A business rarely disappears all at once. More often, its name remains while the edges are sanded down: who it serves, what it actually offers, and where the service stops.
In one composite observation used by Atelier Syntheclair, a French B2B software company was easy to find in the source set. Its pages described scheduling, invoicing, client follow-up and a narrow audience of small professional firms. The final answer kept the company name. It even placed the business in roughly the right category. Then, in the next sentence, the product became “a management platform for SMEs,” with no mention of appointments, invoicing, client reminders or the professional-service niche that made the company different from a dozen broader tools.
That was not a simple omission. The answer had enough material to speak more precisely, yet it chose a safer, thinner description. A nearby run did the opposite in a small way: it preserved “invoicing” and “client follow-up,” but dropped scheduling, which was the entry point of the user’s original question. The lab treats this kind of result as a useful fracture line. The business survives, but not all of its attributes come through the same door.
What counts as an attribute that survives
Atelier Syntheclair uses “attribute” in a plain sense. It means a business detail that changes how a reader understands the company: service area, customer type, product boundary, eligibility condition, regulatory limit, feature, geography, pricing frame, sector fit or evidence claim. The lab does not count every adjective. “Modern,” “simple” or “trusted” usually says more about the source’s selling tone than about the business itself. “For independent accountants in France,” “requires a compliance review before onboarding,” or “handles appointment reminders but not payroll” are different. Those details have consequences.
Attribute survival is the movement of a specific business detail from retrieved evidence into final answer wording, because synthesis decides which facts remain usable. This working definition matters because the answer can look correct while carrying only the easiest pieces of a company. A name may be selected. A category may be kept. The edges may still be lost.
The lab’s early materials around the composite software company, Study Object A, show the same rough pressure in several prompt families. When the prompt asks for “tools for small firms,” broad attributes survive well. “French software,” “SMB,” “management,” and “client follow-up” tend to fit the answer’s expected shape. When the prompt asks for a more practical distinction, such as whether a tool handles appointment flow and invoicing together, the answer is more fragile. Sometimes it retains the combined feature. Sometimes it splits the company into a generic scheduling tool and gives the invoicing strength to a more familiar neighbour in the evidence.
The regulated-service consultancy, Study Object B, behaves differently. Its pages contain eligibility, compliance steps and advisory limits. These details are heavier. They slow the sentence down. In final synthesis, the answer often prefers a cautious shell: “a French consultancy for compliance support,” “may assist with regulated procedures,” or “offers guidance for eligible businesses.” The phrase is not false, exactly. It is like a label stuck over the small print. The reader gets a category, not the operational boundary that would help them act.
The attributes that travel easily
Some attributes are built for travel. A company name with a stable spelling travels well, unless the answer substitutes another entity. A broad category travels well when several sources repeat it. Geography travels well when it is simple: “France,” “Paris,” “French-speaking,” “for French SMEs.” A basic audience label also tends to survive if it is short and ordinary. The answer can carry these facts without changing its rhythm.
The lab sees this most clearly when a source describes the same attribute in several nearby ways. If a software page says “for small professional firms,” a directory says “SMB software,” and a comparison page says “management tool for independent service businesses,” synthesis has a neat cluster. The answer may compress it into “for small businesses,” but the attribute remains recognisable. It has enough repeats to feel safe.
Features survive less reliably, yet some do make it through. In Study Object A, invoicing survives more often than client follow-up when the prompt contains accounting language. Scheduling survives when the prompt contains appointment language. A feature that matches the query is more likely to be carried forward than a feature that merely sits in the source. This sounds obvious, but the lab is cautious with the simple version of the claim. The feature does not survive only because it is present. It survives because it helps the answer complete its chosen story.
That chosen story can be narrower than the user intended. If the final answer decides it is recommending “business management tools,” the software company’s scheduling feature may appear decorative. If it decides the answer is about “booking tools,” invoicing becomes secondary. The same source page has not changed. The synthesis frame has.
Softening is the most common loss
The lab’s anchor classification gives the material a small vocabulary for this passage through synthesis: selected, softened, borrowed or erased. A business is selected when the final answer names it directly. It is softened when a specific attribute becomes a broad category. It is borrowed when a detail from one business is attached to another. It is erased when the attribute or business remains in evidence but disappears from the answer.
For this work-item, softening is the central case. The business is still visible, but the practical detail has been rubbed into something easier to say. “Offers appointment scheduling and invoice follow-up for small clinics and consultants” becomes “provides business management software.” “Explains eligibility steps for regulated advisory work” becomes “helps with compliance.” These are not dramatic hallucinations. They are quieter, and for that reason easier to miss.
A softened attribute can still be useful. A founder scanning the answer may only need the broad category. A marketer, however, will care that the phrase no longer explains the company’s actual advantage. The answer has not lied in the loud way. It has reduced the business until it fits the mould of the paragraph.
Borrowing sits nearby but should not be confused with softening. In one typical paired-source scene, the software company’s client-reminder feature appears near another provider’s invoicing feature. The final answer says the second provider offers “invoicing and client follow-up,” while the first is described only as scheduling software. The feature has not vanished from the answer. It has moved. The lab treats that as a different pattern, because the damage is not thinning alone; it is misplacement.
Erasure is more severe. This happens when a business remains in a cited page or visible source passage, but the final wording leaves out the attribute entirely. In regulated-service prompts, eligibility boundaries often meet this fate. The answer may recommend caution or professional review, yet omit the concrete condition that the source page gave. For a reader, the missing boundary can matter more than the visible category.
Why limits disappear faster than categories
Limits are awkward material for synthesis. They interrupt fluency. They often require conditions, dates, exceptions or careful verbs. A category can be stated in one breath. A limit asks the answer to slow down.
Study Object B makes this visible. A composite regulated-service consultancy may explain that it advises only on preliminary eligibility, that formal filing must be handled by a qualified professional, and that certain cases fall outside its scope. In an answer about “regulated business help in France,” these limits are often compressed into “check professional requirements” or “consult a qualified adviser.” That caution may be sensible, but it no longer represents the business’s own boundary. The answer has replaced a source-specific limit with a general safety phrase.
The lab suspects that answer engines prefer limits when the user directly asks for risk, eligibility or compliance. The wording then has a place to go. When the user asks for a recommendation or comparison, limits compete with selling points and category labels. They can look like friction. The final answer may keep the attractive part of the business and drop the conditions that define it.
This is where source writing and synthesis meet. A detail buried in a long paragraph, written once, framed as an aside, is easier to lose. A detail repeated near the service description, the FAQ and the page title has more chances to be carried into the answer. The lab does not turn that into a promise. It simply treats repetition and proximity as visible reasons why one attribute survives and another does not.
A small roughness matters here. In one composite run, the answer preserved an eligibility limit but assigned it the wrong tone. The source said the consultancy “does not replace formal legal review.” The final answer said it “may be unsuitable for complex legal cases.” That is not identical. The first phrase defines scope; the second sounds like a weakness. The attribute survived, yet changed its posture.
Reading loss without overclaiming it
A marketer may want a clean diagnostic: which attributes should be written where, how many times, and in what format. Atelier Syntheclair avoids giving that kind of certainty. Their method is closer to close reading than measurement. A small prompt family can show how an attribute behaved in related answers. It cannot prove how every answer engine will treat that attribute across the French market.
The lab’s labels help keep that modesty intact. “Observed in this run” marks a visible case. “Recurring across related runs” marks a pattern seen more than once within a controlled prompt family. “Plausible synthesis tendency” marks an interpretation that fits the evidence but still needs caution. These labels are deliberately unspectacular. They keep the material from sounding more measured than it is.
The method also depends on what the interface reveals. Some answer engines show citations. Some show source cards without exact passages. Some rewrite evidence summaries, and some hide the chain almost entirely. If the lab cannot see the retrieved evidence, it can still study final wording, but it cannot confidently separate retrieval loss from synthesis loss. That distinction belongs more fully to the later material on diagnosing where visibility is lost.
Even when citations are visible, they may be partial. A source card may support the broad category but not the feature beside it. A cited passage may mention the company but not the eligibility boundary. That makes attribute survival a source-to-answer question, not just a content-quality question. The lab reads the passage, the claim and the final sentence together.
What businesses can learn from the surviving details
The practical lesson is not to stuff every page with every feature. That would make the source less readable and may produce new confusion. The clearer lesson is that synthesis rewards attributes that are stable, repeated, close to the business name and easy to use in the answer’s task. If a feature matters, it should not appear only as a decorative phrase deep in a page. If a limit matters, it should not be trapped in a legal-sounding afterthought that no recommendation sentence can easily carry.
Study Object A suggests one path: keep the company’s category, audience and feature boundaries close together. “Scheduling, invoicing and client follow-up for small professional firms” is more synthesis-ready than three separate claims scattered across a homepage. Study Object B suggests another: keep eligibility and advisory limits near the service claim, so the model does not replace them with generic caution.
None of this guarantees final-answer survival. Answer engines change. Prompt wording changes. A competitor with cleaner source language may still be easier to compress. But a business can make its attributes less slippery. It can give the synthesis layer fewer excuses to keep the name and lose the substance.