Following retrieval before answer synthesis

In the lab’s founding case, a supplier outside Tours had a clean French website, current opening hours, product pages and local mentions, and still appeared in AI search through an old directory page. That kind of mismatch is where Indexe Clair begins. The lab does not treat a fluent answer as the main object of study. It asks what had to happen before that answer appeared: which pages were discovered, which entity was indexed, which evidence ranked, and which source was finally selected.

For the lab, an observation is a visible retrieval event. That can be a business name appearing in a response, a page listed as a source, a location signal surviving the query, or a stale record taking priority over an owned site. A conclusion comes later. The team only moves from observation to conclusion after the same pattern appears across comparable prompts, systems or time windows. This keeps the work from becoming a collection of screenshots with dramatic captions.

Samples are formed around French business categories where evidence is already scattered across several public places. A small service firm may have an owned website, a review profile, a municipal mention, a sector listing and a few regional references. Indexe Clair does not pretend that one such case stands for all French commerce. The sample is useful because the same business leaves different traces, and AI search systems often show which traces they can reach and which ones they treat as more dependable.

Repeatability is handled like a field notebook with strict margins. The same query wording is kept, the language is recorded, the location framing is noted, and the system conditions are described as clearly as the interface allows. Variation is expected. AI search is not a vending machine. What matters is whether the variation changes the business retrieved, the source selected or the ranking order. A different sentence is less important than a different evidence trail.

The lab also separates retrieval from answer synthesis. A company can be mentioned without proof that its own site was retrieved. A company can be absent from one answer while still present somewhere in the index. A source can be selected because it is crawlable, fresh, structured, popular, geographically clear, or simply easier for the system to parse. Indexe Clair records these possibilities without flattening them into one quick rule.

The limits are stated in the work itself. AI search interfaces change. Personalization may be partly hidden. Live retrieval can mix with cached knowledge. Some systems expose source trails plainly, while others show only a finished answer. Forecasts are therefore marked as interpretation, not finding. When the lab suggests that a signal may help retrieval, it describes the observed mechanism and the uncertainty around it. A useful claim should leave a handle for someone else to test.

Working principles

1. Observation before conclusion

   A visible page, listing, source trail or business signal is recorded before the lab explains what it may mean. The explanation stays tied to the trace.

2. Retrieval stays separate

   Indexe Clair distinguishes being retrieved from being described well. A polished answer is not enough evidence that the right source was found.

3. Runs must be comparable

   Queries are recorded with wording, language, location framing and system conditions. The lab expects variation, but it checks whether the underlying retrieval path changes.

4. Conflicts are preserved

   Duplicate listings, stale records and regional mismatches are not cleaned up too early. The team first notes which source was chosen, which was ignored and what kind of conflict shaped the result.