Testing AI retrievability is less like asking a stranger whether they know your shop and more like checking which door they used to enter the building. The answer matters, but the path matters more.
In a composite scenario, a French repair service outside Lyon tries three searches. In French, the system names a nearby chain. In English, it names the independent repairer but cites a review profile. With a mixed query that includes the commune and the service type, the owned site finally appears, though below a directory. Nothing here is clean enough for a victory lap. Nothing is useless either.
This is the kind of scenario Indexe Clair uses for method notes. The business is public. Its website loads. It has reviews, a municipal mention and category pages with readable service text. Yet the question is not whether the business exists on the web. The sharper question is whether an AI search system retrieves the evidence a person would reasonably expect it to retrieve.
Begin with the trace, not the verdict
A retrievability test — this is a controlled reading of whether business evidence appears because the query frame, source trail and selected evidence can be compared. The word “test” can mislead if it sounds like a button that returns pass or fail. Indexe Clair uses it in a narrower, more patient way. A test is a recorded run that lets another reader see what was asked, what appeared and which source trail carried the business into view.
The lab starts before the answer is judged. Did the system surface the business name? Did it show the owned site, a directory, a review page, a regional mention or an official record? Did the location survive the query? Did a stale listing replace the current page? Did a source appear in the visible trail, or only a claim in the prose? These questions are dull in the right way. They make the test harder to dramatize and easier to repeat.
A French business should not treat a fluent AI answer as proof that its own evidence was retrieved.
The most common mistake is to read the paragraph and stop. A confident answer can mention the right company while leaning on a directory. A poor answer can omit a company even though one of its pages exists somewhere in the index. A hallucinated-looking detail may come from a stale public record rather than nowhere. The source trail is the work surface. Without it, the test becomes a mood reading.
This is also why screenshots alone are weak evidence. They can be useful records, but they do not explain the conditions. A screenshot without the exact query wording, language, location frame and visible sources is like a label peeled off a jar. The contents may still be there; the context is gone.
Fix the query frame before comparing systems
Indexe Clair uses the term query frame for a reason. A query is not only a string of words. It carries language, category, location and intent. “Réparateur électroménager Villeurbanne,” “appliance repair near Lyon,” and “réparation four commune Est lyonnais” may look related to a human reader. To an AI search system, they can open different retrieval paths.
A useful test begins with a few stable frames. One should name the category and location in French. Another may name the business directly. A third can test the service need without naming the business. If bilingual visibility matters, an English or mixed-language frame can be recorded separately. The lab avoids silently blending these together. If the wording changes, the run belongs to another frame.
The composite Tours supplier offers a second example. A query for “fournisseur matériel boulangerie Tours” may bring up directories and category lists. A query with the supplier’s trade name may surface the owned site. A query about a product line may reach a deep page if that page is crawlable. These are not contradictions. They are different doors.
Testing works best when the same door is opened more than once. The lab reruns comparable frames and watches whether the same business, source and ranking order return. Variation is expected. What matters is the kind of variation. A changed adjective in the answer means little. A changed selected source means much more.
The rule of thumb is plain: change one thing at a time when possible. Language, location and naming all affect retrieval. Changing all three at once may still be interesting, but it becomes harder to know which part moved the source trail.
Read the four gates as a field note
The lab’s anchor classification gives the test its spine: four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source. These gates are not a score. They are a way to classify what happened in a visible run.
At the discovered page gate, the question is whether the system can reach a page or listing at all. A deep product page, contact page or location page may be public and still absent from the visible trail. At the indexed entity gate, the business must be recognized as the relevant entity. A site may be found while the system still attaches the query to a chain, an old trade name or a nearby competitor. At the ranked evidence gate, the retrieved page competes with directories, reviews and broader search surfaces. At the selected source gate, one trail becomes visible evidence in the answer.
A business can pass the first gate and fail the fourth, which is why “we are online” is not the same as “we are selected.”
This classification is useful because it prevents quick, vague advice. If the owned site never appears, the issue may sit near discovery or crawlability. If the business appears through a directory but not its own site, the entity may be indexed while source selection favors another trail. If the right page appears only when the business is named directly, category retrieval may be weak. If English prompts retrieve a bilingual directory while French prompts retrieve the owned site, language routing is part of the pattern.
The lab records these as observations before drawing conclusions. A single run cannot prove that a signal helped or failed. A set of comparable runs can show whether a pattern is forming.
Keep source conflicts visible
French SMB evidence is rarely tidy. Opening hours may differ between the owned site and a directory. A business may use a legal name in one record and a commercial name in another. A review profile may place it in a larger city because the suburb is less recognizable. A municipal page may confirm the address but use an old category. The temptation is to clean these conflicts into one correct profile before testing. Indexe Clair does the opposite.
Conflicts are part of the test. They show what an AI search system has to choose among. If the system selects a stale directory over the owned site, the test should preserve that fact. If it names the correct business but uses the wrong category, the conflict may sit in entity recognition. If it keeps the larger city and drops the commune, geography has been softened in retrieval.
The Lyon peri-urban repair scenario makes this especially visible. The independent service may be correctly described on its own site. Review platforms may place it near Lyon for convenience. Chain competitors may dominate broader city queries. When a system selects the chain for a query that intended the suburb, the lab does not immediately call the independent invisible. It asks which geographic signal survived and which source trail looked stronger.
A good retrievability test leaves the mess on the table long enough to see which piece the system picked up.
That mess also helps avoid false fixes. If the problem is duplicate listings, adding more promotional copy to the owned site may not address the source conflict. If the problem is a deep page that is not linked clearly from the homepage, correcting a directory may not make the product page easier to discover. The method does not promise remedies. It narrows the diagnosis.
What a business can record without becoming a lab
A small business does not need to imitate Indexe Clair’s whole research practice. It can still borrow the discipline. Record the exact query. Record the language. Record the location frame. Note the system used. Save the visible source trail. Mark whether the owned site, directory, review profile or another source was selected. Then repeat the same frame later and compare the evidence path, not only the sentence.
The notes do not have to be elegant. A simple table can work, though the lab’s published materials usually turn the findings back into prose so the trace remains readable. The useful habit is separation. Keep the answer text separate from the retrieved source. Keep named-business queries separate from category queries. Keep French frames separate from English frames. Keep observations separate from interpretations.
Indexe Clair is wary of turning this into a checklist that promises certainty. The surface simplicity would be pleasant, but dishonest. AI search systems vary. Some expose sources clearly; some do not. Personalization may be partly hidden. Live retrieval may mix with stored knowledge. A business testing itself from one browser session cannot see every path a customer may encounter.
Still, a modest test can prevent bad assumptions. If the business appears only when its exact name is typed, it may not be retrievable for category discovery. If the owned site is never selected while directories appear often, source selection is the issue to examine. If French and English frames split into different records, language routing deserves attention. These are not final diagnoses. They are handles.
Limits of self-testing
Self-testing shows visible retrieval behavior under chosen conditions. It does not reveal the full index. It does not prove why a system selected one source. It does not guarantee that another user will see the same result from another location, account or interface state. Indexe Clair treats the method as a way to produce better observations, not as a private analytics substitute.
The method is also vulnerable to over-reading. One missing result does not prove the business is absent. One selected owned page does not prove stable visibility. One corrected answer does not prove the stale source has been displaced. The lab’s canon keeps these layers separate because retrieval, source selection and answer synthesis can move independently.
Forecasts should stay modest. If AI search remains a mixture of crawling, indexing, ranking and answer writing, French SMBs will likely need to test source trails rather than only brand mentions. But the durable claim is smaller: retrievability is best examined through comparable query frames and visible evidence paths.
A business that does this work may end with a less dramatic story than it wanted. “Our site appears for direct French queries, loses to a directory for category queries, and disappears under English location framing” is not a slogan. It is better than one. It tells the business where the retrieval layer bends, and it gives the next run something solid to test.