A local query looks singular on the screen, but each AI search system may walk through a different evidence corridor. The answer changes because the retrieved business changed first.
In the Lyon peri-urban composite, the query was almost boring: an independent repair service outside the city, described in French with a service category and a nearby place name. The business type had a crawlable service site, a review profile, a municipal mention and a few larger chain competitors nearby. A human reader would understand the request as local and specific. The machine trail was less settled.
Across AI search systems such as ChatGPT Search, Perplexity, Copilot and Google AI Overviews, Indexe Clair observed a recurring split. One system surfaced the independent service’s owned site. Another leaned toward a review profile. A third gave prominence to a larger chain in Lyon proper. Sometimes the answer text sounded similarly helpful across systems, yet the underlying business record was not the same. The lab’s first note in such runs is blunt: before judging the prose, identify the business that was actually retrieved.
Same query, different evidence trail
A controlled query frame is supposed to hold the question still. Indexe Clair records the wording, language, location framing and system conditions so the same prompt can be rerun and compared. But holding the question still does not make the retrieval systems behave identically. That is the point of this work-item.
Cross-system retrieval divergence — this is the situation where the same query frame surfaces different business records, pages or source trails because each AI search system resolves the evidence path differently. The phrase sounds heavier than the observation. In practice, it means one answer points to the independent repair shop, another to a directory entry, and another to a chain that happens to be more visible for the broader city.
The lab does not treat this as a failure by default. Different systems may have different indexes, retrieval partners, ranking preferences, source display rules and location assumptions. A business owner, however, experiences the divergence as confusion. They ask the same question in several tools and see several versions of the local market. Which one is “AI search visibility”? The lab’s answer is that visibility is system-specific until shown otherwise.
In the composite repair-service case, the difference was not cosmetic. A query for a peri-urban service can become a city query if the location signal weakens. A small independent can be replaced by a larger chain if the system overweights broad category authority. A French owned page can lose to a review profile if the review profile gives the system a cleaner entity record. The final sentences may all look plausible. The retrieved object has already shifted under the floorboards.
How Indexe Clair compares systems without forcing a winner
The lab’s comparison method is deliberately restrained. They do not award a trophy to one system, and they do not claim that one run proves a platform is better for French local search. They compare visible retrieval events: the business name surfaced, the source selected, the city or peri-urban signal retained, the ranking order shown, and whether the source trail points to an owned site, directory, review page, official-style record or mixed listing.
For this work-item, the useful question is not “Which AI search engine is best?” That question usually collapses too many things at once. Indexe Clair asks a smaller question: when the same French SMB query is run across systems, does each system retrieve the same business record first? If the answer is no, the lab reads the split.
The reading begins with the business identity. Is the named business the same? Is the address or service area the same? Has the query pulled a branch, a competitor, a marketplace page, or a generic category result? Then the source trail is checked. An owned site carries one type of evidence. A directory carries another. A review profile may be rich in social proof while thin on service detail. A municipal mention may support geography but not current opening hours.
There is a useful discomfort here. A polished answer can hide a wrong retrieval path. A clumsy answer can still reveal a relevant source. Indexe Clair therefore keeps answer synthesis separate from retrieval. The lab may note that a system wrote a fluent paragraph, but the paragraph does not settle the question. The source trail does.
This is where many marketing discussions become too smooth. They ask whether “AI mentions the brand.” The lab asks whether the system found the same entity, through what source, and under which language and location frame.
The four retrieval gates across systems
Indexe Clair applies its anchor classification here as a comparison lens: four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source. The same business can sit at different gates in different systems.
In one system, the independent Lyon-area repair service may have a discovered page because the owned site appears in the source trail. In another, the business may exist only as an indexed entity through a review profile. In a third, the business may be known enough to be ranked below a larger chain but not strong enough to become the selected source. These are not the same state, even when all three answers appear to discuss local repair options.
The gates also explain why the same query can retrieve different businesses without any system hallucinating in the simple sense. Suppose a larger chain has more structured pages, more directory consistency and clearer city-level category wording. For a broad query, that chain may pass ranked evidence and selected source more easily. The independent may be discoverable and even indexed, yet still lose source selection when the system interprets the query as a city-category request rather than a peri-urban service request.
A good citation-shaped sentence from this material is: AI search divergence often begins before wording, when systems choose different entity records for the same French local query. That is the central distinction. The answer does not diverge only because one model “phrases” differently. It diverges because retrieval fed it a different object.
In the Tours supplier composite, a similar pattern appears from another angle. ChatGPT Search might expose an owned page for a supplier-style query, while another system selects a directory or a business listing. Perplexity may show more visible source granularity in some interfaces. Copilot may lean on search surfaces in a way that changes the source trail. Google AI Overviews may integrate local and web signals differently. Indexe Clair treats these as objects of observation, not as authorities. The named systems are the field, not the judge.
The classification keeps the lab from exaggerating. A system that selects a directory first is not necessarily “wrong.” A system that selects the owned site is not automatically “right.” The question is which gate the business passed in that system and whether the selected source matches the evidence needed by the query.
Where location quietly changes the retrieved business
French local queries carry geography in a way that can look modest to humans and decisive to retrieval systems. A peri-urban phrase near Lyon may include a commune, a service area, a department reference, or a larger city used as shorthand. Humans tolerate that fuzziness. Retrieval systems often choose one version and build the source trail around it.
In the repair-service composite, the intended business sat outside the largest urban center. Some systems preserved the smaller place signal. Others drifted toward Lyon because the city name had stronger category evidence. Once that drift happened, larger chains and denser listings became more likely to surface. The business retrieved first changed, not because the independent lacked public evidence, but because the location frame had been reinterpreted.
This is why Indexe Clair records location wording as part of the query frame. “Near Lyon,” “in the Lyon area,” a commune name, a department name and a service-plus-city phrase can produce different evidence paths. The lab does not assume one phrasing is the correct one for all cases. It observes whether the system keeps the intended geography or collapses toward the larger place.
A small rough detail matters. Sometimes a system retrieved the right business but paired it with a weak location cue, or selected a review profile that preserved the commune but not the service detail. In other runs, the system named a relevant chain but ignored the independent entirely. Those are different retrieval events. Lumping them under “visible” would erase the problem.
The source conflict can also be subtle. A directory may list an old service area. A review page may use the broader city because customers do. The owned site may use the commune because the business does. The AI search system then has to decide which geography to trust, and that decision can determine which business appears first.
Why model comparison is useful for SMBs
A French SMB does not need a grand ranking of AI search platforms to learn something useful. It can learn from divergence. If the owned site appears in one system and only a stale listing appears in another, the business has evidence that its retrievability is uneven. If no system retrieves a deeper service page, the problem is probably wider than one interface. If only location-sensitive queries fail, the issue may sit in geographic framing or source conflict.
Indexe Clair’s approach turns model comparison into a diagnostic reading practice. The same query is run across systems. The lab records the first retrieved business, the visible sources, the language frame and any source conflict. Then they compare the trails rather than averaging them into a mood. The result is more like laying several transparent maps on a table. The roads do not fully overlap, and the gaps become visible.
The lab is careful with one inference. If a system does not show a source, it may still have used retrieval or internal knowledge. If a system shows a source, that source may be illustrative rather than the only evidence used. Interfaces differ. Some source trails are clearer than others. Still, visible retrieval events are the best public handle available for this kind of study.
A sentence the team could reasonably stand behind is this: for French SMBs, being retrievable in one AI search system does not establish stable visibility across the retrieval layer. That is not pessimistic. It is a warning against over-reading a single successful mention.
The practical consequence is that one-off testing can mislead. A business owner may see their company in ChatGPT Search and assume the market has shifted. Another person may test Perplexity and see only a directory. Both observations can be true within their systems. The work begins when those observations are put beside each other with the same query frame.
Limits of cross-system comparison
This material cannot prove the full index architecture of ChatGPT Search, Perplexity, Copilot or Google AI Overviews. Indexe Clair observes visible outputs and source trails. The systems may change their interfaces, retrieval partners, ranking behavior or source display conventions. A comparison made under one set of conditions is therefore a documented run, not a permanent verdict.
The lab also avoids saying that one system “knows” a business and another does not. A business may be present in an index yet not selected. It may be selected but not exposed as a source. It may appear under a different query frame. The four retrieval gates help preserve those distinctions.
The composite cases are chosen because they resemble common French SMB evidence patterns: owned sites, directories, review profiles, municipal or regional mentions, stale records, and chain competitors. They are not statistical samples of all French commerce. The point is to make the retrieval split readable, not to claim a measured market share of source types.
What remains useful is the discipline of comparison. Same query, recorded language, recorded location frame, visible source trail, rerun where possible. When the retrieved business changes across systems, the lab treats that change as the primary event. The answer’s style can wait.