# RankBrain

> Source: https://aiwiki.ai/wiki/rankbrain
> Updated: 2026-11-18
> Fact-checked: 2026-11-18
> Categories: Google, Information Retrieval, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "RankBrain." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/rankbrain
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**RankBrain** is a [machine learning](https://aiwiki.ai/wiki/machine_learning) system that [Google](https://aiwiki.ai/wiki/google) uses to help rank web pages in [Google Search](https://aiwiki.ai/wiki/google_search), and the first [deep learning](https://aiwiki.ai/wiki/deep_learning) system the company deployed in Search [2]. Google made it public in October 2015, when Bloomberg reported that the system turned words and phrases into vectors so it could make guesses about queries it had never seen before, and that it had become the third-most important of the hundreds of signals Google used to order results [1]. In its own descriptions, Google says RankBrain helps it understand how words relate to real-world concepts, so that it can return relevant pages even when they do not contain the exact words of a search [2][3]. Evidence from the U.S. government's antitrust case against Google, released in 2023 and later, added details the company had not published: RankBrain examines only the top 20 or 30 results for a query and may adjust their scores, and it is trained on click and query data and then fine-tuned on ratings from human evaluators [10][12][14].

## Announcement

RankBrain was first described publicly in an October 2015 Bloomberg article by Jack Clark, in which Greg Corrado, a senior research scientist at Google, described the system [1]. According to the article, RankBrain converts written language into mathematical objects called vectors that a computer can process. When it meets a word or phrase it does not recognize, it guesses which words or phrases might have a similar meaning and adjusts the results accordingly [1]. Bloomberg reported that the system was especially useful for the roughly 15 percent of daily queries that Google's systems had never seen before, and that a "very large fraction" of queries had been interpreted by RankBrain over the preceding few months [1].

Corrado contrasted RankBrain with Google's other ranking signals, which he described as based on discoveries made by information retrieval researchers with no learning involved; RankBrain, by contrast, learns from data [1]. The article said the system was one of "hundreds" of signals and that, within a few months of deployment, it had become the third-most important of them [1].

Bloomberg also reported two internal results. In one test, Google search engineers and RankBrain were each asked to guess which pages the search engine would rank highest: the engineers were right 70 percent of the time and RankBrain 80 percent [1]. In experiments, Google found that switching RankBrain off would hurt users about as much as forgetting to serve half the pages on Wikipedia [1]. According to the article, RankBrain came out of a yearlong effort that began with about five engineers, and Amit Singhal, then Google's senior vice president of search, approved its rollout [1].

## Expansion and leadership change

In early February 2016, Singhal announced that he was leaving Google after about 15 years, and [John Giannandrea](https://aiwiki.ai/wiki/john_giannandrea), who had led the company's artificial intelligence and machine learning work, took over Search [5]. In his coverage of the change, Danny Sullivan noted that Google had called RankBrain the third-most important signal involved in ranking pages [5].

In June 2016, Steven Levy reported in Backchannel that RankBrain had gone online in April 2015 and quoted [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean) as saying it was "involved in every query" and affected the rankings "probably not in every query but in a lot of queries" [4]. Levy wrote that RankBrain was then rated the third most useful of Google's hundreds of ranking signals [4]. Search Engine Roundtable contrasted this with the roughly 15 percent of queries described at launch [4].

Google's own dating of the system has varied. Bloomberg's October 2015 report described it as already in use for months [1], and Pandu Nayak, a Google vice president for Search, wrote in 2022 that Google launched RankBrain in 2015 [2]. In March 2019, however, Google's Search Liaison account described RankBrain as a system Google "began using in 2016" [8].

## Hardware

In May 2016, when Google announced its [Tensor Processing Unit](https://aiwiki.ai/wiki/tpu), a custom chip for machine learning, it named RankBrain among the applications the chips already powered, describing it as used to improve the relevancy of search results [9]. The TPU generation in service at that point is covered in [TPU Chip](https://aiwiki.ai/wiki/tpu_chip).

## How it works

Google has never published a technical paper on RankBrain, and much of what is known comes from company statements and from testimony and exhibits in the antitrust case.

### Google's public description

Nayak's February 2022 overview of AI in Search described RankBrain as a system that relates the words in a search to concepts, learned from seeing those words on many pages. It gave the example of a query asking for the title of the consumer at the top of a food chain, which RankBrain could connect to the concept of an apex predator [2]. Nayak said RankBrain is used to help decide the order of top results, and that it remained one of the major AI systems in Search alongside neural matching (introduced in 2018) and [BERT](https://aiwiki.ai/wiki/bert) (introduced in 2019) [2]. The same post said MUM, a multitask model Google had introduced the previous May, was not then used to rank results in the way RankBrain, neural matching and BERT were [2].

Google's Search Central documentation lists RankBrain in its guide to ranking systems, describing it as an AI system that helps Google understand how words are related to concepts, so that it can return relevant content that does not contain all the exact words used in a search [3]. As of November 2026 the guide, last updated in December 2025, still includes RankBrain [3].

### RankBrain and neural matching

In March 2019, Google's Search Liaison account distinguished RankBrain from neural matching. It summarized the difference as RankBrain helping Google "better relate pages to concepts" and neural matching helping it "better relate words to searches", and said there was nothing special that searchers or site owners needed to do for either system [8].

### Details from the antitrust case

The Department of Justice's search case against Google, *United States v. Google LLC*, produced the most specific public account of how RankBrain runs.

| Detail | Source |
|---|---|
| One of Google's three main deep learning ranking models, with DeepRank and RankEmbed BERT, all trained in part on click and query data | Nayak trial testimony, 2023-10-18 [10] |
| Looks at the top 20 or 30 documents for a query and may adjust their initial scores | Nayak trial testimony [10] |
| More expensive to run than some other ranking components, so it is not run on hundreds or thousands of results | Nayak trial testimony [10] |
| Trained on queries across all languages and locales Google operates in (Nayak: "I think so"), then fine-tuned on human rating (IS) data | Nayak trial testimony [10] |
| Cannot be trained on human rater data alone; retrained regularly with new data so it is not blind to new events | Nayak trial testimony [10] |
| Fine-tuned on IS rating data | Plaintiffs' rebuttal expert slides (Douglas Oard), 2023-11-15 [14] |
| Described with DeepRank as an LLM-based signal, unlike most hand-crafted signals | Notes of a 2025 call with Google engineer HJ Kim [16] |

At trial on October 18, 2023, Nayak agreed that RankBrain examines the top 20 or 30 documents and may adjust their scores, and that its cost is one reason Google waits until the candidate list is that short before running it [10]. Asked whether RankBrain had for years been trained on 13 months of click and query data, he answered that it initially started with the same amount as Navboost, Google's click-based signal [10]. Plaintiffs' counsel also read from a 2019 internal Google review of DeepRank that asked whether DeepRank would replace RankBrain and answered "No, complementary strengths". Nayak said that was not an unreasonable position, but added that the two were becoming less complementary over time and that DeepRank was taking on more and more of that capability [10]. Search Engine Land's summary of the testimony described DeepRank as BERT applied to ranking [11].

An internal Google presentation entered as a plaintiffs' exhibit, titled "Logging & Ranking", said that all of Google's major machine learning ranking systems, naming RankBrain, RankEmbed and DeepRank, relied on search logs [13]. In his August 2024 liability ruling, Judge Amit Mehta grouped RankBrain with DeepRank, RankEmbed, RankBERT and MUM as newer "generalization" signals that rely less on user data than older signals such as Navboost, but found that they were designed with user data and continue to be trained on it, in smaller volumes [12]. The ruling also described a data reduction experiment by Google's expert Edward Fox that retrained several signals, RankBrain among them, on an estimate of Bing's volume of user data; the court gave the experiment little weight [12].

In notes of a January 2025 call with Google's lawyers and expert, entered as an exhibit in the case, Nayak described Google's shift toward machine learning and deep learning signals, naming RankEmbed, RankBrain and DeepRank, while saying that Google avoids simply predicting clicks because clicks are easy to manipulate [15].

## Reception and commentary

Google staff have described RankBrain in more candid terms than its launch coverage. At SMX West in March 2016, Paul Haahr, a Google ranking engineer, said during a question-and-answer session that Google understood how RankBrain works but not entirely what it is doing, according to attendees quoted by Search Engine Roundtable [6]. In a February 2019 Reddit question-and-answer session, Gary Illyes of Google called RankBrain a "PR-sexy machine learning ranking component" that uses historical search data to predict what a user would most likely click on for a previously unseen query, and said it relied on data about what happened on the results page itself rather than on the landing page [7].

Search industry writers questioned whether RankBrain should be called a ranking signal at all. Sullivan argued in 2016 that it appeared to work mainly by interpreting what people type and relating it to broader, similar queries, rather than by judging which pages are best [5], and Barry Schwartz of Search Engine Roundtable wrote that he still did not see how it was a ranking signal rather than a query interpreter or modifier [4]. Google's Search Liaison account said in 2019 that there was nothing special searchers or site owners needed to do for RankBrain or neural matching [8], and Search Engine Roundtable reported that RankBrain cannot be optimized for [4].

## Significance

When it was announced, Corrado contrasted RankBrain, which learns, with ranking signals that did not [1], and Google now describes it as its first deep learning system in Search [2]. It represents language as vectors [1], the approach behind [word embeddings](https://aiwiki.ai/wiki/word_embedding), and applies a [neural network](https://aiwiki.ai/wiki/neural_network) to the [information retrieval](https://aiwiki.ai/wiki/information_retrieval) problem of matching queries to pages by meaning rather than exact words, which is also the aim of [semantic search](https://aiwiki.ai/wiki/semantic_search) [1][2]. Google has since added neural matching and BERT to its [search engine](https://aiwiki.ai/wiki/search_engine) [2], and Nayak testified in 2023 that DeepRank, its BERT-based ranker, was taking on more and more of RankBrain's capability as the two became less complementary [10][11].

## See also

- [Google Search](https://aiwiki.ai/wiki/google_search)
- [BERT](https://aiwiki.ai/wiki/bert)
- [Search engine optimization](https://aiwiki.ai/wiki/seo)
- [Word2vec](https://aiwiki.ai/wiki/word2vec)
- [Google Brain](https://aiwiki.ai/wiki/google_brain)

## References

1. Clark, Jack (Bloomberg). "Google Turns Web Search Over to Artificial Intelligence Machines". Claims Journal (reprinted from Bloomberg), October 27, 2015. https://claimsjournal.com/news/national/2015/10/27/266615.htm
2. Nayak, Pandu. "How AI powers great search results". Google The Keyword, February 3, 2022. https://blog.google/products/search/how-ai-powers-great-search-results/
3. Google Search Central. "A guide to Google Search ranking systems". Google for Developers, last updated December 10, 2025. https://developers.google.com/search/docs/appearance/ranking-systems-guide
4. Schwartz, Barry. "RankBrain Used In All Google Searches & Impacts Search Results A Lot". Search Engine Roundtable, June 27, 2016. https://www.seroundtable.com/rankbrain-used-for-all-google-queries-22300.html
5. Sullivan, Danny. "What's Next, As Google's Head Of Search Leaves & Its Machine Learning Chief Takes Over?". MarTech (Search Engine Land), February 3, 2016. https://martech.org/google-machines-running-the-search
6. Schwartz, Barry. "Google's Paul Haahr: We Don't Fully Understand RankBrain". Search Engine Roundtable, March 8, 2016. https://www.seroundtable.com/google-dont-understand-rankbrain-21744.html
7. Schwartz, Barry. "Google Gary Illyes: RankBrain Is PR-Sexy Machine Learning". Search Engine Roundtable, February 8, 2019. https://www.seroundtable.com/google-rankbrain-is-pr-sexy-machine-learning-27082.html
8. Schwartz, Barry. "Google On Neural Matching vs RankBrain". Search Engine Roundtable, March 22, 2019. https://www.seroundtable.com/google-explains-neural-matching-vs-rankbrain-27300.html
9. Jouppi, Norm. "Google supercharges machine learning tasks with TPU custom chip". Google Cloud Blog, May 18, 2016. https://cloud.google.com/blog/products/ai-machine-learning/google-supercharges-machine-learning-tasks-with-custom-chip
10. U.S. District Court for the District of Columbia. "Transcript of Bench Trial, Day 24, Afternoon Session" (testimony of Pandu Nayak), United States v. Google LLC, No. 1:20-cv-3010, October 18, 2023. https://thecapitolforum.com/wp-content/uploads/2023/10/101823-USA-v-Google-PM.pdf
11. Goodwin, Danny. "How Google Search and ranking works, according to Google's Pandu Nayak". Search Engine Land, December 5, 2023. https://searchengineland.com/how-google-search-ranking-works-pandu-nayak-435395
12. Mehta, Amit P. "Memorandum Opinion", United States v. Google LLC, No. 1:20-cv-03010-APM, Document 1033, U.S. District Court for the District of Columbia, August 5, 2024. https://storage.courtlistener.com/recap/gov.uscourts.dcd.223205/gov.uscourts.dcd.223205.1033.0_2.pdf
13. U.S. Department of Justice. "Trial Exhibit UPX0219: Logging & Ranking", United States v. Google LLC. https://www.justice.gov/d9/2023-11/417518.pdf
14. Oard, Douglas W. "Rebuttal Testimony of Professor Douglas W. Oard" (Trial Exhibit UPXD105), United States v. Google LLC, November 15, 2023. https://www.justice.gov/d9/2023-11/418021.pdf
15. U.S. Department of Justice. "Trial Exhibit PXR0357: January 31, 2025 Call with Google Engineer Pandu Nayak", United States v. Google LLC. https://www.justice.gov/atr/media/1398866/dl
16. U.S. Department of Justice. "Trial Exhibit PXR0356: February 18, 2025 Call with Google Engineer HJ Kim", United States v. Google LLC. https://www.justice.gov/atr/media/1398871/dl

