Google Antitrust Case Reveals AI Overviews Use FastSearch, Not Links
Court documents show Google’s AI Overviews rely on FastSearch, powered by RankEmbed signals, not traditional link-based ranking.
2 min readHighlights
Google’s AI Overviews use FastSearch, not the main Search algorithm.
FastSearch relies on RankEmbed deep-learning signals rather than links.
Shift suggests semantic relevance and user data may outweigh backlinks in AI ranking.
Source: Image created by Martech Scholars_Google’s AI Overviews Depend on FastSearch Algorithm, Not Links.
A recent twist in the ongoing Google antitrust case has shed new light on how the tech giant powers its AI Overviews feature. Instead of relying on the traditional web search algorithm that factors in backlinks, Google’s FastSearch technology uses semantic relevance and deep learning signals to generate quick grounding results.
The revelation comes from a Memorandum Opinion highlighting how Google grounds its Gemini models for AI answers. Search marketer Ryan Jones spotted a passage explaining that FastSearch retrieves fewer documents than regular search, prioritizing speed over traditional ranking signals such as links.
Grounding Generative AI With FastSearch
Ordinarily, one might expect AI Overviews to pull data from the same web ranking system used by Google Search, with backlinks playing a key role. But according to testimony, FastSearch works differently.
The memorandum states:
“FastSearch is based on RankEmbed signals—a set of search ranking signals—and generates abbreviated, ranked web results that a model can use to produce a grounded response.”
Because it skips many signals, including spam detection layers, this explains why early AI Overviews occasionally displayed spammy or low-quality pages.
Jones notes:
“For grounding Google doesn’t use the same search algorithm… They just need text that backs up what they’re saying.”
What Is RankEmbed?
RankEmbed is a deep-learning ranking model that helps Google identify semantic relationships between queries and documents. It uses:
- Click-and-query data from search logs.
- Human rater quality scores to train its models.
This model was designed to handle long-tail queries, providing more relevant results even if query terms don’t exactly match.
Importantly, the memorandum highlights that RankEmbed is trained on a fraction of the data of older ranking systems, yet offers higher efficiency.
Implications for AI Search
If FastSearch avoids traditional link signals, it suggests a significant shift: Google’s AI Overviews may no longer depend on backlinks as a major ranking factor. Instead, semantic understanding and user interaction data appear to guide result selection.
This raises broader questions about:
- How AI-driven ranking changes SEO strategies.
- Whether multiple indexes exist, one for FastSearch and another for core search.
- How much influence human raters’ training data has in shaping AI results.
For now, one thing is clear: Google’s AI Overviews represent a new frontier in search, prioritizing speed, semantics, and user-side data over traditional link-based ranking.
