The page's best passage scores clearly lower cosine similarity to its target queries than the passages on pages that rank for them. Testimony and the 2025 call notes describe RankEmbed as a dual encoder that retrieves and ranks by distance between query and document embeddings. Open embedding models are only a proxy for Google's model, so this check compares relative fit, not RankEmbed scores.
Why does low semantic similarity between target queries and page passages (rankembed proxy) matter for AI search?
Testimony and the 2025 call notes describe RankEmbed as a dual encoder that retrieves and ranks by distance between query and document embeddings. Open embedding models are only a proxy for Google's model, so this check compares relative fit, not RankEmbed scores.
How do you fix it?
Rewrite the weakest sections to address the query's meaning directly: state the topic, name the entities and relationships involved, and answer the likely questions. Split mixed-topic sections.
How do you find it on your site?
- Take the page's top 10 queries from Search Console and the top 5 ranking URLs for each.
- Split all pages into passages of roughly 100 to 200 words.
- Embed queries and passages with the same open sentence embedding model and record each page's best passage similarity per query.
- Flag queries where the site's best passage is in the bottom quartile of the ranking set.
Which ranking and citation factors does it touch?
Written for the RankEmbed and RankEmbedBERT page (evidence: Trial or leak evidence). Google has not documented this system in a way a site can test directly, so this check measures a proxy for it, not the system itself.
What is the impact?
Indirect. Estimate only: better semantic fit should help retrieval and ranking for long-tail variants.
What is the evidence?
PXR0357 call notes (31 January 2025), Nayak's 18 October 2023 testimony, the 5 August 2024 and 2 September 2025 opinions, and the RankEmbed module in the May 2024 leak.
Sources
- PXR0357, notes of 31 January 2025 call with Google engineer Pandu Nayak, US Department of Justice
- Transcript of bench trial, Day 24 afternoon session, 18 October 2023, United States v. Google LLC
- Memorandum Opinion, United States v. Google LLC (D.D.C., 5 August 2024), CourtListener
- Memorandum Opinion on remedies, United States v. Google LLC (D.D.C., 2 September 2025), CourtListener
Frequently asked questions about low semantic similarity between target queries and page passages (rankembed proxy)
What does "Low semantic similarity between target queries and page passages (RankEmbed proxy)" mean?
The page's best passage scores clearly lower cosine similarity to its target queries than the passages on pages that rank for them.
How do you fix low semantic similarity between target queries and page passages (RankEmbed proxy)?
Rewrite the weakest sections to address the query's meaning directly: state the topic, name the entities and relationships involved, and answer the likely questions. Split mixed-topic sections.