Google BERT: what natural-language understanding means for content

On October 25, Google confirmed that it had started applying BERT, short for Bidirectional Encoder Representations from Transformers, to search ranking and to featured snippets. Google called it the biggest leap forward in search understanding in five years, since Hummingbird in 2013, and said it would affect roughly 10 percent of English queries in the United States. Featured snippets are getting the change in around 24 languages at the same time.

The technical idea is the "bidirectional" part. Older language models read a sentence one direction at a time, so a word was interpreted mostly from what came before it. BERT looks at the words on both sides of each term at once, which is how it works out that "2019 brazil traveler to usa need a visa" is about a Brazilian going to the US, not an American going to Brazil. Google's own examples were all like that: prepositions and function words that earlier systems treated as noise but that carry the meaning of the question.

Two things follow from that. First, the update acts on the query side, interpreting what the searcher meant. Second, it skews toward longer queries, which is where voice search and typed questions live. Short head terms like "plumber chicago" barely change.

Not every site will notice anything. Based on what Google has said and what we are seeing in client accounts in the first two weeks, these are the places to look.

The last row is the important one. For years a page could rank for a long query simply because it contained most of the words in it. If a searcher typed "can you get medicine for someone pharmacy" and your page mentioned pharmacies and medicine, you might have appeared. BERT understands that the question is about picking up a prescription on someone else's behalf, and it will favor the page that addresses exactly that. Traffic lost this way was never qualified traffic.

Before you rewrite anything, find out what actually moved. This takes an hour in Google Search Console.

1. Open the Performance report and compare the 14 days from October 26 against the 14 days before October 12, so you avoid the rollout week itself.
2. Switch to the Queries tab and sort by click change. Export to a spreadsheet if you have more than a few hundred rows.
3. Flag queries of five or more words. If the losses and gains concentrate there, BERT is the probable cause. If they are spread across short terms, look elsewhere first: the September core update, a technical regression, or a seasonal dip.
4. For each lost long query, search it yourself in an incognito window and read the pages now ranking. Ask honestly whether they answer the question better than yours.
5. Check the Pages tab for the same period. A handful of pages usually account for most of the movement, which tells you where to spend editing time.

Do not confuse BERT with the featured snippet changes that landed alongside it. If your position held but clicks fell, you may have lost a snippet rather than a ranking.

There is no checklist that makes a page "BERT-friendly," but the update rewards habits that good writers already have and that a lot of SEO content has drifted away from.

Answer the question in the first paragraph. State the direct answer, then explain. Pages that spend 300 words on context before getting to the point are harder to match to a specific question and less likely to win the featured snippet.

Use the words people actually say, including the small ones. "How to return a mattress bought online" is a better heading than "Online Mattress Returns Guide." The preposition and the verb are now meaningful signals.

One question, one section. A page built as a series of clear H2 questions with complete answers underneath maps cleanly to how queries are now interpreted. Avoid headings that are clever rather than descriptive.

Stop forcing exact-match phrases. "Best plumber Chicago cheap" as a heading was always awkward. It now has no advantage over a sentence that reads naturally, and it may read as lower quality to a human.

Cover the follow-up questions. If someone asks whether they need a permit for a deck, they also want to know how much it costs and how long approval takes. A page that handles the cluster of related questions is more useful and will match more of the variations BERT now recognizes.

Be precise about who the page is for. Many long queries carry an implicit subject: a parent, a tenant, a non-resident. Say explicitly who your answer applies to. Ambiguity used to be harmless; now it costs matches.

• Do not buy a "BERT optimization" service. There is nothing to optimize; vendors selling one are selling the old keyword-density work under a new label.
• Do not rewrite pages that are performing. Sites that lost nothing should keep their process and apply the writing habits above to new content.
• Do not strip out all short keywords. Head terms still work the way they did last month. This update is additive, not a replacement.
• Do not treat it as a reason to publish more. Volume was never the lever, and thin pages written around long-tail phrases are exactly what gets re-routed now.
• Do not panic over single-day fluctuations. Give any comparison at least two full weeks on each side.

BERT is one of a series of moves that make Google better at matching meaning rather than strings, following Hummingbird, RankBrain in 2015 and neural matching last year. Each one shifts value toward pages that genuinely answer a specific need and away from pages engineered to match a phrase. Voice assistants, which handle almost entirely conversational queries, push the same direction. The reasonable planning assumption for 2020 is that the share of queries interpreted this way grows and that the approach spreads to more languages, so the writing habits above are worth adopting whether or not you saw movement this month.

Search is one input into a connected plan across brand, growth, web and technology, so we do not treat an algorithm update as an isolated fire drill. Discover starts with the data: Search Console query and page analysis across the update window, a content inventory, and an honest read of which pages answer a question and which merely mention it. Plan maps your audience's real questions, in their words, to pages that will own them, including which existing pages to consolidate and which to retire. Build is the editorial work: restructured pages, question-led headings, clear first-paragraph answers, and the technical fixes such as schema and internal linking that help the right page get matched. Improve is a monthly review of query-level performance against the plan, so you learn what is working in your category rather than in general.

The content program engagement is the representative pattern. A business had grown a large library of articles written around keyword phrases, with rankings that drifted every time Google changed how it read queries. After a discovery that matched real customer questions to existing pages, we consolidated overlapping articles, rewrote the strongest ones to lead with direct answers, and established an editorial process built on questions rather than keywords. Organic traffic became more stable through subsequent updates, and more of it came from visitors who found what they were looking for.