Does BERT actually reduce traffic to websites or simply enhance targeting?

What changes does BERT bring to how Google operates?

BERT (Bidirectional Encoder Representations from Transformers) is a natural language processing model that enables Google to better understand the context of words in a query. Before BERT, the engine analyzed words in a more isolated manner, resulting in approximate results for long or complex queries.

With BERT, Google processes the relationships between words in a bidirectional manner. The meaning of a term now depends on the words that come before AND after it. This radically changes the game for conversational queries, natural language questions, and any formulations where nuance matters.

Why do some pages lose traffic with BERT?

Danny Sullivan puts it plainly: BERT can lead to a drop in traffic for certain pages. But this is not a penalty. It’s a targeting correction. If a page was receiving traffic on queries where the user intent didn't really match the content, BERT detects this and redirects that traffic elsewhere.

In practical terms, a page optimized for “buy running shoes” might have captured traffic on “how to choose running shoes” simply because the keywords matched. BERT now understands the difference between transactional and informational intent—and adjusts the results accordingly.

Should this decrease be viewed as a negative signal?

No. Sullivan insists: the redirected traffic is more relevant for the user. In other words, if you're losing traffic with BERT, it was probably poorly qualified traffic that wasn't converting or that generated a high bounce rate. Users arrived, didn't find what they were looking for, and left.

For an SEO, this requires a shift in perspective. Measuring gross traffic volume becomes less relevant. What matters is the alignment between the query, intent, and content. A 20% drop in traffic can mask a 30% increase in conversion rate if the remaining traffic is better targeted.

• BERT processes the bidirectional context of words in a query, not just their isolated presence.
• Pages may lose traffic if they were capturing poorly aligned queries with their actual content.
• This loss is not a sanction—it’s a relevance correction.
• Traffic volume alone is no longer enough as a KPI: one must analyze quality and intent.
• BERT favors content that precisely addresses the intent behind the query formulation.

Is this statement consistent with what we observe on the ground?

Yes and no. Sullivan's theory makes sense: BERT improves relevance, so poorly targeted traffic declines. On the ground, it is indeed observed that generalist sites or those too vague in their positioning lose traffic on long-tail queries that they captured by default. More specialized sites, with precise and targeted content, often gain these positions.

But the nuance that Sullivan does not mention: BERT does not correct all targeting errors. There are still SERPs where transactional pages rank for informational queries and vice versa. BERT is progress, not a miracle solution. [To be verified] in certain underdocumented niches, where the training dataset volume was probably insufficient.

What are the blind spots of this statement?

Sullivan talks about “more relevant traffic”, but he doesn’t provide any concrete metrics. More relevant according to what criteria? The bounce rate? The time spent on the page? The conversion rate? This lack of transparency is typical of Google's communications: they validate a general principle without providing the data that would allow it to be measured.

Another point: Sullivan says nothing about the content that now captures this redirected traffic. If BERT redirects better, who benefits? Authority sites? Recent pages? Long-form content? We lack visibility on the real beneficiaries of this redistribution. [To be verified] by cross-referencing data from various sectors to identify recurring patterns.

In which cases does this logic not hold?

BERT performs well on natural language queries, long questions, conversational formulations. But for short queries (1-2 keywords), its impact is negligible or marginal. If a site loses traffic on short queries after BERT's deployment, it’s probably not BERT that’s to blame—one needs to look elsewhere (competition, content update, concomitant Core Update).

Likewise, BERT doesn’t really help with ambiguous intent queries. “Avocado” can refer to the fruit or the legal professional. BERT understands the context of a sentence better, but on a single word, it still relies on classic signals (search history, location, aggregated data). Don’t overestimate its capabilities.

How can I check if BERT has affected my site?

The first step: analyze the queries that have lost traffic in Google Search Console. Filter for long queries (5 words and more) and natural language questions. If these queries dropped sharply after BERT's integration, it’s a good indicator that your content wasn't perfectly aligned with user intent.

Next, look at the associated landing pages. Does the content truly answer the question posed in the query? Or is the page optimizing a broad lexical field without precisely addressing the intent? If your page “complete guide to running shoes” was capturing traffic on “what size to choose for running shoes”, BERT has probably redirected that traffic to a more specific page.

What concrete steps should be taken to adapt?

Rethink the content architecture by better segmenting intents. Instead of one catch-all page, create dedicated pages for each intent: one for buying, one for choosing, one for comparing. Each page should answer a specific question, phrased in natural language in the H1 or introduction.

Optimize conversational formulations. BERT understands natural phrases, so write as your users speak. Integrate questions into your titles, subtitles, and paragraphs. “How to choose the right size?” rather than “Size guide.” This approach enhances both BERT’s understanding and the user experience.

What mistakes should absolutely be avoided?

Do not try to stuff your pages with long-tail variations in hopes of capturing all nuances of intent. BERT detects content that attempts to cast a wide net without really responding. Better to have a short and ultra-targeted page than a long and vague one.

Avoid also neglecting traffic quality metrics. If your traffic declines but your conversion rate increases, this is a positive signal. Don’t panic over a downward curve without analyzing what’s happening downstream. BERT may cause you to lose 1000 poorly qualified visits and gain 200 visits that convert—the net result is positive.

• Audit long and conversational queries that have lost traffic
• Check the alignment between query intent and landing page content
• Segment content into pages dedicated to a specific intent rather than general pages
• Integrate natural language formulations into titles and subtitles
• Measure traffic quality (conversion, engagement) and not just volume
• Don’t attempt to cast a wide net—favor precision on each page