RESONEO has published a list of 305 groups of signals that Google records in RankLab, the tool where the search engine logs the variables calculated by its ranking systems, called “features.” According to the agency, 167 of these groups apply to each result and 138 to each query. The schemas give the names of the systems, but not their weight.
Olivier de Segonzac presented the list on November 11, 2026, in a LinkedIn post and in an article published on think.resoneo.com. This text is the eleventh part of the agency’s “Understanding Google” series. According to him, the list comes from Google’s internal code that RESONEO analyzed, and he writes that he has not “seen it published anywhere else.”
The author starts from a figure known to all SEOs. In 2008, Hal Varian, Google’s chief economist, mentioned “more than 200 signals” used by the ranking algorithms. In September 2021, John Mueller explained that Google had moved away from this, as the figure suggested a spreadsheet where signals would be ranked by importance.
According to RESONEO, Google instead organizes its signals by system, such as RankBrain, SafeSearch, topical authority, YouTube videos, or AI Mode responses. Based on the documentation from the 2024 leak, Google records these features in production and uses them for its “ranklab models,” its ranking models. The agency notes limitations, however: the schemas come from several packages dated 2025 and 2026, and only three groups per result detail their fields. The French labels and categories were established by RESONEO.
The web page concentrates 48 groups, generative AI accounts for 34
Per result, 48 groups relate to a web page, of which 31 are used to rate it: quality, demotions, freshness, personalization. Generative AI comes second with 34 groups (AI Overviews, AI Mode, Web Guide), ahead of Discover and news (19), images and Lens (17), or video and YouTube (14). Each format has its own systems: a video goes through Toledo, which classifies it by type (learning, entertainment, information seeking).
RESONEO then describes the relevance calculation in four steps. RankEmbed retrieves candidate documents by semantic proximity, including when the query does not contain some of their terms. Ascorer, the scorer of Mustang, the web page ranking system, assigns a base score. RankBrain re-ranks the top results, then DeepRank, i.e., BERT applied to ranking, scores the whole set. Finally, the twiddlers, programs that modify already-scored results, intervene last. The agency cites the testimony of Pandu Nayak, then Vice President of Google Search, during the US antitrust trial on October 18, 2023: RankBrain examines the top 20 or 30 documents.
Google has also cited several quality notions, and RESONEO associates a distinct signal group with each: expertise, topical authority, first-hand content, “hidden gems,” authentic points of view, YMYL topics (health, money, safety) or discussions and forums. Seventeen systems rate both the page and the query. Topical authority, presented by Google in May 2023 for news, actually applies to any web page. Google thus measures authority topic by topic: a running site is authoritative on a runner’s injury, not on a mortgage.
RESONEO’s Tips for SEOs
The agency draws three recommendations from this, which it presents as stemming from its analysis and not the schemas themselves. It first advises targeting the quality notion that corresponds to the page, and demonstrating expertise on YMYL topics, with an identified author, sources, and a specialized site. It then suggests prioritizing updates for pages targeting queries that require freshness. Finally, it believes one must first rank in the results before aiming for a citation by AI: for “knee pain while running,” five of the seven sources in the AI Overview also appear among the first-page blue links.
The list does not say what weight to give to each system. RESONEO clarifies this in its method: the page describes what each system does, not the importance it holds in the ranking. It therefore remains impossible to know, for example, what place corroboration occupies in AI Mode relative to other signals.
