SEO Glossary · Algorithm

BERT

Google still lists BERT among its active ranking systems in the documentation in force in December 2025, so filing it under 2019 news is a mistake. What it means operationally: query and page are matched on meaning, not tokens. A page that answers the intent outranks one that repeats the keyword, whatever its backlink profile looks like.

Key takeaways The essentials in 30 seconds
  • BERT is not a historical update, it is a named ranking system Google still documents as active in the version of its ranking systems page in force in December 2025, alongside RankBrain, neural matching and link analysis.
  • You cannot optimise for BERT. It has no dial, no threshold, no score you can inspect. What you can do is remove the ambiguity it has to resolve: explicit entities, explicit qualifiers, one clear intent per URL.
  • DOJ trial documentation summarised in September 2025 describes RankEmbed BERT as trained on 70 days of search logs plus quality rater scores, which is a very different animal from a link-graph signal.
  • Semantic relevance now filters what links can do. An off-topic link from a strong domain moves less than it did five years ago because the receiving page has to pass topical matching first.
  • The commercial consequence is measurable on informational queries: an Ahrefs analysis of 300,000 keywords on December 2025 data found the top organic result loses roughly 58% of its CTR when an AI Overview appears.
  • Anchor text is still parsed in context. Over-optimised exact-match anchors read as a pattern, not as a relevance signal, and they were already a liability before BERT existed.
3 questions to test your knowledge Read first, the quiz is waiting at the bottom.
Four numbered steps tracing the release of BERT in October 2018, its rollout in Search in October 2019, its extension to more than 70 languages in December 2019 and its status as an active system in the December 2025 documentation.
The four public milestones that turn BERT from a research paper into a permanent ranking component.

What BERT really is, beyond the acronym

BERT stands for Bidirectional Encoder Representations from Transformers. Google open-sourced the model in 2018 and shipped it into Search in 2019, starting with English before extending it to other languages. The acronym is the least interesting part. The mechanism is what changed the job: BERT reads a token conditioned on everything to its left and everything to its right at the same time, so « bank » in « river bank » and « bank » in « bank transfer » stop being the same word, and the preposition in « 2019 brazil traveler to usa need a visa » stops being a stop word Google can safely throw away.

For anyone who was writing pages before 2019, that is the whole shift. Pre-BERT ranking rewarded lexical overlap plus authority. You matched the string, you pointed links at it, you ranked. Post-BERT, the string matters far less than whether the passage on the page resolves the question the query encodes. Two pages using entirely different vocabulary can compete for the same query, and a page stuffed with the exact query terms can lose to one that never uses them in that order.

The honest framing for a 2026 audit: BERT is not a ranking factor you can influence directly, it is the interpretation layer sitting between the query and everything else Google scores. You do not optimise for it. You either give it an unambiguous document to parse or you make it guess, and when it guesses you inherit whatever it guessed.

List of five operational takeaways for writing: function words carry meaning, polysemy is resolved by neighbouring words, keyword density no longer produces anything, the model understands but does not write, and intent matters more than line-by-line fixes.
Five concrete consequences of how BERT works for the way you write a page.

Where BERT sits in Google's stack in 2026

The most common error in client decks is treating BERT as legacy, something superseded by MUM, then by Gemini. Google's own ranking systems documentation, in the version in force in December 2025, still lists BERT as an active named system next to RankBrain, neural matching, passage ranking, PageRank and link analysis, and spam detection. Systems Google actually retires get moved to a separate list, which is what happened to the Helpful Content system when it was folded into core ranking on 5 March 2024. BERT was not moved. It is still there.

Technical write-ups based on the 2024 Google API leak place BERT in the pipeline as one of the models that adjusts an initial score before final re-ranking, with Ascorer doing primary scoring and Twiddlers doing the last pass. More useful, because it is sourced from sworn material: DOJ trial documentation summarised in September 2025 by Marie Haynes describes RankEmbed BERT as a deep learning model trained on 70 days of search logs plus quality rater scores, producing an additional ranking signal that reflects user preference and E-E-A-T rather than link metrics. The same analysis notes that PageRank is now nowhere near as important as the information stored about the page itself. Read that sentence twice if you sell links for a living.

The other 2026 context is the answer layer. Gemini 3 Flash became the standard model for AI Mode worldwide between December 2025 and January 2026, and AI Overviews were estimated to reach around 1.5 billion monthly users across more than 200 countries by late 2025. Those answers sit on top of a retrieval and ranking pipeline that still uses BERT-family semantics to decide what is relevant in the first place. The generation model changed. The matching layer underneath did not get thrown away. When a broad core update rolls out and a site loses positions on informational queries, the diagnosis is almost never « BERT changed », it is « the semantic fit between our pages and those queries was always thin and something else stopped compensating for it ».

How BERT works: attention, masking, context

Three ideas carry the whole model. Self-attention lets every token compute a weighted relationship with every other token in the sequence, so the representation of a word is built from its actual neighbourhood rather than from a fixed dictionary vector. Masked language modelling is the pre-training trick: hide 15% of the tokens and force the model to reconstruct them from both sides, which is what makes the encoder genuinely bidirectional instead of two one-directional passes glued together. Next sentence prediction, the second pre-training objective in the original paper, asks whether sentence B plausibly follows sentence A, which is where a crude sense of discourse coherence comes from. Later research questioned how much NSP actually contributes, and several successor models dropped it.

This short overview covers the components in sequence if you want the visual version before we go further into architecture.

Put it into practice?

Nautilinks operates an owned network of editorial media. In-house written articles, transparency disclosures respected, anchor mix calibrated.

See pricing → Buy backlinks service
BD
Benoit Demonchaux Founder · Nautilinks

Founder and operator of Nautilinks. Edits and writes the site's editorial glossary, as well as the content published across the Nautilinks network of editorial media.

Frequently asked questions

Is BERT still used by Google in 2026, or was it replaced by MUM and Gemini?

Still used. The version of Google's ranking systems documentation in force in December 2025 lists BERT among active named systems, alongside RankBrain, neural matching, passage ranking and link analysis. Retired systems are moved to a separate list, which is where the Helpful Content system went after being merged into core ranking on 5 March 2024. MUM and Gemini added capabilities on top, particularly for AI Mode and AI Overviews, but the encoder layer that matches queries to documents was not removed.

Is BERT supervised or unsupervised, and does that matter for SEO?

Pre-training is self-supervised: the labels are generated from the text itself by masking tokens and asking the model to reconstruct them, so no human annotation is needed. Fine-tuning for a specific task is supervised. It matters indirectly, because RankEmbed BERT, described in DOJ trial documentation summarised in September 2025, was trained on 70 days of search logs plus quality rater scores. That is supervised signal derived from human judgement, which is closer to E-E-A-T than to anything in your backlink profile.

How is BERT different from word2vec for keyword clustering?

word2vec gives one fixed vector per word regardless of context, so « jaguar » is a single blurry average of the car and the animal. BERT produces a different vector for the same word depending on the sentence it sits in. For clustering a keyword list, that means BERT-family encoders separate intents that word2vec merges. If you are building this in-house, a distilled multilingual encoder run locally is enough, and for French corpora CamemBERT is the reference model.

Does BERT read anchor text, and should that change how I write anchors?

Anchor text is text on a page, so it is processed in context like everything else. The practical consequence is that a natural phrase inside a real sentence carries its topical meaning fully, without needing exact-match repetition. Stacking identical exact-match anchors across placements no longer buys extra relevance, it just makes the pattern legible. Vary anchors, let the surrounding paragraph do the topical work, and pick placements where that paragraph is genuinely about the target subject.

Can BERT generate text, and could Google use it to write AI Overviews?

No. BERT is encoder-only: it sees the full sequence at once and outputs representations, which is exactly why it is good at matching and classification and structurally incapable of generating fluent continuations. Generation needs a decoder that predicts the next token. AI Overviews and AI Mode are produced by Gemini models, with Gemini 3 Flash becoming the standard model for AI Mode worldwide between December 2025 and January 2026. The retrieval and ranking underneath still relies on encoder-side semantics.

How does BERT handle a brand name or a word it has never seen?

It splits unknown strings into sub-word units using WordPiece, so nothing is discarded, and it builds a representation from the surrounding context. For a new brand this is good news and a constraint at the same time: the model can learn what your brand is from the sentences around it, but only if those sentences actually define it. A brand mention dropped into a paragraph with no descriptive context contributes almost nothing at the semantic level.

Quiz

Test your knowledge

Quiz: BERT

1/3

What does the bidirectional part of BERT actually refer to?

Newsletter

GEO + SEO analyses and network case studies, in your inbox

Once or twice a month at most. No filler. One-click unsubscribe.

By subscribing you agree to receive our emails. See our privacy policy.

One detail with direct SEO consequences: BERT tokenises with WordPiece, meaning an unknown word is broken into known sub-units rather than dropped. A brand name it has never seen gets decomposed and still receives a representation from context. That is precisely why a new brand can be understood correctly when the surrounding copy defines it, and why an entity mentioned once in a link-bait paragraph with no context around it contributes almost nothing semantically.

The other detail is the sequence limit. The classic model handles 512 tokens, which is roughly why passage-level analysis exists as a separate system: long documents get read in chunks, and a strong passage in the middle of a mediocre page can rank on its own. Practically, that means burying the answer under 800 words of throat-clearing does not just annoy readers, it puts the useful passage in a different window from the query terms and the surrounding signals.

Two-column comparison between BERT, a Transformer encoder that understands and matches, and GPT or Gemini, autoregressive decoders that produce text.
Generative models did not replace BERT: they do not do the same job.

Architecture, variants, and why BERT is not GPT

BERT-base is 12 stacked encoder layers with 768-dimensional hidden states and 12 attention heads, around 110 million parameters. BERT-large is 24 layers, 1024 dimensions, 16 heads, around 340 million parameters. By 2026 standards those numbers are small, which is the point: the model is cheap enough to run on every query at Google scale, which a frontier generative model is not.

This walkthrough goes deeper into the encoder stack and the training procedure if you want the mechanics at a lower level.

The comparison people keep getting wrong: BERT is encoder-only, GPT is decoder-only. BERT sees the whole sequence at once and produces representations, which is ideal for classification, entity recognition, and matching a query to a document. GPT predicts the next token and cannot look ahead, which is what makes it generative. Asking BERT to write your article is a category error, and so is assuming that because Google now generates answers with Gemini, the encoder side of the pipeline became irrelevant. Retrieval still needs representations. Generation consumes them.

Variants matter mostly if you are building your own tooling: RoBERTa retrains without NSP on more data, DistilBERT trades a few points of accuracy for roughly half the size, ALBERT shares parameters across layers, CamemBERT and FlauBERT are the French-language versions worth knowing if you classify French corpora in-house. For clustering keyword sets or deduplicating a content plan, a distilled multilingual encoder on your own machine is still the sane default in 2026, and it costs nothing per call.

What BERT changes for a netlinking operation

The blunt version: semantic matching moved from a filter applied after authority to a gate applied before it. A link from a domain with a strong profile pointing at a page whose topic sits nowhere near the linking page's topic still passes something, but it no longer compensates for a document that does not answer the query. That is consistent with what the trial documents describe, where information about the page itself outweighs raw PageRank.

Operationally this reshapes three things. Placement selection stops being a metrics exercise and becomes a topical one: the surrounding paragraph, the category of the host site, and the entities already present on the linking page all feed the interpretation of your link. Anchor text stops rewarding exact-match repetition, because a natural phrase in a coherent sentence carries the same topical signal without the pattern risk. And the article hosting the link has to be a real article, since a passage assembled to hold a link reads as exactly that once the surrounding context is what determines meaning. That is the reasoning behind running 50 owned French editorial media in-house rather than buying whatever placement is available: you control the context the link lives in, not just the domain it comes from. Same logic when you pick placements from a catalogue you can filter by topic before you commit, and when you spread a campaign across several months instead of a single burst.

The traffic side deserves a number. An Ahrefs analysis of 300,000 keywords on December 2025 data found the CTR of the top organic result is around 58% lower when an AI Overview is present. Seer Interactive's longitudinal data on informational queries between June 2024 and September 2025 shows organic CTR on AI Overview queries falling 61%, from 1.76% to 0.61%. Zero-click searches went from 56% to 69% of Google queries between May 2024 and May 2025 according to Similarweb data. And yet an analysis of 40,000 US sites found organic search traffic down only 2.5% year over year comparing February to December 2024 against January to November 2025. Distribution shifted, demand did not collapse. The practical read: informational rankings convert to clicks less reliably than they did, so link campaigns pointed exclusively at top-of-funnel content are buying a smaller asset than the position implies.

What we see go wrong

Selling « BERT optimisation » as a deliverable is the first one, and it is fluff. There is no BERT score, no BERT audit, no setting to tune. Any agency invoicing a line item for it is selling readability work under a technical label. Write unambiguously, disambiguate entities on first mention, keep one intent per URL, and you have done everything that can be done.

The second is confusing BERT with the systems that punish. BERT interprets, it does not sanction. Ranking drops after a broad core update or after a quality signal folded into core ranking get blamed on semantics when the cause is quality evaluation, and the two are handled by different parts of the stack even though they run in the same core system. The March 2024 core update ran 45 days, Google's longest, and was reported to remove around 45% of unhelpful content from results. That was not BERT changing its mind about your pages.

The third is over-correcting on semantics and abandoning entity coverage. Because BERT works on context rather than exact strings, some teams concluded that keyword research no longer matters and started writing without any lexical target at all. Wrong conclusion. The model still needs the entities and the qualifiers to be present in the text to build a representation of them. What died is exact-match repetition, not topical coverage.

The fourth, the expensive one, is treating relevance and authority as separate budgets. Teams buy links on the basis of domain metrics, then wonder why a page that has never once been rewritten to match the query it targets stays on page two. The sequencing that works: fix the document so the interpretation layer has something unambiguous to score, then send links at it. Reversed, you pay for authority that the semantic gate discards before it can help.