A citation can sit beside a recommendation like a label on a jar. The lab opens the jar and checks whether the contents match, or whether synthesis has poured in something else.
A composite French B2B software company is described in a cited passage as offering scheduling, invoicing and client follow-up tools for small professional firms. The final answer recommends it as “the best choice for agencies needing a complete client operations platform.” The source supports parts of that sentence. It does not support all of it. “Best choice” is new. “Agencies” is narrower than the source. “Complete platform” is a larger claim than the page makes.
This is not the same problem as a missing citation. The citation is there. The source is relevant. The business may even be a reasonable candidate. The drift happens inside the final stretch, when the answer engine turns source material into a recommendation. Atelier Syntheclair studies that stretch because many readers treat a citation as a small stamp of proof. In recommendation answers, the stamp can be real while the proof is thinner than it looks.
What source drift means in a recommendation
A recommendation asks the answer to do more than summarize. It must compare, select and justify. That extra pressure makes source drift more likely. The system may retrieve a page showing that a business has certain features, then write a sentence implying that those features make it the strongest option. The source supports existence and capability, but the final answer adds judgment.
Recommendation source drift is the gap between what a cited passage supports and what the final recommendation claims because synthesis strengthens, narrows or redirects the conclusion. The definition is deliberately claim-level. A source can support the business name and still fail to support the reason the answer recommends it.
The lab’s concern is not that answer engines make judgments. Users ask for judgments. A recommendation that merely lists sources would be evasive. The issue is whether the judgment is traceable to evidence or whether it appears because the answer needs a neat ending. The neater the recommendation, the easier it is for drift to hide.
Study object A is useful here: a typical French B2B software company serving small professional firms with scheduling, invoicing and client follow-up tools. In one composite pattern, the source page describes features in a measured way. The final answer recommends the business for “teams seeking an all-in-one operational suite.” That phrase may sound natural, but it shifts the category. A small toolset becomes a suite. A target audience of small professional firms becomes teams in general. The recommendation has not broken free from the source entirely. It has stretched the fabric.
Study object B shows a stricter version of the same problem. A regulated-service consultancy source may say it helps clients prepare documents and understand eligibility. The final answer recommends it “for handling compliance procedures.” That wording is risky because “handling” may imply responsibility for the process. The source supports assistance, not control. In regulated contexts, a small verb can carry a large burden.
Where the drift enters the sentence
Atelier Syntheclair often finds drift in four places: the recommendation verb, the category label, the audience, and the reason clause. The answer may move from “offers” to “is best for,” from “scheduling tool” to “operations platform,” from “small firms” to “growing agencies,” or from “includes invoicing” to “ideal for managing finances.” Each step is plausible. Together they can produce a recommendation the source would not have written.
The lab reads these moves through the anchor classification from its canon. A business is selected when the answer names it directly as the recommendation. It is softened when the reason becomes generic, such as “good support” or “useful services,” without carrying the source’s specific feature. It is borrowed when the recommendation reason appears to come from another business or nearby source. It is erased when a business present in evidence disappears while another receives the final recommendation.
Source drift often begins after selection. The business gets the visible place, but the reason beside it changes. That is why recommendation drift can be flattering and still inaccurate. A business may be recommended for a capability it only partly has, or for a client type it does not primarily serve. To a marketer, this may look like visibility. To the lab, it is unstable visibility because the match between evidence and claim is weak.
A composite Study object A example makes the mechanics visible. The cited passage says the tool includes appointment booking, invoice templates and reminders. A second source, discussing a different provider, emphasizes team workflows and reporting dashboards. The final answer recommends Study object A for “teams that need scheduling, invoicing and reporting in one place.” The reporting feature has crossed the seam. If the citation beside the sentence points to the first source, the reader sees proof where the proof is not actually present. This is borrowing inside a recommendation.
The more ordinary version is not borrowing but stronger reasoning. The cited source says the business serves “cabinets indépendants et petites structures.” The answer says it is “especially suitable for agencies that want a scalable solution.” “Especially suitable” is the model’s judgment. It may be defendable if other evidence supports it, but if the cited passage does not, the answer has drifted from source support into synthesis confidence.
Why recommendations invite stronger wording
Recommendation prompts create a pressure toward closure. A user asks which business to choose, which provider is better, or which option fits a need. The answer feels incomplete unless it gives a reason. The system then writes a reason using the materials at hand. If the source is modest, the reason may become slightly less modest in order to satisfy the prompt.
The lab does not treat this as malice or hallucination in the dramatic sense. It is more like a tailor tugging a sleeve to make a jacket look finished. The cloth was real. The tug changed the fit. Recommendation synthesis often takes true fragments and arranges them into a more decisive shape than the source supports.
This is where the difference between “claim support” and “recommendation support” becomes useful. A cited page may support that a firm offers scheduling. It may support that the firm serves small professional clients. It may support that it has French-language documentation. But a recommendation claim says something else: given the user’s need, this firm is a good or preferred answer. That claim depends on comparison, fit and sometimes exclusion. The source passage rarely carries all of that by itself.
In the lab’s readings, drift becomes more visible when the answer compares several French businesses. One option receives careful detail, another receives a generic phrase, and a third receives the final recommendation. The cited passages may be similar in strength, but the final answer gives one source a more confident narrative. This overlaps with work on ordering and prominence, but the present material stays with the source-to-recommendation seam: does the reason for selection exist in the evidence the answer shows?
A useful answer can make its own judgment, but it should mark the basis. “Based on the cited feature list, this appears suitable for small firms needing scheduling and invoicing” is a different sentence from “this is the best all-in-one solution for agencies.” The first keeps the judgment tied to visible evidence. The second asks the reader to trust a conclusion the citation may not carry.
The role of clean source wording
Some sources are easier for answer engines to recommend from. They contain stable category sentences, clear feature boundaries and audience language that can be lifted without much strain. Messier sources require more interpretation. That does not make them worse businesses. It makes them harder to carry through synthesis.
The lab has seen composite cases where a source with cleaner wording wins the recommendation even when another source appears equally relevant. A business that says “software for client scheduling and invoicing for independent consultants” gives the answer a ready-made reason. A business that describes the same capability across three paragraphs, with examples and exceptions, may be softened or omitted. In recommendation answers, tidy phrasing can behave like a handle.
This creates a particular risk for French SMBs. Their pages may be written for human reassurance, local trust or sector nuance. They may include careful explanations but lack one compact sentence that ties category, audience and supported features together. During synthesis, the answer may then borrow a cleaner category from a directory or competitor. The final recommendation looks evidence-based, but the evidence has been rearranged.
The lab remains cautious here. It cannot say that cleaner wording always wins. Some answer engines do preserve detailed pages well, and some over-compressed pages lose important boundaries. The observed tendency is narrower: when a prompt asks for a recommendation, source wording that already contains a usable fit statement is easier for synthesis to carry without drift.
That insight has a practical edge, but the material is not a content-advice checklist. The research point is that source drift is not only about false information. It can come from missing connective tissue. If the source states features but not audience fit, the answer may invent the fit. If the source states audience but not boundaries, the answer may overextend the service. Good synthesis should resist that temptation. It does not always do so.
How the lab judges support without pretending certainty
Atelier Syntheclair compares the recommendation sentence against the cited passage phrase by phrase. The team asks what the source directly supports, what it indirectly suggests, and what the answer adds. Direct support is usually visible in matching or closely paraphrased wording. Indirect support requires more caution. Added judgment is not automatically wrong, but it must be named as synthesis rather than source fact.
The team uses modest labels: observed in this run, recurring across related runs, and plausible synthesis tendency. If one answer recommends a composite software provider as “best for agencies” without source support, that is observed in this run. If related prompt variants repeatedly strengthen the same source into the same recommendation, the pattern becomes recurring across related runs. If the team sees similar movement across several composite cases, it may describe a plausible synthesis tendency.
Those labels are intentionally plain. They prevent a case note from pretending to be a market measurement. They also protect the reader from the opposite mistake, which is to dismiss every mismatch as random. Synthesis can be unstable and still patterned enough to study.
The lab also distinguishes drift from legitimate inference. If a source says a business offers scheduling, invoicing and client follow-up for small professional firms, an answer may reasonably infer that it fits a small practice seeking administrative coordination. But “best,” “complete,” “leading,” “guaranteed,” “for all agencies,” or “handles compliance” are stronger claims. They need stronger support. The line is not mathematical. It is editorial, which is why close reading is required.
Limits of this source-drift material
This material cannot prove that a recommendation is wrong in the real world. A business may indeed be the best fit for a user even if the cited passage does not prove it. Atelier Syntheclair studies the visible relationship between the final answer and the evidence shown, not the full commercial truth of the market.
The method is also constrained by partial citations. Some systems may base a recommendation on sources not displayed to the user. Others may cite a page as a general reference while the reasoning came from hidden retrieval, model memory or another source card. The lab can only inspect what can be placed before the reader: answer text, citations, source passages, prompt variants and repeated-output differences.
Composite study objects keep the analysis focused on mechanism rather than accusation. Study object A and Study object B are not named clients. They are typical scenarios assembled from observed page structures and answer behaviours. The lab uses them to show how recommendation wording can move beyond source support without making public claims about identifiable firms.
The conclusion stays narrow. A citation beside a recommendation is not the same as evidence for the recommendation. It may prove the business exists, or that it has one feature, or that a source used a certain category. The reader still has to ask whether the final reason for recommending the business is selected from the source, softened into a generic fit, borrowed from a neighbouring context, or quietly invented by the pressure to answer.