What Is Machine Learning and How Does Google Use It?
RankBrain, BERT and MUM are what people search for and the least useful way to understand this subject. They are components inside an assessment nobody outside Google sees the whole of. None of them is something a site is optimised for. Everything below is restricted to what was actually published, with dates.
You Cannot Optimise For Any Of These
None of these systems is a filter to satisfy. They are components of an assessment, they operate on queries and language rather than on sites. There is no setting, structure or technique that addresses any of them individually.
Google has said so directly. About BERT, in October 2019.
Danny Sullivan stated at the time that BERT does not assign values to pages and that it is a way to better understand language. That distinction is the whole of this block. It also comes from Google rather than from us.
Why the misunderstanding is so durable. Everything else in search had a lever.
An update to how links are assessed produced link work. A speed signal produced speed work. Businesses reasonably expect a named system to produce corresponding work. The industry supplies the expectation whether or not the work exists.
What is actually being described. Interpretation, not judgement.
These systems are largely concerned with establishing what a query means and what language conveys. They sit before the assessment of whether your page is any good rather than inside it.
What that leaves you. One thing, repeatedly.
Every one of them rewards the same response, which is covered in block seven and is considerably less exciting than a named system suggests.
What Machine Learning Is Doing Here
A machine learning system improves at a task from examples rather than from rules somebody wrote. That is the entire concept. Search is an unusually good application of it.
Why rules fail at this. Language will not hold still.
Nobody can write rules covering how people phrase questions, because the phrasings are effectively infinite, they change constantly and they differ by region, age and trade. Any rule set is out of date before it is finished.
What examples provide instead. Patterns nobody articulated.
Given enough instances of queries and what satisfied them, a system can identify regularities no person would have thought to write down. That is genuinely different from a very large rule book.
The consequence people find uncomfortable. Nobody can fully explain it.
A system that learned from examples does not carry a readable list of reasons. Google engineers can describe what a system was built to do without being able to state why it treated one particular page as it did.
Why that matters to how you read this industry. Confident explanations are suspect.
Anybody explaining precisely why a system did something specific is inferring. That is fine as inference and misleading as fact. The distinction is the reason this page states dates and sources.
RankBrain
RankBrain was first described publicly in a Bloomberg article by Jack Clark on 26 October 2015, based on an interview with Greg Corrado, then a senior research scientist at Google. It was described as a machine learning system for processing queries.
What it was said to address. Queries never seen before.
A meaningful share of searches every day have never been made in that exact form. A system that can only recognise queries it has encountered handles none of them. That gap was the stated problem.
The claim that outlived everything else. A ranking of factors.
In that same 2015 interview, Corrado described RankBrain as the third most important ranking signal after content and links. Google has not ranked its top factors that plainly since, which is precisely why the statement survives in industry folklore a decade later.
How to treat that claim now. As a dated statement.
It was accurate as a description of a position in 2015. Repeating it today as a current fact about how search works is presenting an eleven year old interview as present tense, which happens constantly.
Why the block is short. Little else was published.
Beyond those points, most of what circulates about RankBrain is inference. We are not going to fill that space, because filling it is the failure this page exists to avoid. The system was also reconfirmed as playing a role during United States Department of Justice proceedings, which is a rare instance of sworn testimony rather than commentary. That is worth more than a decade of articles restating the same 2015 interview.
BERT
BERT was applied to search from 25 October 2019. Google's announcement described it as the biggest change in five years and said it affected about one in ten queries in English in the United States. It extended to over seventy languages from 9 December 2019.
What it addressed. Relationships between words.
Rather than treating a query as a collection of terms, it considers how the words relate to each other, in both directions rather than reading left to right. That sounds academic until you see what it changes.
Google's own launch example. One preposition.
A query about a Brazilian traveller to the United States needing a visa. The word "to" establishes the direction of travel. A system disregarding small words answers the opposite question. That is why prepositions suddenly mattered.
Why it was a bigger deal than it sounded. Small words are everywhere.
The words that carry direction, negation, possession and condition are precisely the ones earlier systems treated as noise. Handling them correctly changes the meaning of an enormous number of ordinary questions.
The boundary with our other material. Deliberate, not an oversight.
BERT also appears in our algorithm updates material, treated as a dated event in a timeline. Here it is treated as a language understanding system and what it changed. That duplication is intentional and neither version should be removed to tidy it.
MUM
MUM, the Multitask Unified Model, was announced by Pandu Nayak on the Google blog on 18 May 2021 at Google I/O. Google stated it was roughly one thousand times more powerful than BERT, trained across seventy-five languages and multimodal, meaning it works across formats rather than text alone.
What was promised. Complex, multi-step questions.
The capability highlighted was answering questions requiring several pieces of reasoning, of the kind a person would otherwise break into a sequence of separate searches.
What actually appeared. Considerably less than the announcement.
Its visible applications have been narrow features rather than general ranking. The first applied use was vaccine information from June 2021. Subsequent uses have been specific product features and particular verticals rather than the assessment of pages generally.
Why we are saying that plainly. Almost nobody does.
An enormous quantity of writing treats MUM as a force reshaping how pages are ranked. It is a textbook case of a loud announcement without a corresponding consequence for ordinary search work. A page repeating the announcement rather than checking what followed is not worth reading.
What that teaches beyond MUM. Announcements are not deployments.
A capability being demonstrated is not the same as that capability affecting your results. The gap between the two is where most confident industry commentary lives.
Newer Systems Keep Arriving
Named systems have arrived steadily for a decade and will continue to. Attempting an exhaustive current list is a losing exercise, so this page deliberately does not attempt one.
Why the list approach fails. Two reasons.
Systems are announced without deployment detail, deployed without announcement, then sometimes named retrospectively. And any page claiming to be current becomes wrong within months without appearing to change.
What has stayed constant. The pattern.
Each arrival has moved understanding closer to how a person reads. Each has been followed by industry writing describing how to optimise for it. The second half of that pattern has never once produced a durable technique.
What we do instead of listing. Describe the direction.
The direction is stable enough to plan around even when the named components are not. That is a more useful thing to give a business than a list that ages.
When this page is reviewed. Quarterly.
Checked on 15 August 2026 and reviewed quarterly thereafter, plus after any significant announcement. The owner is Andrew Odgers, named rather than assigned to the business, because a review owned by nobody in particular stops happening during a busy quarter.
What All Of Them Have In Common
Every one of these systems moved understanding closer to how a person reads. Not one moved it towards a mechanical property a site could be adjusted to satisfy. That has held across a decade.
What the direction has been. Consistently one way.
From matching terms towards understanding questions. From treating queries as bags of words towards reading them as sentences. From handling what had been seen before towards handling what had not. Every step has been in the same direction.
What follows from that. The answer has not changed.
Across a decade of these systems, the practical response has been identical: write as somebody who genuinely knows the subject would write for somebody who needs to understand it. That is unglamorous and it has outlasted every technique invented alongside it.
Why it keeps being resisted. It is not actionable in the way people want.
It cannot be delegated to a specification, checked against a target or completed on a schedule. Businesses want a lever and this is a standard, which is a harder thing to buy.
What it connects to. The rest of this cluster.
The mechanism behind it is in what is semantic SEO and the strategic consequence in why topical authority outperforms keywords.
AI Tools Support A Writer Rather Than Replace One
This is the position we hold everywhere on this site. It belongs on this page in particular because of the connection between the two halves of the subject, which is rarely made.
The connection. Both concern reading like a person.
Systems that assess content the way a reader would are not fooled by content produced without a reader in mind. The same decade of development that made language understanding better made unread content easier to recognise as unread.
What responsible use looks like. Judgement stays with a person.
Structuring an argument, drafting something a person rewrites, tidying prose, getting past a blank page. In each case somebody decides what is true and what is worth saying.
What irresponsible use looks like. Publishing output.
Generating pages and putting them live without anybody deciding they should exist. The failure is the absence of a person rather than the presence of a tool.
The specific trap on this subject. Machines assessing machines.
There is a tempting symmetry in the idea that content produced by a language model would suit systems built on language models. Nothing supports it, the assessment is still ultimately about whether a reader was served. Our blogging material sets out the position in full.
What This Does Not Mean
A page arguing that the answer is always good writing invites a conclusion that nothing else matters. That is wrong. It is also a conclusion businesses reach with expensive results.
Technical work did not stop mattering. The obvious point, frequently missed.
A page nothing can reach is not assessed by any of this. No amount of language understanding evaluates a page that was never fetched. None of these systems compensates for a site that cannot be crawled.
Which makes the order clear. Reachable first.
Being findable is a precondition rather than a competing priority. Businesses that read about semantic understanding and stop maintaining their site have removed the floor from underneath the argument.
Structure still matters too. For the same reason.
How a site is organised affects what gets reached and what a section is understood to be about. That is a separate contribution rather than an alternative to good writing.
The accurate summary. Language understanding improved. The prerequisites did not go away.
Everything that made a site legible to a machine still applies. It now determines whether any of the interesting assessment happens at all. The full series is on the advanced SEO guide.
There is no
lever. There
is a standard.
A decade of named systems has produced one consistent answer. It cannot be delegated to a specification. What we do instead is cover a subject completely enough that the assessment has something to work with. We will tell you plainly if your subject is too broad to cover properly.
What the national tier covers:
The national tier is £1,550 a month. It suits businesses competing beyond one town. We will say so if a lower tier would serve you better.
Every guide.
One practice.
Semantic search, machine learning, topical authority, programmatic production, rendering, log files, crawl allocation, information architecture, content pruning and how strategy changes with scale.
