They’re two separate systems addressing two different parts of the same problem, and treating them as interchangeable (a common shorthand in SEO writing) obscures what each one actually does. RankBrain, which Google confirmed publicly in 2015, is a machine-learning system that helps interpret queries, particularly ones Google hasn’t seen before, by relating them to patterns found in similar queries it has processed previously. Its job is on the query side: taking an ambiguous or novel string of words and forming a usable interpretation of what’s being asked, largely by finding conceptual neighbors among queries with known interpretations. Neural matching, which Google confirmed in 2018 (reported at the time via Danny Sullivan’s public comments and Google’s own discussion of Search improvements), works on a different part of the pipeline: matching queries to pages based on the broader concept being represented rather than requiring the literal words in the query to appear on the page. Google has described neural matching’s role using the example of a page that’s genuinely relevant to a query’s meaning but doesn’t use the specific words in the query itself, where prior word-based matching systems would struggle to surface that page and neural matching helps bridge that gap. So RankBrain helps Google understand what a query means, and neural matching helps Google understand which pages are relevant to that meaning even when the vocabulary doesn’t overlap. They’re complementary, not duplicative, and Google has been explicit in distinguishing them rather than presenting one as an evolution or replacement of the other.
What RankBrain actually does with a query
RankBrain’s original stated function, per Google’s 2015 confirmation, is helping process a meaningful share of daily queries, with particular value on the portion of queries Google hasn’t encountered before (a nontrivial share of daily search volume, historically, consists of novel query strings). For those novel or ambiguous queries, RankBrain’s approach is to map the query into a mathematical representation (a vector, in the embedding-space sense used broadly in machine learning) and find similar queries whose intent is already understood, using that similarity to infer what the new query is probably asking for. This is fundamentally a translation and clustering function: it doesn’t evaluate pages, and it doesn’t decide rankings on its own. It’s one signal among the broader ranking system, feeding an interpretation of query intent into the rest of the pipeline that then handles retrieval and ranking against that interpretation. Google has described RankBrain as one of the signals used in ranking, not as a separate ranking algorithm that operates independently of everything else.
This matters practically for interpreting SERP behavior: when a query with no exact historical precedent returns results that seem to reflect a broader or adjacent interpretation rather than a literal reading of the words, that’s consistent with RankBrain-style query interpretation doing its job of relating unfamiliar phrasing to familiar intent clusters. It’s not something practitioners can inspect directly (Google doesn’t expose which queries triggered RankBrain-specific processing), but the behavior it produces, ambiguous queries resolving toward a dominant, well-understood interpretation, is observable in aggregate.
What neural matching adds on the page side
Neural matching addresses a distinct limitation: even once a query’s intent is well understood, older retrieval approaches built heavily around matching specific words and phrases could fail to surface a page that used different vocabulary to describe the same underlying concept. Google’s own example when discussing neural matching involved a search related to the concept of a “why does my TV look strange” type of query, where the actual issue (the soap opera effect, a motion-smoothing artifact) doesn’t share vocabulary with how users describe the symptom; neural matching was described as helping connect that gap between how a concept is expressed in a query and how it’s expressed on a page. This is squarely a query-to-document matching function, applied after or alongside the query interpretation RankBrain contributes, working at the level of representing both the query and candidate documents in terms of underlying concepts rather than surface tokens.
Because both systems deal in learned representations of meaning rather than explicit rules, they get grouped together informally as “the AI parts of Search,” which is understandable but imprecise. Google has consistently described them in its public statements as separate systems that were developed, confirmed, and rolled out at different times (2015 versus 2018), targeting different parts of the query-to-result pipeline. Practitioners writing about either one should keep that distinction intact rather than using the names interchangeably, since conflating them tends to produce vague, unfalsifiable claims (“RankBrain reads semantic meaning on the page,” for instance, describes neural matching’s role, not RankBrain’s).
The practical upshot for how queries connect to pages
Together, the two systems describe a pipeline where an ambiguous or unfamiliar query first gets interpreted by relating it to known query patterns (RankBrain’s contribution), and the resulting understanding of intent then gets matched against candidate pages using conceptual rather than purely lexical similarity (neural matching’s contribution), with traditional ranking signals (relevance, quality, authority signals, and the rest of Google’s documented ranking systems) still applying on top of that matched candidate set. Neither system replaces keyword relevance as a baseline signal; Google has never suggested exact-match relevance stopped mattering. What both systems address is the gap between literal query and document text and the actual underlying meaning, from two different directions in the pipeline, which is the most useful mental model for practitioners trying to reason about why a page without exact-match keyword coverage might still rank for a semantically related query.
