RankBrain and search intent: what machine learning changes

The facts are thinner than the commentary. From the Bloomberg interview and the follow-up statements from Google, this is what we know.

• RankBrain uses machine learning to turn words into mathematical representations, so it can guess that an unfamiliar query is similar to familiar ones.
• It has been live for several months and processes "a very large fraction" of searches.
• Google ranks it as the third most important signal, without naming the first two. Most of the industry assumes those are links and content.
• Google engineers tested it against humans at predicting which pages would rank, and the system did better.

That is all. Nobody outside Google knows whether it is applied to every query or only ambiguous ones, how often the model is retrained, or how it interacts with Panda, Penguin and the mobile-friendly update from April. Treat any article that claims to know the weighting as speculation.

The behavior to adapt to is not new. It is the acceleration of a direction Google has been on since Hummingbird in 2013 and the Knowledge Graph in 2012. Three things shift.

Long-tail pages lose their moat. Many sites still carry hundreds of near-identical pages built for query variants: "plumber in Austin," "Austin plumber," "plumbing company Austin TX." If Google can tell that these mean the same thing, it no longer needs a page for each. The variants will consolidate around whichever page serves the intent best, and the rest become dead weight that dilutes crawl budget and internal links.

Phrasing matters less than coverage. A page that answers the question a searcher has, including the follow-up questions, will rank for phrasings you never targeted. That rewards depth over volume.

User behavior is probably part of the feedback loop. Google has not said that click-through rate or dwell time feed RankBrain, and you should be skeptical of anyone who says it has. But a learning system needs training data, and the most obvious source is what searchers do with the results. Writing titles and descriptions that earn the click and pages that satisfy it is a sound bet either way.

Here is the process we use when a client asks what to do about this.

1. Export what you already rank for. In Search Console, download queries and landing pages for the last 90 days. In Analytics, pull organic landing pages with bounce rate and pages per session. Join the two in a spreadsheet.
2. Label the intent of each query. Keep it simple: informational (how, what, why), commercial investigation (best, vs, review, cost), transactional (buy, near me, pricing, quote) and navigational (your brand name). Most lists fall into those four.
3. Check the match. A transactional query landing on a 2,000-word guide with no pricing or call to action is a mismatch. An informational query landing on a product page is a mismatch. Mark each row.
4. Group the variants. Sort by the page, and you will see clusters of queries that mean the same thing. Each cluster gets one target page.
5. Decide: improve, merge or retire. Improve the page that best serves each cluster. Merge the near-duplicates into it with 301 redirects. Retire the pages with no traffic and no links.
6. Write to the full question. For each target page, list the questions a real customer would ask in a sales call and make sure the page answers all of them, in the order they would ask.

• Answer in the first paragraph. State the answer, then explain. People on phones, now the majority of Google searches in the US, do not scroll to find out whether the page is relevant.
• Use the words customers use, not the words your industry uses. Pull them from support emails, sales calls and the Search Console export, not from a thesaurus.
• Use subheadings that are themselves questions or plain statements. They help both readers and the crawler see the structure.
• Cover the adjacent questions on the same page. If someone asking about cost also asks about timing and what is included, put all three on one page rather than three pages.
• Keep title tags honest. A title that over-promises wins the click and loses the visit. If behavior data feeds the model at all, that pattern is what it would learn to punish.
• Mark up what you can. Schema.org structured data for articles, products, reviews, local businesses and events is not a RankBrain factor, but it helps Google understand entities, which is the same direction of travel.

Do not chase "LSI keywords." Stuffing a page with synonyms pulled from a tool is the old behavior in a new costume. Write naturally and completely, and the related terms appear on their own.

Do not delete long-tail pages wholesale. Some of them have links and convert. Merge thoughtfully with redirects, one cluster at a time, and watch Search Console for two to four weeks between batches.

Do not confuse this with Panda. Thin, duplicate and scraped content is a Panda problem, and Panda 4.2 has been rolling out slowly since July. Intent mismatch is a different problem and the fixes overlap only partly.

Do not ignore mobile. The April mobile-friendly update and now RankBrain are both part of Google getting better at serving people on phones. If your site still fails the Mobile-Friendly Test, fix that before you touch content.

Do not expect a fast signal. There is no RankBrain penalty to recover from and no RankBrain boost to claim. Content consolidation typically shows in rankings over one to three months, and only if the consolidated page is genuinely better.

We treat search as part of a connected plan across brand, growth, web and technology rather than a standalone channel. Discover starts with the data: the Search Console and Analytics join described above, a crawl of the site to find duplication and thin templates, and interviews with sales and support staff about the questions customers actually ask. Plan turns that into a content map: one target page per intent cluster, a list of pages to improve, merge or retire, and the structured data and template changes the web team needs to make. Build is the writing, the redirects and the template work, done in batches so we can see the effect of each. Improve is a monthly review of query and landing page data, because the clusters shift as Google's understanding does, and the content map is a living document rather than a one-time deliverable.

The content program engagement is the representative pattern. A business with several hundred pages built up over years of keyword-targeted blogging had traffic that was flat despite publishing more each quarter. The work began with an intent audit of the existing pages rather than a new editorial calendar, consolidated overlapping posts into a smaller set of thorough guides with redirects, and rebuilt the publishing process around customer questions gathered from the sales team. Publishing volume went down, organic landing page performance and lead quality went up, and the editorial calendar now starts from a question rather than a keyword.