DeepRank is Google's name for BERT when it is applied to ranking, according to Pandu Nayak's sworn testimony in the US antitrust trial. It is a transformer model trained partly on user data that re-scores a small set of top results for language understanding. Google has publicly announced BERT in Search but, as far as we can verify, not the DeepRank name. For sites, it means the meaning of every word in a query, including small words, can decide whether a page is the best answer.
What is DeepRank?
DeepRank is the internal name for Google's use of BERT as a ranking model. In the 18 October 2023 trial session, DOJ counsel asked Nayak: "When BERT is used for ranking, that's DeepRank?" He answered: "That's correct." He also agreed DeepRank is trained on user data.
A Google document read into the record at the same session describes DeepRank as "the latest edition" to a family that began with RankBrain and then RankEmbed.
DeepRank not only gives significant relevance gains, but also ties ranking more tightly to the broader field of language understanding.
Google document (UPX2029) read into the record, Transcript of bench trial, Day 24 afternoon session, 18 October 2023, United States v. Google LLC
How does DeepRank work?
The testimony places DeepRank at the end of the ranking pipeline, working on a small candidate set after traditional signals such as Navboost and topicality have narrowed tens of thousands of documents to a few hundred. Nayak explained that DeepRank uses transformers, which are more expensive to train than the feedforward networks behind RankBrain.
A 2019 Google review admitted as UPX860 says DeepRank "understands language and has common-sense" and answers the question whether it replaces RankBrain with "No, complementary strengths." Nayak added that over time the two were "becoming less complementary" as DeepRank took on more of that capability (transcript).
The 2025 HJ Kim call notes mention an eDeepRank that "tries to take LLM-based signals and decompose them into components to make them more transparent", and the PXR0357 Nayak notes say BERT-based DeepRank signals could be decomposed into signals resembling traditional ones.
Does DeepRank replace RankBrain? No, complementary strengths.
Google 2019 review (UPX860) read into the record, Transcript of bench trial, Day 24 afternoon session, 18 October 2023, United States v. Google LLC
What evidence is there that Google uses DeepRank?
The label is Trial or leak evidence, specifically DOJ trial testimony and exhibits, plus remedies phase call notes. In October 2023 Nayak agreed with DOJ counsel that RankBrain, DeepRank and RankEmbed BERT are the main three deep learning models Google uses in ranking. Judge Mehta's August 2024 opinion lists DeepRank among newer signals that rely less on user data. A 2020 Google presentation, UPX0219, lists it among machine learning systems that rely on logs.
Two limits. First, we found no leaked Content Warehouse model named DeepRank in the v0.4.0 API reference index, so the leak adds nothing here. Second, Google's public BERT announcement is covered on our BERT page; the DeepRank name comes only from the litigation record.
eDeepRank is an LLM system that uses BERT, transformers.
Notes of call with HJ Kim, PXR0356, notes of 18 February 2025 call with Google engineer HJ Kim, US Department of Justice
What does DeepRank mean for your site?
A transformer reads the whole query in context, so prepositions, negations and qualifiers matter. A page about "flights to Madrid" is a poor answer to "flights from Madrid", and a page about dog food is a poor answer to "can dogs eat grapes without harm" unless it addresses that relationship.
The falsifiable test: for queries containing relation words (to, from, without, for, versus, near), check whether the ranking page's answer states the same relationship. Where it answers a neighbouring question, expect weaker positions than a page that matches the relation.
- Answer the exact relationship the query asks about.
- Keep qualifiers such as audience, location and time explicit in the answer.
- Write naturally; a model trained on language gains little from keyword repetition.
Which Laurelin audit checks test for DeepRank?
DeepRank cannot be observed from a site. The new check, answer ignores the query's qualifying words, is a manual proxy for the language understanding DeepRank is described as adding.
Related checks: content not matching search intent is the broad version of the same failure; answer buried below the lead delays the matching answer; no clear answer to the page's question leaves nothing for the model to match; low readability for the audience and poor auto-translated content degrade the language a transformer reads.
What are the key dates for DeepRank?
- 2020-05-08: Google 'Logging & Ranking' presentation (UPX0219) names DeepRank among log-trained systems (source)
- 2023-10-18: Nayak testifies that DeepRank is BERT used for ranking (source)
- 2024-08-05: Liability opinion lists DeepRank among newer generalization signals (source)
- 2025-02-18: HJ Kim call notes describe eDeepRank (source)
Frequently asked questions about DeepRank
Is DeepRank the same as BERT?
Nayak testified that DeepRank is BERT when BERT is used for ranking. It is an application of BERT, not a separate public model.
Did DeepRank replace RankBrain?
A 2019 Google review said no, citing complementary strengths. Nayak testified in 2023 that DeepRank was taking on more of that capability over time.
Has Google publicly named DeepRank?
We could not verify any public Google statement using the name. It comes from trial testimony and exhibits.