BERT

What BERT is

BERT (Bidirectional Encoder Representations from Transformers) is a language model Google announced for Search on October 25, 2019, that reads words in both directions at once — so context from before and after a word shapes its meaning. The stake for SEOs was never "optimize for BERT"; it was that a whole class of long, conversational, preposition-heavy queries suddenly started matching pages by meaning instead of by shared strings, redistributing traffic accordingly.

Google called it the biggest leap forward in Search in five years. At launch it affected roughly 10% of English queries in the US — Google's own figure — plus featured snippets in the two dozen languages where snippets ran. By late 2020, Google stated BERT was involved in almost every English query. BERT sits in the NLP lineage that runs from keyword matching through to MUM.

The documented example that explains everything

Google's launch example: "2019 brazil traveler to usa need a visa." Pre-BERT, Search ignored the word "to" and returned results about US citizens traveling to Brazil — the exact wrong direction. BERT understood that "to" makes the traveler Brazilian and the destination the USA, and returned US embassy visa information. Another launch example: "can you get medicine for someone pharmacy" — BERT caught that "for someone" was the whole point of the query. Small words, damn big meaning shifts. That's the entire model in two queries: function words stopped being throwaways.

Under the hood, in one paragraph

BERT was pretrained with masked language modeling: hide ~15% of the words in a sentence and train the network to predict them from everything around the blank — left context and right context simultaneously. That's the "bidirectional" part, and it's why BERT handles function words so well: predicting a masked "to" or "for" forces the model to internalize what those words do. Google open-sourced the research model in 2018, a year before the Search rollout, which is why you'll find BERT-descended encoders inside half the content-scoring tools you already pay for. Same family, very different tuning from whatever runs in production Search.

What BERT changed vs. what it didn't

How to check whether BERT-era matching affects you

1. In Search Console, filter queries containing function words: "for," "without," "vs," "can i," "how to ... with." These are the query class BERT re-interpreted.
2. Compare that segment's impression trend against your head terms. If long conversational queries are growing while your pages answer them only implicitly, you're leaving matches on the table.
3. Pick five such queries and read your ranking page aloud. Does any sentence actually answer the question as phrased? If the answer only exists "between the lines," make it explicit.
4. Audit your snippet losses: for queries where you rank top-5 without the snippet, check whether the winner states the answer in one plain sentence under a matching heading. That pattern is what post-2019 passage systems reward — our passage indexing explainer covers the mechanics.
5. Kill any lingering "write like a robot" habits in your style guide — dropped articles, unnatural keyword phrasings ("plumber Denver cheap") now parse as exactly the gibberish they are.

Common mistakes

• Buying "BERT optimization" services. There is nothing to optimize. Anyone selling it is selling the name, not a method.
• Blaming BERT for a traffic drop. If your rankings moved in late 2019, correlation isn't diagnosis — check the algorithm update history before assigning blame; several core updates landed the same year.
• Concluding keywords no longer matter. Meaning-based matching still starts from words. Topic and phrasing research matter as much as ever; robotic exact-match insertion is what died (and it died before BERT).
• Writing longer because "BERT reads context." Bidirectional context is a parsing capability, not a length preference. Padding dilutes the passages that would have won.
• Ignoring negations in your content. Queries like "parking on a hill with no curb" (another Google launch example) now surface pages that address the no-curb case specifically. Cover the negative and edge cases your competitors gloss over.

FAQ

Is BERT a ranking factor?

No. It's a language-understanding system used in query interpretation, passage understanding, and featured snippets. It changes which pages are seen as relevant; it doesn't score your domain.

Is BERT still used now that MUM and Gemini-era models exist?

Yes. Google has repeatedly described its ranking stack as many models running together — BERT-class encoders remain workhorses because they're efficient at query-scale, while larger models handle narrower jobs.

Can I run BERT myself?

The original model is open-sourced (google-research/bert on GitHub), and it powers half the SEO tooling that scores "semantic relevance." Useful for research; just remember Google's production systems are tuned far beyond the public checkpoint.

Did BERT affect languages other than English?

At launch, snippet improvements shipped in about 25 languages; query understanding expanded to 70+ languages by December 2019. One documented strength: transfer — improvements learned in English lifted other languages too.