A local query looks simple from the keyboard. In retrieval, it becomes a geography test: does the system keep the village, suburb or service area, or does it slide toward the larger place that has denser evidence?
A repair service can be ten minutes outside Lyon and still disappear into Lyon. That is the shape of the problem Indexe Clair keeps returning to. The business has a site, a review profile, a municipal mention, and service pages that name several nearby communes. A person asks an AI search system for help near the city. The answer comes back with chains inside Lyon, a directory page for the department, or a business in a neighboring suburb with a stronger listing trail.
In the lab’s composite Lyon peri-urban scenario, the independent repair service is not imaginary in the sense of a clean classroom example. It is assembled from recurring patterns: small service firms outside major French cities, crawlable French websites, uneven directory records, and competitors with broader city labels. The rough edge matters. The business is not perfectly documented. One page says “intervention autour de Lyon,” another names the suburb, a review profile uses a department label, and a municipal page gives the older address format. AI search has to decide which geography counts.
French local intent is not just a map pin
People often speak about local search as if it were a pin dropped on a map. AI search behaves more like someone sorting envelopes on a kitchen table: town names, service areas, departments, directories, review snippets and query wording all land in the same pile. A precise location can get covered by a larger one if the larger one has more retrievable evidence.
Geographic intent in AI search is the system’s interpretation of place, distance and service relevance because a query rarely states all three cleanly. A user may write “near Lyon,” “autour de Lyon,” “réparateur Villeurbanne,” “dépannage électroménager Rhône,” or a mixed phrase with “near me.” Each frame suggests a different geography. The system then has to choose whether to privilege exact place, broader metro area, administrative region, service radius, or source authority.
This is where French geography becomes particularly fiddly. France has dense local naming: communes, arrondissements, departments, historical place names, peri-urban belts, and business service areas that do not match administrative borders. A rural business may identify itself through a nearby town because customers recognize that town. A directory may assign it to a department. A review profile may tie it to a postal code. The owned site may use a phrase such as “près de Tours” or “secteur Lyon Est.” AI search has to flatten or preserve those signals while retrieving sources.
Indexe Clair does not treat a wrong city in an answer as only a writing mistake. It can be a retrieval event. If the selected source is a directory that groups several communes under a larger city, the answer may inherit that geography. If the owned site is selected, the location may stay finer. The written sentence is the last trace of a choice made earlier.
City gravity and the peri-urban slide
The first pattern the lab watches is city gravity. A major city name in a query can pull retrieval toward sources that are organized around that city, even when the user’s intent allows nearby areas. Lyon, Marseille, Lille, Bordeaux, Nantes, Toulouse and Paris have this effect in different ways. The city name becomes a magnet on a table full of paper clips. Smaller places nearby may be relevant, but the denser city trail moves first.
In the composite repair-service case, a French prompt naming the exact suburb may retrieve the independent business or its service page. A broader prompt such as “réparation électroménager près de Lyon” often makes the field noisier. Chains with many location pages, directories with thick category listings, and review surfaces with city-level pages may appear before the independent. The business has not failed to exist. Its location signal has been diluted by the query frame.
The lab classifies the pattern through the four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source. A peri-urban service firm can have a discovered page and an indexed entity, but its evidence may be ranked under a narrower geography than the user’s broader city query. Source selection then goes to a city-level directory or chain page, because that source looks better aligned with “near Lyon” as the system interpreted it.
This is different from saying AI search “prefers cities” in every case. A precise named-business query can still retrieve the right source. A strongly structured local page may survive the broader frame. Some systems handle smaller places better when the prompt includes a category and a town together. The finding is narrower: broad city framing can erase the practical geography of businesses that operate around the city but are not labeled as city-center entities.
The slip is often visible in wording. The answer may say a business serves Lyon when the source trail points to a suburb. It may group several surrounding firms under a city heading. It may select a directory page whose geography is broader than the actual business address. None of this requires malice or hallucination. It is the ordinary compression of local evidence into a retrievable shape.
Rural queries have a different failure mode
Rural retrieval does not always suffer from city gravity. Sometimes it suffers from evidence thinness. A business in a small commune may have an owned site, but fewer external mentions, weaker review surfaces, and directory entries that point to a department or nearest larger town. The system then faces a sparse trail. It can preserve the small place and risk a thin answer, or widen the geography and retrieve richer sources.
Indexe Clair has seen this in composite cases around suppliers, trades and local services. A French query with a precise commune may return the owned site or no confident source trail. The same query with “near Tours” or “near Angers” may produce a more fluent answer, but the selected sources become less local. The answer improves in polish while the retrieval path drifts away from the business the user might actually need.
A useful clue is whether the system keeps the location signal from the query in the cited or visible sources. If the prompt names a commune but the selected source names only a department, the retrieval event has broadened. If the prompt asks for a business “near” a city and the answer selects only firms inside the city, the system has narrowed in a different direction. Both are geographic reinterpretations, and both can change which business becomes visible.
This is why Indexe Clair separates the geographic signal from the business name. A system may retrieve the right category and the wrong place. Or it may retrieve the right place and a weak category match. Or it may retrieve a strong directory that blends several places into one page. The answer can look reasonable while the source trail tells a more crooked story.
The lab avoids a neat rural-versus-urban morality tale. Some rural businesses are very retrievable because their public evidence is unusually clear. Some city businesses are buried under duplicates and stale listings. Geography matters, but it works through source density, wording, and ranking conflicts rather than through a simple rule.
“Near me” without a stable “me”
The phrase “near me” is awkward in AI search because the system may not expose how it handles location. It may use device location, account setting, IP-level inference, a query location, or none of these in a visible way. For a lab that cares about repeatability, that creates a problem. A “near me” prompt is only useful if the location frame can be stated or controlled.
Indexe Clair therefore rewrites “near me” tests into explicit location frames where possible. Instead of relying only on “near me,” a run might compare “réparateur électroménager près de Lyon,” “réparateur électroménager à Bron,” and “appliance repair near Lyon.” The point is not to make the query less realistic. It is to make the observed retrieval path readable. If a run cannot be repeated with the same location meaning, it becomes a loose anecdote.
The lab also watches how AI search handles French prepositions and proximity words. “À Lyon” is not the same as “près de Lyon,” and “autour de Lyon” is not identical to “dans le Rhône.” A human buyer may use them loosely. A retrieval system has to turn them into candidate sources. In some runs, “près de” opens the field to surrounding businesses. In others, it still collapses toward the main city because the city has stronger evidence surfaces.
Mixed-language prompts add another layer. “Near me réparation Lyon” may blend English proximity with French category wording and a city name. The system may keep the city and category but substitute sources that are easier to parse in English, especially directories or review pages. That behavior overlaps with the language-routing question studied elsewhere in the lab’s index, but the geographic part deserves its own reading: which place survived the prompt?
The most useful local prompt is often not the most natural one. A person may say “near Lyon,” but a repeatable test needs to ask whether the business should be found by commune, by service area, by department, or by category plus city. Those frames answer different questions.
What a French business can inspect in its evidence
For a French SMB, geographic retrievability begins with the boring parts of public evidence. Does the owned site state the address in crawlable text? Does it name the real service area without hiding it in a map image? Are local pages internally linked, or are they orphaned? Do directory records use the same town, category and address? Is the business described as being in the suburb, the city, the department, or all three in conflicting ways?
The lab does not present these as a checklist that guarantees retrieval. They are places to look when a geographic signal keeps slipping. If the owned site says “Lyon Est,” the directory says “Lyon,” and the review profile says a postal-code commune, an AI search system may select whichever version is easiest for the query. The business may think it has broad local coverage; the retrieval layer may see three competing geographies.
Service-area wording is especially delicate. A trade business may honestly serve twenty communes. Listing all of them in a footer can look crawlable but noisy. Naming none of them leaves the system dependent on external records. The lab’s view is conservative: the strongest pages tend to make the main location and service area legible in ordinary text, with internal links that connect category, place and business name. A page that reads like it was written for a person is often also easier to retrieve.
The owned site is only one piece. Municipal mentions, sector pages, review profiles and directory records can reinforce or distort geography. A stale listing with an old town can still be selected if it is structured and prominent. A regional article can keep a location signal alive if it names the business clearly. In local AI retrieval, geography is rarely a single field. It is a chorus with a few people singing late.
Limits of the geographic reading
This method cannot show the private distance calculations of AI search systems. It cannot always tell whether a result appeared because of live location, cached knowledge, account personalization, query rewriting, or source authority. Some interfaces expose source trails; others do not. Even when sources are visible, the ranking process behind them remains partly hidden.
Indexe Clair therefore treats geographic behavior as observed retrieval, not as a map-science claim. The lab can say that a query frame preserved, broadened or collapsed a place signal in visible results. It should not claim that a system measured distance in a particular way unless the interface or source trail clearly supports that reading. The difference matters. A local answer can be geographically plausible and still rely on a source trail that is too broad for the business question.
The interpretation is strongest when comparable prompts are run with stable wording, language, location framing and system conditions. A single “near me” screenshot tells very little. A set of French, English and explicit-place prompts can show whether the same business survives different geographic frames. If it disappears only when the larger city is named, the problem is likely not absence from the web. It is the way local evidence gets ranked when city gravity enters the room.