A recent memorandum opinion in the U.S. Google search antitrust case details Google's FastSearch technology used to ground Gemini models. The court describes FastSearch as based on RankEmbed signals and returning abbreviated, ranked web results. It also notes that FastSearch retrieves fewer documents faster than standard Search, with lower resulting quality.
Google antitrust memorandum details FastSearch and RankEmbed
The opinion identifies FastSearch as the grounding mechanism for Gemini models and explains how it works with RankEmbed signals.
"To ground its Gemini models, Google uses a proprietary technology called FastSearch."
"FastSearch is based on RankEmbed signals - a set of search ranking signals - and generates abbreviated, ranked web results a model can use to produce a grounded response."
The opinion adds that FastSearch is faster because it retrieves fewer documents, but the resulting quality is lower than fully ranked Search results.
Key details
- The opinion cites testimony identifying FastSearch as the grounding mechanism for Gemini models, with references to the remedy phase transcript.
- FastSearch rapidly generates limited organic search results for specific use cases, such as grounding of LLMs, and is derived primarily from the RankEmbed model.
- RankEmbed is an AI-based deep-learning ranking model with strong natural language understanding.
- An internal document quoted in the opinion notes, "Embedding based retrieval is effective at semantic matching of docs and queries."
- RankEmbed and RankEmbedBERT rely on two main data sources - a redacted portion of 70 days of search logs and scores from human raters used to measure organic search quality.
- RankEmbed was trained on approximately 1/100th of the data used for earlier ranking models while producing higher quality search results.
- RankEmbed particularly improved performance on long-tail queries.
- RankEmbedBERT requires retraining to reflect fresh data.
- The opinion characterizes the data underlying RankEmbed as user-side data, including click-and-query information and human rater scoring.
Background context
United States et al. v. Google LLC challenges Google's conduct in general search and search advertising markets. In 2024, the court found that Google unlawfully maintained monopolies in those markets under Section 2 of the Sherman Act. The case moved to remedies, where additional technical evidence was admitted.
The FastSearch and RankEmbed descriptions relate to grounding outputs from Gemini models. Google introduced AI Overviews in Search in 2024 and has publicly described using Gemini for generative features. The memorandum does not characterize product performance beyond the statements quoted above.
Additional context
SEO practitioner Ryan Jones shared these insights on LinkedIn. His LinkedIn profile provides background on his role.
Source citations
- United States, et al. v. Google LLC, No. 1:20-cv-3010 (D.D.C.), Memorandum Opinion on remedies - excerpts describing FastSearch and RankEmbed.
- Remedy phase transcript citations referenced in the opinion, including Rem. Tr. 3509:23-3511:4 (Reid).
- Internal exhibit references cited in the opinion, including PXR0171 at -086.
.svg)
.svg)
.svg)