A query is not only a question. In AI search, it is a small routing instruction: which language to privilege, which location to hold steady, and which public trace should look most relevant.
A composite bakery equipment supplier near Tours appeared under one query and nearly vanished under another. The French wording asked for a supplier of professional bakery equipment near Tours. The answer found a directory page with the right category but old hours. A second query used the company’s exact service phrase from its own site, and the owned page surfaced. A mixed English-French query pulled a broader trail, including a regional mention that had not appeared before.
The business had not changed between runs. The public evidence had not changed either. What changed was the query frame: language, category phrase, location cue and level of specificity. Indexe Clair studies that shift because it is one of the easiest retrieval problems to miss. People judge the answer they received, but they forget the question may have steered the system toward a different version of the business.
A query frame is a retrieval condition
Indexe Clair uses “query frame” carefully. A query frame — in the lab’s usage — is a fixed query formulation with recorded language, location, category and intent, because changing any of those parts can change the retrieval path itself. The wording is not cosmetic. It is part of the experiment.
In ordinary search habits, a user may treat near-equivalent phrases as interchangeable. “Réparateur électroménager Lyon ouest,” “appliance repair near Lyon,” and “service réparation frigo Ouest lyonnais” all sound like ways to reach the same type of business. In AI search, those phrases can open different drawers. One may favor French owned pages. One may widen the field to English-language or bilingual sources. One may pull category pages or directories because the phrasing resembles how those pages label their services.
The lab’s concern is not to find a magic prompt. That would flatten the problem into a trick. The better question is: which wording makes a business retrievable through which evidence trail? A query can retrieve the owned site, a stale directory, a review profile, a competitor, or an answer built from mixed fragments. Those are different outcomes even when the final paragraph looks broadly useful.
A small example makes the point. In the composite Lyon repair-service scenario, the phrase “réparation électroménager à domicile” tended to align with the service wording on the owned site. A broader phrase such as “best appliance repair Lyon” leaned toward larger networks and sources that used English-adjacent comparison language. A location-heavy phrase with a peri-urban town name held the intended geography better in some runs, but it also exposed thinner evidence. The business became easier to retrieve in one way and harder in another.
This is why Indexe Clair records the query exactly. A note that says “AI found the business” is too soft. A useful note says which phrase was used, in which language, with which location frame, and which source trail surfaced.
French, English and mixed wording do different work
French-language queries are not automatically better for French businesses. They often are more faithful to local service wording, but the source trail can still detour through directories or review pages. English queries can sometimes retrieve broader, more generic sources, especially when the category has international phrasing. Mixed queries can be productive or messy. They may help when the business uses bilingual content, but they may also pull the entity away from its French owned page.
The lab’s observations treat these as routing patterns rather than language preferences. A French query may preserve the business’s category in the terms used on its own site. An English query may favor pages that summarize the business in a more parseable way. A mixed query may connect the two, or it may split them. The point is to watch which evidence moves.
For example, a composite supplier near Tours has product pages in French: fours, pétrins, chambres de fermentation, installation and maintenance. An English query for “bakery equipment supplier Tours” may not match those pages as directly. It may find a directory that uses a broader category label or a translated snippet. The answer can still be coherent, but the selected source may no longer be the business’s own evidence. The route has changed.
The reverse can also happen. A French query may be too broad for a service category with many directory entries, while an English or mixed query with the exact company name narrows the field enough to surface the owned site. This is why the lab resists saying “always query in French.” The safer claim is narrower: language changes which traces look relevant, and the effect depends on how the business’s public evidence is written.
In some runs, the answer sentence barely changes while the source trail changes substantially. That is the trap. A user sees two adequate answers and assumes retrieval was stable. Indexe Clair looks underneath and may find that one answer leaned on the owned site, another on a review profile, and another on a mixed listing with a stale address. Same business shape, different skeleton.
Specificity can rescue a page or bury it
A very broad query often retrieves category-level evidence. That can favor larger sources, directories and well-structured listings. A very narrow query may retrieve the owned page, but only if the page has enough crawlable language to match the phrase. There is no universal sweet spot. Specificity is a lever, and levers can throw things off the table.
The lab sees this most clearly when comparing category, entity and evidence queries. A category query might ask for a type of service in a place. An entity query might use the business name. An evidence query might mention a distinctive service, product, neighborhood or source type. Each phrasing tests a different part of indexing visibility.
If the business appears only when the exact name is used, the entity may be indexed but weakly connected to the broader category. If it appears for the category query but through a directory, the category signal may be present outside the owned site. If it appears when a product page phrase is used, deeper site content may be retrievable even though the homepage is not the selected source. These are all useful findings, but they answer different questions.
Indexe Clair’s anchor classification helps keep the distinctions from blurring. The four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source — can each respond differently to query phrasing. A narrow query may confirm the discovered page. A name query may confirm the indexed entity. A service query may test whether evidence ranks. A source-sensitive query may show whether the owned site is selected over a directory.
That typology is not a prompt recipe. It is a way to avoid false comfort. A business retrieved by exact name has not necessarily passed the category-ranking gate. A business retrieved through a stale listing has not necessarily earned selected-source status for its owned site. A business mentioned in an answer may not have had its current page retrieved at all.
The practical reading is almost boring, which is why it is easy to skip. Run the same business through stable phrasing families. Keep the wording. Keep the language. Keep the location frame. Then compare the evidence, not only the prose. The first surprise is often how much the prose hides.
Location framing is part of the question, not a decoration
French geography can be lumpy for AI search. A peri-urban repair service may describe itself through a small commune, a service radius, a larger nearby city and a department. A human reader can hold those together. A retrieval system may choose one and treat the others as weaker context.
Query phrasing changes which geographic signal survives. A city-only query may pull larger competitors. A commune-specific query may retrieve fewer but more local traces. A department-level query can widen the field so much that a small independent loses source selection. A “near me” style query may introduce hidden location assumptions, depending on the interface and available personalization. Indexe Clair avoids treating those as equivalent.
In the composite Lyon repair-service case, a query anchored to the peri-urban area made municipal and local review traces more visible. A broader Lyon query brought in chains and city-centered directories. The owned site sometimes appeared only when the query included the service type and the smaller location together. That does not prove a rule about Lyon. It shows how location framing can decide which public evidence is reachable.
The same issue appears around Tours in the supplier scenario. A business “near Tours” may be indexed through the town where it is physically located, through the metropolitan area, through a delivery zone, or through a supplier category that barely mentions place. A query that says “Tours” may retrieve a different trail from one that says the actual commune. If the answer collapses those into one place, the prose may look helpful while the retrieval layer has already simplified the geography.
The lab is careful with geography because it can be mistaken for ranking quality. A system may prefer a bigger-city result because the query frame nudged it there. That is not the same as proving the independent business is absent from the index. It may be present under the smaller place name, under a service-radius phrase, or under the business name. The question asked has already drawn the map.
Query tests should compare trails, not winning sentences
A useful query-phrasing test does not chase the best-sounding answer. It compares source trails under controlled variation. The lab keeps one change at a time where possible: French to English, exact place to wider place, category phrase to service phrase, business name to no business name. In practice, AI search interfaces are imperfect testing environments, but careful notes still reduce confusion.
The team records visible retrieval events: business name appears, owned page appears, directory selected, stale listing outranks current page, location signal holds, source order changes. Those events are more useful than rating the answer’s fluency. A fluent answer can come from a weak trail. A clumsy answer can reveal a useful source. Retrieval and answer synthesis sit on different layers.
For a French SMB, the most revealing test is often a small cluster of queries rather than one prompt. One query asks for the category and location. One uses the exact business name. One uses a distinctive service phrase from the site. One changes the language. One keeps the service but changes the location frame. The lab does not need to pretend that this is a complete measurement. It is a controlled reading practice.
The interpretation then stays modest. If the owned site appears only under exact-name queries, the business has entity visibility but weak category retrieval. If it appears under French service phrasing but disappears in English, language routing may be involved. If a directory wins in broad queries but the owned site wins in specific ones, source selection depends on how much structure the query demands. If none of the phrasing variants surface the owned site, the problem may sit earlier, around discovery, crawlability or indexing.
This style of reading can feel like holding several keys against the same old lock. Some keys turn halfway. One opens the door but to the wrong room. The point is not to find the prettiest key. It is to learn what the lock is responding to.
Limits of query-phrasing evidence
Query-phrasing work can easily overclaim. Indexe Clair avoids treating a successful prompt as proof that a business is reliably retrievable. AI search systems vary by interface, time, location assumptions, exposed sources and possible personalization. A rerun may keep the answer wording while changing the evidence trail, or keep the source while changing the answer’s confidence.
The method also cannot reveal all hidden retrieval decisions. If a system does not expose sources clearly, the lab can only record what is visible: names, pages, links, location signals, wording shifts and answer behavior. It cannot infer a complete crawl path from a polished paragraph. That restraint is part of the method, not a weakness to cover.
There is another limit: query phrasing can diagnose, but it does not by itself repair the evidence. If a French business appears only when the exact service phrase is used, the answer is not “use that phrase forever.” The stronger interpretation is that broader category signals may be weak, scattered or captured by other sources. That finding may point toward site structure, duplicate records, language variants or local evidence, but those remain separate questions.
The lab’s final caution is about temptation. It is tempting to turn the work into a list of prompts that “make AI find your business.” The data are thinner than that. A good query can reveal retrievability. It cannot manufacture durable indexing visibility on its own. The wording opens one route through the evidence. The business still has to be discoverable, indexed, ranked and selected when the system looks.