RankEmbed is a Google deep learning model that places queries and documents in the same embedding space so it can retrieve and rank by meaning rather than exact words; RankEmbedBERT is its later version. It is known from trial testimony, Judge Mehta's 2024 and 2025 opinions and a leaked module name, not from Google's public documentation. For sites, it means pages can be retrieved for queries they never state verbatim, so clear topical meaning matters more than keyword repetition.
What is RankEmbed?
RankEmbed is a Google ranking model that represents both the query and the document as vectors in a shared embedding space. Notes of a 31 January 2025 call with Pandu Nayak (exhibit PXR0357) describe it as "a dual encoder model that embeds both query and document into embedding space", with retrieval and ranking done by a dot product. RankEmbedBERT is its later iteration, per the 2025 remedies opinion.
In the 2024 liability opinion, Judge Mehta lists RankEmbed with RankBrain, DeepRank, RankBERT and MUM as newer "generalization" systems that rely less on user data than older signals.
RankEmbed is a dual encoder model that embeds both query and document into embedding space.
Notes of call with Pandu Nayak, PXR0357, notes of 31 January 2025 call with Google engineer Pandu Nayak, US Department of Justice
How does RankEmbed work?
Unlike most ranking signals, RankEmbed also retrieves. Nayak told the court in October 2023 that RankEmbed is one of the deep learning systems that "also does retrieval", which means a document can be found without containing every query term. The remedies opinion says this lets the model identify good documents "even if a query lacks certain terms".
Its training mix is click-and-query logs plus human rater scores. The remedies opinion says the data is a (redacted) percentage of 70 days of search logs plus rater scores, and that RankEmbed was trained on 1/100th of the data used by earlier ranking models while producing higher quality results. The PXR0357 notes say Google trained it on "a sample from a single month of search data". The two accounts differ on the window; both describe far less data than Navboost's 13 months.
The leak includes a QualityRankembedMustangMustangRankEmbedInfo model with a compressed "quantized document embedding" attached to documents in Mustang, Google's scoring system, which fits a precomputed document vector.
Known as “generalization” systems, these signals “may not be so good at memorizing facts, but they’re really good at understanding language.”
Judge Amit P. Mehta, quoting Eric Lehman, Memorandum Opinion, United States v. Google LLC (D.D.C., 5 August 2024), CourtListener
What evidence is there that Google uses RankEmbed?
The label is Trial or leak evidence. In October 2023 Nayak agreed with DOJ counsel that the three main deep learning models Google uses in ranking are RankBrain, DeepRank and RankEmbed BERT, and that all three are trained in part on click and query data. A 2020 Google presentation admitted as UPX0219 lists RankEmbed among the machine learning ranking systems that "rely on logs". The 2025 remedies opinion treats RankEmbed as one of Google's top-level signals, finds that Google's Gemini grounding tool FastSearch is based on RankEmbed signals, and requires Google to share RankEmbed training data (not the model) with qualified competitors.
The testimony on tail queries is mixed. The remedies opinion quotes Nayak saying RankEmbedBERT "particularly helped with long-tail queries", while the 2025 call notes say the model is "high quality on common queries but can perform poorly for tail queries". Both are in the record; neither is a public Google statement.
As I mentioned, not one system, but a great many within ranking are built on logs.
Google presentation speaker notes, UPX0219, Google presentation: Logging & Ranking, US Department of Justice trial exhibit
What does RankEmbed mean for your site?
Embedding retrieval rewards pages whose overall meaning is close to the query's meaning. A page that covers the topic clearly, with the entities, relationships and questions a searcher expects, sits closer in embedding space than a page that repeats the phrase but drifts off topic.
This is testable with open models. Embed your target queries and your page's passages with any sentence embedding model and compare cosine similarity against the pages that rank. The claim to falsify: pages with materially lower query to passage similarity than the ranking set should underperform for those queries. Open models are not RankEmbed, so treat the scores as relative, not absolute.
- State the topic plainly in the title, H1 and first paragraph.
- Keep each section about one sub-topic so passages embed cleanly.
- Cover the questions users ask, not only the head term.
Which Laurelin audit checks test for RankEmbed?
RankEmbed cannot be observed directly. The new check, low semantic similarity between target queries and page passages, measures a proxy with open embedding models.
Related checks: content not matching search intent catches the largest meaning gap; passages not self-contained affects how sections embed; no clear answer to the page's question weakens query to document match; vague headings blur section meaning; low information gain versus competitors addresses pages that match but add nothing.
When did RankEmbed become public?
RankEmbed was named in open court during Nayak's testimony on 18 October 2023. It appears in Judge Mehta's liability opinion of 5 August 2024, in the leaked API documentation made public in May 2024, and at length in the remedies opinion of 2 September 2025. Google's public ranking systems guide does not name it.
What are the key dates for RankEmbed?
- 2020-05-08: Google 'Logging & Ranking' presentation (UPX0219) lists RankEmbed among log-trained systems (source)
- 2023-10-18: Nayak names RankEmbed BERT as one of three main deep learning ranking models (source)
- 2024-08-05: Liability opinion lists RankEmbed among generalization systems (source)
- 2025-09-02: Remedies opinion requires sharing of RankEmbed training data (source)
Frequently asked questions about RankEmbed
What is the difference between RankEmbed and RankEmbedBERT?
The 2025 remedies opinion describes RankEmbedBERT as a later iteration of RankEmbed. Testimony does not publicly detail the architectural differences beyond the BERT name.
Does RankEmbed use click data?
Yes, according to testimony and the remedies opinion: it is trained on a sample of search logs plus human rater scores, far less data than Navboost uses.
Can I optimise directly for RankEmbed?
No tool exposes it. You can measure semantic similarity between queries and your passages with open embedding models as a rough proxy.