RankBrain

RankBrain is Google's machine learning ranking system, launched in 2015, that relates the words in a query to the concepts behind them so that a page can be matched to a query it does not literally contain. Google's own ranking systems guide defines it in a single line: "RankBrain is an AI system that helps us understand how words are related to concepts."

Google's post on how AI powers great search results fixes both the date and the significance: "When we launched RankBrain in 2015, it was the first deep learning system deployed in Search." Everything Google has since shipped into ranking — neural matching in 2018, BERT in 2019, MUM after it — sits downstream of that decision.

The mechanism, in the terms Google uses

Google illustrates RankBrain in its AI in Search post with the query "what's the title of the consumer at the highest level of a food chain". The literal reading of "consumer" is a shopper. The system recognises the query is about animals and about the concept of an apex predator, and matches pages that discuss that concept without necessarily using the words in the query.

That is the whole idea, and it is narrower than the folklore around it. RankBrain is a component that improves the mapping from query wording to concept. It is not a general-purpose scorer of page quality, and Google does not describe it as one.

Google is also explicit that these systems do not operate in sequence, one owning an outcome. Writing about BERT in the same post, Google says that "while BERT plays a major role in Search, it's never working alone — like all of our systems, BERT is part of an ensemble of systems that work together to share high-quality results." Attributing an individual ranking movement to RankBrain is therefore not something the published documentation supports.

Why it still matters

Because it is the point at which the industry's model of retrieval had to change, and the change has not been reversed since.

Before 2015, the mental model of matching was lexical: a page ranked for a phrase because the phrase was on it. After 2015, a page can rank for a query whose words it never uses, and can fail to rank for a query whose words it uses repeatedly. That is the same shift now happening again in semantic search and in embedding-based retrieval inside AI systems, where the unit of comparison is a vector rather than a string.

The practical implication is the same in both eras and it is a negative one. Repeating a target phrase does not make a page a better match for the concept behind it. Stating the concept clearly — naming the entity, the category, the relationship — does. See entity SEO for the structured version of that argument.

Where the concept fails

• Exact identifiers. Concept matching is good at paraphrase and weak at literals. A part number, an error code or a version string has no useful conceptual neighbourhood; it needs to appear verbatim in text.
• Attribution. Because the systems run together, no ranking change can honestly be traced to RankBrain from outside. Anyone claiming a specific RankBrain effect is inferring it.
• Retired-system confusion. Google maintains a separate retired list — Hummingbird (2013), Panda (2011, folded into core systems in 2015), Penguin (2012, integrated in 2016), and the helpful content system (2022, integrated into core ranking in 2024). RankBrain is not on it. Advice that treats it as historical is describing something Google still lists as active.
• Over-application. RankBrain is a query understanding component. It does not evaluate expertise, freshness, links or spam; Google lists separate systems for each of those.

How to act on it

1. Write the concept, not the string. If the page is about apex predators, the words "apex predator" should be in it, in a sentence that says so.
2. Keep literals literal. Model numbers, API names and versions appear as text, not as something a reader is expected to infer.
3. Stop treating keyword density as a lever. It has not been the matching mechanism since 2015, and in AI systems it is actively counterproductive.
4. Answer the concept-level question directly. Both a 2015 ranking system and a 2026 retrieval pipeline are looking for the passage that states the answer, not the page that mentions the topic.