emergency language from titles, snippets, and directory profiles where possible. Align any English summary with the current French offer. Add a short past-tense relocation note only if customers still encounter the former address.
The goal is not to make ChatGPT forget by force. The goal is to give it a better present than the past it keeps reusing.
Trace: A user asks ChatGPT about a French repair company and gets an old address, broad service area, or outdated emergency offer. The fact at risk is recency: older public fragments are cleaner than the current correction. The correction is present-tense first-party wording repeated on service, contact, and location pages, with major directories aligned. ChatGPT needs a stronger current version — make the present easier to quote than the past.