Hummingbird, RankBrain and BERT: Why You Cannot Be Penalised by Them

Two families of ranking system

Almost every diagnostic mistake in this area comes from treating these as one thing. They are not.

	Understanding systems	Quality systems
Examples	Hummingbird, RankBrain, BERT, MUM	Core ranking, helpful content, spam, reviews
What they decide	What the query means and which pages are relevant	Whether your content deserves to rank
Measurement (the signature)	Query mix changes, some terms lost and others gained	Broad loss across a page group or the whole site
Can you be demoted?	No. You are matched differently	Yes
How to respond	Follow the queries you now match	Fix the quality problem, wait for reassessment
Announced?	Rarely, and updated continuously	Core and spam updates are confirmed

Read the fourth row carefully, because it is where the money gets wasted. There is no BERT penalty and no RankBrain penalty. If your traffic changed shape after one of these systems shipped, what changed is which searches send you visitors.

Hummingbird, 2013

Announced in September 2013 and already running for some weeks by then, Hummingbird was a rewrite of the core algorithm rather than a filter bolted onto it. Panda and Penguin adjusted an existing engine. Hummingbird replaced it.

The shift was from matching keywords to interpreting meaning. Google began treating a query as a question with intent behind it rather than a bag of words to find on a page. This is where the idea of entities rather than strings entered ranking, so that a page about the same concept could rank without repeating the exact phrase.

Its legacy is the reason keyword density stopped working. Writing the same phrase fifteen times stopped helping around this point, because the system was no longer counting the phrase.

RankBrain, 2015

Google’s first machine learning component in ranking, built for a specific problem: a meaningful share of daily searches had never been seen before. With no historical data on a query, keyword matching has nothing to work from.

RankBrain interprets unfamiliar queries by relating them to similar ones it has seen. Google described it at the time as one of the most important ranking signals, and later took care to downplay the idea that it was a single dominant factor. Both statements are consistent: it matters, and it is one system among many.

Practically, RankBrain is why long-tail and conversational queries began finding pages that never targeted those exact words. It expanded the query set a good page could serve, which is a gain rather than a risk for most sites.

BERT, 2019

BERT reads a query in both directions at once, so a word is interpreted using everything around it rather than the words before it. That sounds abstract until you see what it fixed.

Before BERT, small function words were largely ignored. Prepositions in particular carry the whole meaning of a search: to and from, with and without, for and against. A query about travelling to a country and one about travelling from it were treated as near-identical. BERT made those distinct.

At launch it affected around one in ten English searches in the US and was extended to more languages afterwards. The pages that gained were ones answering the precise version of a question. The pages that lost were ones ranking for a query they only approximately matched, which had been getting traffic they never really earned.

MUM and what came after

Announced in 2021 and far more capable than BERT, MUM works across languages and formats, so information learned in one language can inform results in another. Google has deployed it selectively rather than as a broad ranking change.

The same lineage now underpins AI Overviews, which is where language understanding stopped being invisible. Earlier systems changed which results you saw. This one generates an answer above them, which is why it produces a traffic problem the previous generation never did. That is covered in impressions stable, clicks falling.

What a query-matching change looks like

Distinguishing this from a quality demotion is straightforward once you know what to compare, and it is the whole point of understanding these systems.

• Compare your query list before and after, not just your totals. Why: A quality demotion loses the same queries at lower positions. A matching change loses some queries entirely and gains others. Different shapes, same headline number.
• Check whether impressions moved with clicks. Why: A matching change usually shifts impressions too, because you stopped appearing for those searches at all rather than appearing lower.
• Look at what you gained. Why: Demotions do not come with gains. If new queries appeared as others vanished, the system reassigned you rather than judged you.
• Check whether the lost queries were ever a good fit. Why: Traffic lost because you were only approximately relevant rarely converted anyway. Losing it looks worse in a chart than it is in revenue.

The drop analyzer separates a clicks-only loss from a genuine ranking loss, and the page scope analyzer shows whether the loss is site-wide, which a matching change almost never is.

What this means for how you write

• Answer the precise question, not the topic. Why: BERT made small qualifiers meaningful. A page covering a subject broadly loses to one answering the exact question with its conditions and exceptions.
• Stop writing for keyword variants. Why: Separate pages for near-identical phrasings are a Hummingbird-era mistake that has since become an active liability, because near-duplicate pages are what recent core updates narrow hardest.
• Write the way people ask. Why: These systems were built for conversational language. Natural phrasing is now a match rather than a compromise.
• Cover the whole entity, not the phrase. Why: Related concepts, common follow-up questions and the parts people get wrong. That is what completeness means to a system reasoning about meaning.

None of this is optimisation for BERT, which is not a thing that can be done. It is writing that survives a system which understands what it reads.

Frequently asked questions

Can my site be penalised by BERT or RankBrain?

No. They are language understanding systems, not quality judgements. They change which queries your pages match. If traffic changed after one shipped, you were matched differently rather than demoted.

Can you optimise for RankBrain or BERT?

Not directly, and Google has said so repeatedly. What helps is answering precise questions in natural language and covering a topic completely, which is what these systems are built to recognise.

Are Hummingbird, RankBrain and BERT still running?

Yes. Google lists RankBrain and BERT among its active ranking systems, and Hummingbird’s architecture underlies everything since. None were switched off, and the same lineage now powers AI Overviews.

What is the difference between Hummingbird and Panda?

Hummingbird rewrote how Google interprets queries. Panda judged content quality and demoted sites that failed. One decides relevance, the other decides worth. The Panda lineage is covered in the Google Panda update.

Why did my rankings change without a confirmed update?

Understanding systems are updated continuously and almost never announced. A shift in which queries you match can happen at any time with nothing on any update list. Compare your query mix before and after rather than only your traffic totals.

Does keyword density still matter?

No, and it has not since Hummingbird. Repeating a phrase does not make a page more relevant to a system reasoning about meaning, and heavy repetition now reads as a quality problem instead.