A missing business is not one problem. It can be absent from the evidence shelf, or it can be on the shelf and still left out of the paragraph the user reads.
In a composite local-services query, the answer did not name the French business the marketer expected. The first reaction was familiar: “the system cannot find us.” But the visible citations told a more awkward story. The business appeared in a directory page, and a service description close to its own wording sat inside the evidence. The final answer still named another provider.
A different composite run looked similar from the outside and completely different under the glass. The expected business was absent from the final answer, absent from the citations and absent from repeated source traces. Related prompts found competitors, category pages and local media, but not the business itself. This was not a final-answer survival problem. It was a discovery problem.
Two losses that look the same to the reader
For a user reading only the final answer, both cases are identical. The business is missing. For Atelier Syntheclair, they belong to different parts of the visibility chain. The lab studies the synthesis layer, but that layer can only be diagnosed clearly when retrieval is separated from final composition.
Retrieval loss is the absence of a business from visible or reconstructable evidence, because the answer engine does not appear to gather usable source material about it for the prompt. Synthesis loss is the disappearance or weakening of a business inside the final answer, because evidence is available but the final prose selects, softens, borrows from or erases differently.
That definition keeps the work from blaming the wrong machine part. If a business never enters the evidence set, studying answer ordering will not explain much. If the business is present in citations and still not named, adding more directory listings may not address the final editorial failure. The diagnosis changes the question.
The lab’s position is not that retrieval and synthesis are perfectly separable in every interface. They are often tangled. Some systems expose citations after the answer is written. Some hide sources. Some show summaries rather than passages. Still, the distinction is useful because it forces the reader to ask what can actually be observed.
A French SMB or agency may want a simple verdict: found or not found. The lab’s work suggests a rougher but better question: where did the business stop moving? Did it fail to enter the source set, or did it enter and then fail to survive the final answer?
The retrieval-side failure
Retrieval loss tends to leave a cold trace. The business is absent from citations, absent from repeated answer sources, and absent when the prompt is varied around the same intent. The answer may still be plausible. It may name other businesses in the category. It may cite directories, articles or competitor pages. But the expected business is not visibly there.
Using Object A, the composite French B2B software company, a retrieval-side problem might appear when prompts about scheduling and invoicing tools for small professional firms retrieve broad software lists but never the company’s own pages or relevant mentions. The final answer cannot select what it has not gathered. In that case, the synthesis layer may be doing ordinary composition from a source set that lacks the business.
The temptation is to infer too much from one failed prompt. Atelier Syntheclair avoids that. A single prompt may be poorly phrased, too broad, too local, too ambiguous or too close to a competitor’s category wording. The team looks for prompt families: brand-versus-category questions, narrower sector questions, French and English variants, and comparison prompts. If the business remains absent across related routes, retrieval loss becomes the stronger diagnosis.
Object B, the regulated-service consultancy, can disappear at retrieval for different reasons. Its pages may use formal French wording that does not match common query phrasing. Its service may be described through process terms rather than category terms. Or the available public evidence may be too thin for the answer engine to treat it as a usable source. The lab can observe absence; it cannot always know which underlying retrieval mechanism caused it.
This is where the method stays modest. The lab does not claim access to the engine’s full index or ranking machinery. It works with visible answers, citations, source passages, language variants and repeated-output differences. Retrieval loss, in this framework, is a diagnosis from available traces, not a claim about the whole internal system.
The synthesis-side failure
Synthesis loss feels more irritating because the business was close enough to touch. It appears in evidence, or its source passage appears, or the answer cites a page where the business is clearly present. Then the final prose chooses another name, blurs the business into a category, transfers a feature elsewhere, or drops it.
Object A often shows a clean synthesis-side pattern. The expected software provider appears in a cited comparison page beside two competitors. Its own service page gives a specific fit for small professional firms. The final answer recommends a larger competitor with a cleaner category sentence. The retrieved evidence contained the smaller provider, but the answer treated it as secondary material.
Object B shows another version. The consultancy appears in a source passage about a regulated process. The final answer describes the category and names a different firm, while borrowing the original consultancy’s careful caveat as a general warning. The business did not fail at discovery. It failed when the answer engine turned several pieces of evidence into one readable recommendation.
This is where the canon’s anchor classification is most useful: selected, softened, borrowed or erased. A business is selected when it is named directly. It is softened when it becomes only “a French provider” or “a local consultancy.” It is borrowed from when its feature or limit attaches to another business. It is erased when it remains in evidence but disappears from the final answer.
Synthesis loss can be partial. A business may be named but placed last, described weakly, or stripped of the attribute that made it relevant. The lab does not reserve visibility loss only for total disappearance. Final-answer prominence matters, though the lab treats it qualitatively rather than as a score. A name buried in a flat list is not the same as a name explained as the best fit.
The hard part is that synthesis loss often looks reasonable. The answer is smooth. The named competitor may genuinely be relevant. The citation may appear beside the paragraph. Nothing screams failure. The mismatch appears only when the reader compares the final claim with the source passage and asks why the available business was not carried through.
A practical diagnostic path
The lab’s diagnostic sequence starts with the final answer, not with a theory. What business was expected? What business was named? What category did the answer use? Which claims appeared beside each name? Then the team looks for visible evidence: citations, source summaries, quoted passages, repeated references and language variants.
If the expected business is absent from the visible evidence across related prompts, retrieval loss becomes the working diagnosis. If the expected business is present in evidence but absent or weakened in the final answer, synthesis loss becomes the working diagnosis. If the traces are mixed, the lab keeps both possibilities open.
The bilingual step often matters for French businesses. A company may be retrieved in a French prompt but not in an English one. Or it may be retrieved in both, yet selected only in French because the English answer favours cleaner category wording. That distinction prevents a common mistake: treating English absence as total invisibility when French evidence is alive, or treating French selection as proof that English synthesis will follow.
The lab also looks at the granularity of disappearance. Did the name vanish? Did the category change? Did the feature disappear? Did the evidence remain but the answer order demote the business? These are different failures. Object A may keep the name but lose the product boundary. Object B may keep the category but lose the advisory limit. Both matter, but they point to different questions.
A small imperfection often helps the diagnosis. In one composite case, the answer cited a directory page that included the expected business, but the page’s snippet emphasised a competitor. The business was technically present in evidence, yet the cited passage made it weak. The lab marked this as a mixed case: retrieval had found the broader source, while synthesis followed the passage that was easier to compress. Not every case lands cleanly.
What action follows from each diagnosis
This material is not a remediation checklist, but the diagnosis has consequences. When the failure is retrieval-side, the question is whether the business has usable public evidence for the prompt family. The issue may involve missing category language, thin source presence, weak French-English alignment, or sources that mention the business without explaining it. The lab would not call that a synthesis pattern yet.
When the failure is synthesis-side, adding more of the same evidence may not solve the visible problem. The business is already present. The issue is that the final answer prefers another name, softens the business, borrows its features, or erases it during composition. The relevant reading shifts toward claim support, answer ordering, category clarity and the phrasing that makes one entity easier to carry into prose.
For Object A, a retrieval-side diagnosis might mean the provider is missing from category evidence around “scheduling and invoicing for small professional firms.” A synthesis-side diagnosis might mean the provider is present but loses to a broader “practice management” label. Those are not the same problem. One concerns entrance into the source set; the other concerns survival inside the answer.
For Object B, retrieval loss may reflect sparse public evidence or a mismatch between formal service language and user prompts. Synthesis loss may reflect caveats being stripped from the consultancy and applied as general warnings. In regulated categories, that difference is especially important because a visibility failure can become a trust failure. The answer may make the wrong business sound safer or more complete than the evidence permits.
Atelier Syntheclair is careful not to promise control over either layer. Answer engines change, interfaces vary, and citations can be partial. The lab’s contribution is a way to read the failure before naming it. Misdiagnosis is expensive in attention, even when no money changes hands. Teams can spend months polishing pages for retrieval when their real problem is final-answer substitution. They can also obsess over synthesis when the business is simply not entering the source set.
Limits of the diagnosis
The clean distinction between retrieval and synthesis is a method, not a window into the machine. The lab cannot see every document considered by an answer engine. It cannot know all ranking signals, internal summaries or hidden source transformations. It can only work from observable material: final answers, citations, passages, prompt variants and repeated-output differences.
That means some cases remain ambiguous. A business may appear in a cited page but not in the cited passage. It may be retrieved as part of a directory but not treated as relevant to the specific answer. It may appear in one run and vanish in another because the model sampled a different route. The lab marks these as observed in this run, recurring across related runs, or plausible synthesis tendency, rather than forcing a firm verdict.
The composite objects also keep the conclusion bounded. Object A and Object B are built from typical scenarios that help expose mechanisms. They are not a measured sample of French SMBs, software providers or regulated consultancies. The material describes a diagnostic distinction and shows how it can be applied; it does not report market frequency.
There is one more limit: the reader’s expectation may be wrong. A business owner may expect to appear for a broad query where the public evidence gives stronger support to other firms. The lab does not treat every absence as a failure. It asks whether the absence can be explained by the evidence path and the final synthesis. Sometimes the answer is simply choosing a more supported result.
Still, the distinction holds. A missing French business should not be diagnosed by its absence alone. The evidence shelf and the final paragraph must be inspected separately where possible. Only then can the lab say whether the business was never gathered, gathered and softened, gathered and borrowed from, or gathered and erased.