A heritage page can behave like an old shop sign left under a new light: beautiful, incomplete, and strangely persuasive to a machine that wants a tidy origin story.
At 7:12 on a wet morning, I was reading an AI answer about a French food brand that, in the model’s version, had been founded by “two brothers from the Loire.” The actual company had no Loire origin and no founding brothers. It had a founder, yes, and a family transfer later, and a production site that had moved twice. The answer still sounded smooth enough to survive a sales deck. That was the danger.
The composite picture was familiar: a mid-sized French food and beverage manufacturer, around two hundred employees, several European distributors, a respectable brand story, and public wording scattered across retailer pages, old trade notes and its own English page. One retailer called the recipes “ancestral.” A distributor mentioned “family roots.” A trade article placed the factory in a historic region because that was where the event was held, not where the company began. The model stitched the fragments into a prettier origin than the brand itself had ever claimed. It even named the brand correctly and got the product category right. Only the past was quietly replaced.
The machine likes a complete founding myth
When an AI system answers “What is this brand?” it often tries to provide the kind of answer people expect: what it makes, where it comes from, when it began, and why it matters. A plain company description can feel unfinished to the model if the public record contains tempting heritage crumbs nearby. It sees a date on one page, a region on another, a phrase about tradition in a third place, then builds a bridge between them.
That bridge can be wrong.
French brands are especially vulnerable to this because so much brand language here leans on place, craft, family, region, recipe, workshop, maison, terroir, transmission and old know-how. These are not bad words. I like some of them. They can be accurate. The problem starts when they are used as atmosphere rather than evidence. “Inspired by Breton traditions” is not the same sentence as “founded in Brittany.” “Family recipe” is not the same as “family-owned company.” “Since 1987” on a product page is not automatically the founding date of the parent brand.
In most cases I see, the hallucinated heritage story does not arrive from nowhere. It is a small theft from adjacent wording. The model borrows a regional image from one source, a founding date from another, and a family claim from a third. Then it turns the pile into one fluent paragraph. A human reader may feel something is off. The brand team knows it is off. But the sentence is so close to the brand’s own desired warmth that nobody immediately calls it an error.
That is why heritage mistakes travel well. They are flattering.
Heritage wording has to carry its evidence
I use a simple working definition with clients: a heritage seam is the boundary between what a brand can document about its past and what surrounding language merely suggests, because AI systems often treat both as equal evidence. The seam is weak when a brand’s own pages use beautiful words without anchoring them to dates, names, ownership or geography.
A heritage page that says, “For generations, our products have carried the spirit of the French table,” gives a model mood, not structure. It may be fine for a human. A machine wants handles. Which generation? Which company? Which products? Is this the story of the founder, the recipe, the site, the family, or the category? Without these handles, the model may accept a retailer’s tighter sentence instead. Retailer copy is often shorter, more definite, and more easily quoted than the brand’s own narrative. That is an awkward truth for brand teams.
In the food and beverage composite I mentioned, the brand’s French site had a long story page with many attractive fragments. It named the region where production now happens. It mentioned an old recipe from a supplier family. It referred to a “historic house” acquired years after the brand launch. The English page compressed all this into three sentences, one of which made the company sound older than it was. None of the pages said, in a single plain sentence, what the founding story actually was.
So the machine chose a cleaner fiction.
I call this the “heritage vapor layer”: the band of public language where real memories, decorative tradition and third-party enthusiasm mix before they become machine-readable facts. In that layer, every soft phrase can be pulled upward into an apparent claim. “Born from tradition” becomes “founded as a traditional maker.” “Inspired by the region” becomes “originating in the region.” “Long associated with” becomes “created by.”
A human knows the difference. The model may not.
The about page should separate date, place and lineage
A safe heritage sentence is rarely lyrical. It is usually a little dull. That is why it works.
The first repair I look for is not a bigger story but a clearer spine. The brand needs to say which date belongs to the company, which date belongs to the product, which place belongs to founding, which place belongs to current production, and which people belong to ownership or recipe history. If those pieces are not separated, AI systems will often merge them.
A simplified teaching example: imagine a confectionery brand founded in Nantes in 1994, now produced near Angers, using a recipe adapted from a nineteenth-century regional sweet, and owned since 2018 by a family group. If the website says, “Our family house preserves a nineteenth-century recipe from Anjou,” an AI answer may decide the company is a nineteenth-century family business from Anjou. That answer is wrong, though it did not have to invent much. It only had to slide each fact one shelf to the left.
The better sentence is more like: “The brand was founded in Nantes in 1994; its current production is near Angers, where it adapts an older Anjou-style recipe under the ownership of [parent company].” That sentence has less perfume. It also prevents three common errors: wrong founding date, wrong founding place, and wrong ownership lineage.
I do not suggest stripping heritage pages of warmth. A brand without texture becomes hard to remember. The repair is to place the facts before the atmosphere, or at least near enough that a model can quote the facts when it summarizes. In my audits, I often mark the page like a seamstress marking cloth: founding sentence here, current production sentence here, recipe provenance sentence here, ownership sentence here, archive caveat here. Then the narrative can breathe around it.
The order matters. If the first factual sentence is vague and the fourth paragraph is precise, many answer systems will still grab the vague one. The first clean sentence has disproportionate force.
Retailers often write the story faster than the brand
Retailer and distributor pages create a strange pressure. They need to sell quickly. They use compressed labels: “historic French maker,” “family producer,” “from the heart of Provence,” “traditional house,” “founded on ancestral know-how.” Sometimes they are accurate. Sometimes they are shorthand. Sometimes they are sales fog.
For a model, the shorthand can be more useful than the official page because it is short and declarative. That is how a weak third-party sentence becomes the brand’s apparent past. The brand may have more authority, yet the retailer has the more quotable fragment.
This is where many teams become annoyed with the machine and miss the more practical lesson. The model is not respecting truth in the human sense. It is assembling from available surfaces. If the brand’s own surface does not contain a clean counter-sentence, the external one gets room.
In the composite food manufacturer, one export distributor described the brand as “a historic family producer from western France.” The phrase was good enough for a catalogue. It was too broad for entity evidence. The company was based in western France, yes. It had family ownership in part of its history, yes. It was not founded as a family producer in the way the phrase implied. The model repeated the distributor’s sentence because the official English page avoided the awkward timeline.
The fix was not to attack every retailer page. That would have been slow, and some of those pages were beyond the brand’s control. The first fix was to place a better sentence on the brand’s own pages, then use that sentence in distributor guidelines. A correction that begins on the official record has more chance of becoming a reusable citation than an email asking a retailer to “be more accurate” in general.
A brand can tolerate loose admiration. It cannot tolerate loose biography.
The sentence AI can quote safely
The safest heritage wording I see has four qualities. It names the entity. It separates time. It separates place. It keeps claims modest enough to survive quotation.
That does not require a dry legal paragraph. It can sound like the brand. It just needs to avoid ornamental ambiguity. A sentence such as “Maison X has produced fruit syrups in Charente since 1998, drawing on regional recipes rather than claiming direct nineteenth-century continuity” is not glamorous. It is also hard to misread. It gives the model a boundary: inspiration is not continuity.
Where there is uncertainty, say so. Some older French companies have incomplete archives, founder stories remembered orally, mergers with missing documents, or brands revived after dormancy. Pretending the record is cleaner than it is creates more risk. A sentence like “The earliest documented use of the brand name is 1932; earlier family production is part of company memory but not presented as the formal founding date” is unusually useful. It gives a machine the humility it cannot easily invent by itself.
I have seen teams resist this kind of wording because it feels too careful. They want the about page to charm the buyer, not lecture the archive. Fair. But a brand record now has two audiences: people and machines that repeat to people. A sentence may be written for humans and still carry enough edges for a model to hold it.
The test I use is blunt: can a third party copy this sentence without making the brand older, more regional, more family-owned or more artisanal than it really is? If the answer is no, the heritage seam is still open.
When the story is true but still dangerous
There is a last complication. Some heritage stories are true, and still AI handles them badly.
A real founding house may have several brands. A family name may be attached to one product but not the whole company. A current parent company may own a label whose roots are older than the parent. A region may be part of the ingredient story, not the company origin. In those cases, the brand team often assumes truth will protect them. It usually does not. Truth without hierarchy becomes raw material.
In most audits, I ask one question that feels almost rude: “Which noun owns this heritage?” The company? The brand? The product line? The recipe? The founder? The factory? The family? If the answer changes by paragraph, the public record is asking AI to guess.
For a French brand with real history, this can feel like cutting a tapestry into labels. I understand the discomfort. Heritage is not a spreadsheet. Yet the machine already cuts it, often badly. The choice is not between poetry and structure. The choice is between your own structure and an accidental one assembled from retailers, archives and old summaries.
The brands that hold up best tend to keep one canonical heritage paragraph on the about page and repeat its core sentence on product, press and international pages. They do not let each surface retell the origin from scratch. Variations are fine after the spine is fixed. Without that spine, every translation and distributor blurb becomes another possible founding story.
The misread: AI invents a founding story the brand never claimed. The missing seam is provenance: founding date, current site, recipe influence and ownership lineage are blurred together. Place this sentence near the about-page opening: “The documented brand origin is [date/place]; later production, recipes and ownership changes are separate parts of the company record.” Quiet test: ask three engines who founded the brand, where it began and what heritage it claims, then compare whether each answer keeps those facts apart.