RankBrain Explained: What It Does, What It Doesn’t, and What the Trial Revealed

RankBrain gets blamed for a lot of drops it could not have caused. It is one of the best-known names in search, and one of the least understood.

Knowing what it actually does, and where in the ranking process it runs, rules it in or out of a diagnosis in minutes. That is more useful than any amount of “optimising for RankBrain”.

What RankBrain does

Google describes RankBrain as an AI system that helps it understand how words relate to concepts. In practice that means Google can return a relevant page even when the page does not use all the words in the search, because it recognises that the page is about the same idea.

It was built for a specific problem. A meaningful share of searches every day are ones Google has never seen before, often long, oddly phrased or full of everyday language. Matching those by exact words alone produced poor results.

A simple example

Imagine someone searches for “shoes for kids that play football on grass”. A page titled “Children’s football boots for firm ground” may not contain the words shoes, kids or grass.

A system that only matched words would struggle. A system that understands the concept can see these mean the same thing and treat that page as a strong answer. That is the gap RankBrain was designed to close.

A short timeline

• Early 2015: RankBrain is already running on a share of searches.
• October 2015: Google reveals it publicly, describing it as a key signal and explaining it helps with the searches it has never seen before.
• 2016: Google reports that RankBrain is involved in processing every search.
• November 2022: it appears in Google’s official guide to ranking systems.
• December 2025: the latest version of that guide still lists it as an active system.

What the 2023 antitrust trial revealed

On 18 October 2023, Google’s Vice President of Search, Pandu Nayak, was questioned under oath in the United States antitrust case against Google. Several answers concerned RankBrain directly.

• It looks at the top 20 or 30 documents for a search and may adjust their initial scores. He confirmed this when asked.
• It is expensive to run, certainly more so than some of Google’s other ranking components, which is why it is applied to a short list.
• It is trained partly on click and query data, as are Google’s other deep learning ranking models.
• It works alongside newer models such as DeepRank, which is based on BERT, rather than being replaced by them.

The first point changes how you should think about it. RankBrain refines the order among pages that are already strong candidates. It does not pull a page from position 80 onto page one.

Where RankBrain fits in the ranking process

The testimony also sketched the broader order of events. Simplified, a search passes through three broad stages:

1. Retrieval. Google narrows a huge index down to a set of candidate pages that could plausibly answer the search.
2. Scoring. Those candidates are scored using many signals, including links, relevance and quality.
3. Refinement. More expensive models, RankBrain among them, adjust the order of the top 20 or 30.

This is a simplification, but a useful one. It explains why foundations come first: a page has to be retrieved and scored highly before any refinement can help it. Quality and relevance get you onto the short list. Systems like RankBrain decide the finer order within it.

How it differs from Google’s other language systems

System	Introduced	What it does
RankBrain	2015	Relates words to concepts, especially for unfamiliar searches
Neural matching	2018	Matches the concepts in a search to the concepts in pages
BERT	2019	Understands how word order and small words change meaning
MUM	2021	Understands and generates language, but Google says it is not used for general ranking

The simplest way to remember it: RankBrain and neural matching connect ideas, while BERT reads context. All of them are about understanding what a search means, not about judging whether your site deserves to rank. The wider picture is in Google’s ranking systems explained.

Can you optimise for RankBrain?

Not directly, and Google has said as much. There is no setting, tag or keyword density that RankBrain rewards.

• Write the way people search and speak, rather than repeating a phrase you want to rank for.
• Cover the concept fully, including the related questions someone would naturally ask next.
• Use the words your audience uses, not only the industry term.
• Make the page clearly about one thing, so its concept is easy to recognise.

Each of these is simply good writing. That is the point: RankBrain exists to reward pages that answer the idea behind a search, whatever words they use. The older systems it built on are covered in Hummingbird, RankBrain and BERT.

RankBrain and traffic drops

RankBrain is a language system, not a quality system, so it cannot penalise a site. When it plays a part in a change, what shifts is which searches your pages are considered relevant for.

• A drop from page one to page five is not RankBrain. It works within the top 20 or 30 results, so a fall far beyond that range has another cause.
• A shift in the queries sending you traffic can be. Long, conversational searches may start finding your pages while an exact-match phrase sends fewer visitors.
• Rankings that move a few places among the top results are consistent with re-ordering, as are many other everyday causes.

Check the query mix before assuming a loss. Positions that shift while clicks hold steady are covered in rankings dropped but traffic is the same.

Spotting a query mix change in Search Console

1. Open the Performance report and compare the four weeks after a change with the four weeks before.
2. Switch to the Queries tab and sort by the difference in clicks.
3. Look for new queries near the top of the list that barely existed before, often longer and more conversational.
4. Compare them with the queries that fell. If one page lost an exact phrase but gained related searches, the page was re-matched, not demoted.
5. Repeat the check page by page for your most important URLs.

Total clicks often tell the real story. A page that swapped one high-volume phrase for dozens of smaller related ones can hold its traffic, or even grow it, while a rank tracker following the single phrase reports a fall.

Common myths about RankBrain

• “RankBrain penalises keyword stuffing.” Stuffing is dealt with by spam and quality systems. RankBrain simply makes it pointless.
• “RankBrain is the third most important ranking factor, so focus on it.” That description dates from 2015 and was never a measurable target.
• “RankBrain uses my bounce rate.” Its training includes click and query data from search, not your analytics. The difference is explained in do clicks affect Google rankings.

How neural matching differs from RankBrain, and why MUM is not used for general ranking, is covered in neural matching and MUM.