Markup is attractive because it looks like a direct instruction to machines. In AI search retrieval, it behaves more like a label sewn inside a jacket: useful when the jacket is already visible, less useful when the garment is still in the wrong cupboard.
A French business owner can add schema markup and feel, for a brief afternoon, that the technical layer has been handled. The business name is in a structured field. The address is marked. The opening hours are tidy. The category has a recognizable type. Then an AI search prompt still selects a directory listing, or an old review profile, or a competitor page with plainer markup but stronger public evidence.
Indexe Clair studies that disappointment carefully because it is easy to misread. In a composite bakery equipment supplier near Tours, the owned site has product pages and local mentions, but a stale directory entry keeps appearing in some AI search trails. If the site later adds structured data, the lab does not ask only whether the final answer improves. It asks whether the owned page crosses a different retrieval gate: discovered page, indexed entity, ranked evidence, selected source. Markup may help at one gate and fail at another.
Structured data is a support signal, not a magic hatch
Structured data is machine-readable markup that describes page entities because ordinary page text can be ambiguous, incomplete or difficult to classify. For a French SMB, it might identify a business name, address, phone number, opening hours, organization type, service area, product category or breadcrumb structure. It can make evidence less smeared across the page.
The trap is assuming that a clearer label automatically changes AI search retrieval. A label attached to a buried page may remain buried. A well-marked business with conflicting directory records may still lose source selection. A markup field that says one category while visible text says another can even add a little fog. The lab’s position is deliberately plain: structured data is worth observing as part of the evidence trail, but it should not be treated as a standalone ranking rule.
In source-trail reading, the first question is whether the page itself appears. If an owned site is selected after markup is added, the lab checks what else changed. Was the page updated at the same time? Were internal links improved? Did the business name become more consistent? Did a directory record also change? Did the query frame shift from broad category to exact name? Without those controls, “schema made it appear” is too clean a story.
This is where many quick tests go soft. A person adds markup, reruns a prompt, sees a better result, and assigns the change to markup. That may be true. It may also be freshness, crawl timing, clearer visible text, a changed system interface, or retrieval variation. Indexe Clair does not discard the observation. It refuses to promote it too early.
The four gates give markup a more precise role
The lab uses the anchor classification of four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source. Structured data can plausibly help each gate in a different way. It may help discovery indirectly if markup sits on pages that are easier to parse and link. It may help entity indexing by aligning name, address and category. It may help ranking when the structured evidence reinforces the visible content. It may help source selection if the owned page becomes the clearest source among competing trails.
But those are separate possibilities. A page can have markup and still not be discovered if it is poorly linked. A business can be recognized as an entity but not ranked for the category query the owner cares about. Evidence can rank but lose source selection to a directory with stronger external signals. The four gates keep the discussion from collapsing into one yes-or-no claim.
In the composite Tours supplier, markup around the organization and product categories would be most interesting if it reduced entity ambiguity. The business has an owned site, local mentions, and a directory with stale hours. If structured data repeats the current address, category and opening hours in a machine-readable way, it may strengthen the owned page as evidence. Yet the stale directory may still be selected if the query is category-first and the directory has a more familiar listing format.
For the composite Lyon peri-urban repair service, the geographic part becomes sharper. Markup may state the address and service area, but AI search still has to interpret “near Lyon,” the suburb name, and competitors with stronger city-level pages. Structured data can help name the service area. It cannot force a system to prefer a peri-urban independent when the query frame and source density pull toward central Lyon.
That is the pattern Indexe Clair tries to preserve: markup can clarify a source; it does not automatically make that source win.
What the lab looks for on marked pages
The lab reads structured data beside the page, not as a hidden talisman. If a French business page marks itself as a local business but the visible copy barely names the service, the evidence is thin. If the markup gives opening hours but the page shows a different schedule, the source conflict becomes part of the record. If breadcrumbs, service pages and product pages all point to the same entity, the structured layer may support what a crawler can already see.
The most useful cases are usually boring. The business name is written consistently. The address is crawlable. The main service or product category appears in human text and in structured fields. Internal links connect the homepage to deeper pages. The page title, headings and body do not fight each other. The markup is not asked to carry the whole identity of the business on its back.
Indexe Clair records whether marked pages appear as visible retrieval events. That can mean the owned page is listed as a source, the business name appears with current details, the location signal survives the query, or the source trail shifts from a directory to the owned site. None of those proves markup alone caused the change. They show where structured evidence may have entered the path.
The lab also watches pages where markup exists but retrieval remains weak. Those cases are useful because they puncture a common assumption. A page can be technically annotated and still lack enough crawlable text. It can mark the business correctly while external sources disagree. It can use generic organization markup where the retrieval problem is actually a product-category problem. It can include a service area so broad that the local signal becomes mush.
There is a craft issue here. Structured data written like a tax form may be correct but unhelpful for the query a buyer uses. The system is not simply looking for a valid field; it is assembling evidence for a task. A bakery supplier needs product and trade language. A repair service needs service and geography language. Markup that does not reinforce those retrieval needs may sit politely in the page and do little visible work.
When markup appears to matter most
In the lab’s qualitative readings, structured data seems most meaningful when the business is already close to retrievable. The page is crawlable. The entity is not deeply confused. The owned site has some internal structure. External records are not completely contradictory. In that situation, markup may help the page become a cleaner candidate for ranking or selection, especially when it repeats the same facts visible elsewhere on the page.
A different pattern appears when a business has many duplicates. Suppose a French repair firm has an old directory address, a review profile under a previous name, and a current site with clear markup. The structured data on the current site may help state the preferred entity, but AI search still has to resolve the conflict across sources. If the stale listing is selected, the problem is not that markup has no value. It is that source conflict remains stronger than page clarification in that run.
Freshness is another nearby mechanism. Adding schema often happens during a page update. The page gets edited, republished, internally linked, maybe resubmitted in search tools, and discussed by the owner as “the schema update.” If retrieval later improves, several signals moved at once. Indexe Clair marks that as interpretation, not finding. The observed mechanism may be structured clarification plus freshness plus crawl timing. Pulling one thread and calling it the whole cloth would be false neatness.
The lab also sees a difference between exact-name prompts and category prompts. Markup may help an exact-name query connect a page to the correct entity. Category prompts still require the page to compete as evidence for a business type in a place. A marked organization page with weak service wording may appear for the business name and vanish for “fournisseur matériel boulangerie Tours.” That split is not a failure of structured data alone; it shows which retrieval question the page answers.
A useful sentence for teams to keep close is this: structured data helps most when it confirms what the page already says clearly. When it tries to replace clear page evidence, it starts to look like a label on an empty box.
How this changes the way French SMBs should test
A responsible test does not begin with “add schema and see what happens.” It begins with a baseline. Which queries currently retrieve the business? Which sources are selected? Does the owned site appear, or only directories and review pages? Are French, English and mixed query frames behaving differently? Is the problem discovery, entity indexing, ranked evidence or selected source?
After that, a markup change can be observed more cleanly. The lab would record the page state, query frame, language, location framing, visible sources and date of the run. Then it would repeat comparable prompts after the change, while noting that AI search systems vary and crawl timing is not fully visible. The question becomes modest: did the visible source trail change in a way consistent with clearer structured evidence?
This kind of reading also prevents overbuilding. A small business does not need to turn every page into a maze of markup. It needs the right facts to be legible where retrieval is most likely to need them. For a supplier, that may be organization, address, product or service categories, contact details and breadcrumbs. For a local repair firm, local business details, service area and service pages may be more relevant. The exact implementation belongs to technical work; the retrieval question is whether the structured layer supports the source trail.
Indexe Clair stays away from claiming that one markup type produces a specific AI search outcome. The lab is not running a controlled measurement platform with private ranking visibility. It is reading visible retrieval events. That means the output is less dramatic but more useful: markup may have helped here; it did not overcome source conflict there; it supported entity clarity in one query frame and failed to change category retrieval in another.
The difference between those statements is the difference between research and reassurance.
Limits of the structured-data reading
The method cannot inspect the private internals of ChatGPT Search, Perplexity, Copilot, Google AI Overviews or other AI search systems. It cannot know exactly whether a system used structured data, visible text, external links, cached knowledge, directory records or a combination unless the source trail makes part of that visible. Even then, source exposure varies by interface.
Structured data is also hard to isolate in real SMB cases. It is often added during broader site maintenance. Pages get rewritten, internal links change, stale records are corrected, and search systems themselves shift. A later retrieval event may be associated with structured data without being caused by it alone. The lab therefore marks causal claims as interpretation unless comparable runs show a repeated pattern and other changes are limited.
There is one further caution. Valid markup can describe a business badly. A field can be technically correct and still too generic for the query. A service area can be so broad that it weakens local intent. A category can be chosen because it fits a schema vocabulary but not the words French buyers use. AI search retrieval reads evidence in context. Markup is part of that context, not a private tunnel into the answer.
The practical conclusion is deliberately restrained: structured data deserves attention when a French business is close to being retrieved but its entity, category or location evidence is messy. It is less likely to rescue a page that cannot be discovered, has little crawlable content, or sits inside a tangle of conflicting public records. The page still has to be found, understood, ranked and selected. The label helps most when the thing it labels is already in reach.