The strange part is that the fact is often real. The problem is its address. Synthesis can carry a claim across the page and leave it on the wrong doorstep.
A composite source set placed two French service businesses side by side. One was a B2B software company with scheduling and client follow-up features. The other was a broader invoicing provider with a cleaner directory description and a shorter name. In the final answer, the second business received the phrase “client follow-up,” while the first was reduced to “appointment scheduling.” The detail had not been invented. It had moved.
Atelier Syntheclair treats this as one of the more treacherous synthesis failures because it can look well supported at a glance. There is a citation. The source set contains both companies. The feature exists somewhere in the evidence. The answer even feels plausible. Only a slower source-to-claim reading reveals the mismatch: the fact belongs to one business, the sentence gives it to another.
The borrowed fact is different from a false fact
The lab separates a borrowed fact from a simple hallucination. In a hallucination, the answer produces a claim with no visible support in the inspected evidence. In a borrowed attribution, the claim has support nearby, but the support points to a different entity. The error is relational. The fact has lost its owner.
Wrong attribution is a synthesis error where a supported fact is placed beside the wrong business, because the answer resolves evidence into prose incorrectly. That definition is narrow on purpose. It keeps the material focused on a specific failure: the answer engine is not merely being vague, and it is not only dropping details. It is transferring them.
This matters for French business visibility because many firms are described in crowded language. Agencies, consultancies, software vendors and regulated-service advisers often share category phrases: accompagnement, gestion, conformité, suivi client, solution pour PME, service aux professionnels. A model reading several passages may treat those phrases as glue. The final answer wants to produce a tidy comparison, and tidy comparisons need short attribute bundles. That is where one company’s detail can slide into another company’s bundle.
In the lab’s anchor classification, this is the “borrowed” case. A business can be selected when it is named directly, softened when its detail becomes a category, borrowed when one of its attributes is assigned to another business, and erased when it remains in evidence but disappears from the final answer. Borrowing is particularly hard to catch because the answer may still mention the harmed business. It just gives part of that business to someone else.
How the transfer usually starts
A typical borrowed-fact scene begins with proximity. Two businesses appear in the same cited page, the same answer source, the same comparison list or the same retrieved summary. The page may be perfectly clear to a human reader. The model still has to compress it into an answer. During that compression, names and attributes become movable pieces.
Study Object A, the composite French B2B software company, shows this well. Suppose a comparison source describes three tools for small professional firms. The first offers scheduling and client follow-up. The second is stronger on invoicing. The third is a general CRM. A human reader can keep the rows separate. The final answer, however, may produce a sentence like: “For invoicing and client follow-up, the second provider is the most complete option.” If the cited page contains all those words, the citation looks plausible from a distance. The row boundary has disappeared.
The boundary can also fail when names are similar in category, not in spelling. A “solution de gestion pour cabinets” and a “logiciel de gestion pour indépendants” may be different businesses with different feature sets. Synthesis may treat them as members of one broad class and distribute attributes according to what makes the answer read smoothly. The company with the cleaner category label becomes the host for several features.
The lab has also seen a subtler version in bilingual prompt families. A French prompt may preserve the local phrasing of each business more carefully because the source language and answer language are aligned. An English prompt may translate both into simpler category language. Once translated, the two businesses look more alike. The feature has less friction to cross from one to the other.
When source layout invites confusion
The lab does not blame the source page automatically. Some pages are clear and still get compressed badly. Yet certain layouts make borrowing easier. Long comparison pages with repeated mini-profiles can blur when snippets are extracted. Directory entries that place several businesses under one heading can do the same. Agency-made pages sometimes describe a market first, then list companies, then return to market-level claims. The answer may attach a market-level claim to the nearest named business.
This is not a moral failure of the page. It is a structural risk. A page written for human scanning can rely on layout, spacing, tables and visual hierarchy. A final answer may not preserve those boundaries. It sees text, source summaries and citations; it must decide which claim belongs with which entity.
The regulated-service consultancy, Study Object B, brings a more serious version. Imagine a French source that discusses two consultancies. One provides preliminary eligibility review. The other handles administrative preparation but explicitly avoids advisory conclusions. A final answer says the second “helps determine eligibility and prepare the file.” Again, the terms are not fabricated. Eligibility was present. File preparation was present. The combined claim belongs to neither business exactly as written.
The lab is careful here because regulated categories attract general caveats. Some wrong attribution may be masked by caution: “may help with,” “can support,” “appears to provide.” A cautious verb does not repair the ownership problem. If the answer places an eligibility function beside the wrong consultancy, the hedge only softens the visible risk. It does not restore the source boundary.
One rough detail often exposes the transfer. The answer may use a phrase that sounds slightly foreign to the named business’s own page. For example, a company that consistently says “client reminders” is described as offering “customer relationship management,” a phrase that appears in the neighbouring provider’s source. That wording fingerprint is useful. It gives the reader a thread to pull.
The role of answer shape
Final answers are not neutral containers. They have shapes: recommendation, comparison, explanation, shortlist, local alternative, cautionary overview. Each shape pulls facts into a different arrangement.
In a recommendation answer, facts tend to cluster around the chosen winner. A business selected as “best for X” may inherit supporting details that help justify the recommendation. If the source set contains nearby claims about X, the model may use them to make the chosen business sound more complete. This is one reason borrowed facts often appear near strong recommendation language.
In a comparison answer, the risk changes. The answer wants each business to occupy a distinct slot. One is “best for invoicing,” another “best for scheduling,” another “best for compliance.” If the source evidence is messier, synthesis may force separation that the sources do not support. A feature shared by two businesses may be assigned to one, while a feature belonging to one may be used to differentiate another.
Explanatory answers can borrow in a quieter way. The user asks, “What kinds of French firms help with regulated service setup?” The final answer names a consultancy and then describes the category. A category-level statement slips next to the business name. The reader may interpret it as a company-specific claim, although the source only supported it as a general market description.
Atelier Syntheclair sees this as an editorial problem inside synthesis. The answer engine is trying to be useful. It trims, groups and assigns. But usefulness at sentence level can damage attribution at business level. A neat paragraph may be less faithful than a clumsy one.
How the lab reads a suspected borrowed claim
The lab starts with the final sentence, not with the whole page. Which business is named? What exact claim sits beside it? Which citation or source reference is attached? Then the team reads the cited passage closely. If the passage supports the claim and the business together, the attribution holds. If it supports the claim but points to another business, the case becomes a borrowing candidate. If it supports the business but not the claim, the case may belong to citation-support analysis instead.
The distinction is small but useful. Work-item 3 asks whether citations support claims. This material asks a narrower question: whether the supported fact has been attached to the right French business. The two problems overlap in practice, yet the lab keeps them apart so the diagnosis does not become cloudy.
A borrowed claim is easier to mark when the source passage contains clear row boundaries, headings or repeated names. It is harder when the source itself speaks loosely. If a directory entry says “these providers support compliance and administrative setup” and then lists several firms, the lab may avoid calling the final answer wrong. The evidence may be too coarse. The better label might be “thin support” or “plausible synthesis tendency,” depending on what is visible.
The lab’s strongest borrowing cases have three parts: the answer places a specific claim beside a named business; the cited evidence places that same claim closer to another named business; and another run or language variant shows the claim staying with its original owner. That last comparison is useful, though not required. It helps show that the transfer is not merely the lab reading too much into one sentence.
What this means for French business pages
A business cannot control every answer composition. It can, however, reduce the number of loose facts available for borrowing. Study Object A suggests that feature statements should stay close to the company name and category. If “client follow-up” is central, it should appear in a sentence that names the business, not only in a bullet far below a comparison block or in a testimonial with unclear subject.
Study Object B suggests the same for limits. Eligibility review, advisory boundaries and compliance steps should be tied to the entity that performs them. A page that speaks generally about “our partners,” “the service,” and “support” may be readable to a human visitor who sees the page context. In extracted evidence, those pronouns can become fog. The final answer may attach the sentence to whichever name is closest or easiest.
This does not mean every sentence must repeat the brand name in an unnatural way. The lab is not recommending a page written like a shipping label. It means that high-value claims need enough local anchoring to survive extraction and recomposition. The name, the claim and the boundary should be neighbours.
A practical reading habit follows. When a business appears in an answer with a feature it likes, the team should still ask whether the cited passage actually assigns that feature to it. Pleasant mistakes are still mistakes. They can create future disappointment, especially when the answer makes the company sound broader, more qualified or more regulated than it is.
Limits of the borrowed-fact diagnosis
Atelier Syntheclair cannot always prove borrowing. Some interfaces show only a source card, not the exact passage. Some answer engines cite a page that contains several relevant sections, and the visible snippet may not include the part used by the model. In those cases, the lab may identify a suspicious attribution but should not state the transfer as fact.
The method also works with small prompt families and composite scenarios. It can show how a feature moved in related observations. It does not measure how often this happens across all French businesses or all answer engines. The lab’s labels remain descriptive: observed in this run, recurring across related runs, plausible synthesis tendency. Anything stronger would pretend to a scale the work does not have.
There is another caution. A source may be ambiguous enough that the final answer’s attribution is defensible. If a page itself mixes company-level and category-level claims, the model may not be the first place where ownership failed. The lab then treats the source as part of the pattern, not as clean ground truth. Synthesis can magnify ambiguity that was already present.
The final difficulty is time. Answer engines change, citations shift, and repeated prompts can produce different wording. A borrowing case documented in one run may not repeat later. That does not make the observation useless. It means the reader should treat it as a visible example of a mechanism, not as a permanent verdict on the business named in the run.