Before Google ranks a single result, every search query passes through a multi-stage pipeline of parsing, AI-powered understanding and semantic expansion.
- Tokenization, stop word analysis, synonym expansion and entity recognition all happen before ranking even begins.
- RankBrain, BERT (internally DeepRank), Neural Matching and MUM work in parallel – each system handles a different part of query understanding.
- Understanding how Google processes queries means writing content that targets concepts, entities and semantic relations – not just keywords.
15% of all daily search queries have never been seen by Google before. That number comes directly from Google – and it reveals a fundamental challenge: a search engine cannot rely on exact keyword matches when a significant share of queries is entirely new.
In my work as a Product Developer at iGaming.com and across my client projects at SEO Kreativ, I regularly see SEOs optimizing content exclusively for keywords – while completely overlooking what Google actually does with a search query before ranking even begins.
Between typing into the search box and seeing the ten blue links (or the AI Overview), there is a complex processing chain: query parsing, tokenization, synonym expansion, entity recognition, intent classification – and only then does the actual ranking begin. Understanding this pipeline explains why some pages rank even though not a single keyword from the search query appears in the text.
In this article, I break down Google’s complete query processing pipeline – step by step, based on official sources like Gary Illyes’ “Search Off the Record” podcast, Google patents and findings from the 2023 DOJ trial.
What Is Query Processing?
When you enter a search query into Google, it is not sent to the index in its original form. Gary Illyes, Google’s Search Analyst, described the parsed version of a search query in the “Search Off the Record” podcast as “an absolute monstrosity” compared to the original.
This is by design. The Google index does not understand natural language – it works with tokens, vectors and posting lists. The job of query processing is to bridge the gap between human language and machine-level index queries.
The process can be roughly divided into three phases:
| Phase | What Happens | Example |
|---|---|---|
| 1. Parsing | Tokenization, stop word analysis, segmentation | “what are the best restaurants in berlin” → [best, restaurants, berlin] |
| 2. Understanding | Synonyms, anti-synonyms, plural forms, entity recognition | “book holiday” → [book, holiday] OR [book, vacation] OR [book, trip] |
| 3. Expansion | Related terms, intent signals, language context | “book holiday” + [holidays, booking, travel agency] + Intent: transactional + Language: EN |
From my work on technical SEO audits, I know that most content strategies only address the ranking stage. But if you understand what happens before ranking, you can create content that ranks not just for one keyword – but for the entire semantic field around it.
The Query Parsing Pipeline in Detail
Tokenization & Segmentor
The first step: Google breaks the character string into individual words. For English, this seems trivial – spaces separate words. But Google processes queries in over 100 languages, and in languages like Thai or Chinese, there are no spaces between words.
For this, Google developed a system called the Segmentor, as Gary Illyes explained in “Search Off the Record.” The Segmentor splits character strings into individual words that can then be looked up in the index.
Segmentation is also relevant for compound-heavy languages like German: words like “Bundesligatabelle” (Bundesliga table), “Kaffeemaschinentest” (coffee machine test) or “Autoversicherungsvergleichsportal” (car insurance comparison portal) need to be broken into their components so Google can correctly match them against the index. This is non-trivial – “Kaffeemaschinentest” could mean “Kaffeemaschinen Test” (two words), “Kaffee Maschinen Test” (three words), or the compound form. Google’s Segmentor must recognize all variants as semantically identical.
Stop Word Analysis
In the next step, Google removes stop words – words like “the,” “a,” “and,” “in,” “of.” According to Gary Illyes, these words are irrelevant for search in most cases because they appear on virtually every web page.
But – and this is crucial – Google recognizes cases where stop words change the meaning. Consider “Who Wants to Be a Millionaire”: the query consists almost entirely of stop words, yet every single one is necessary to identify the entity. Remove “Who,” “Wants,” “to,” “Be,” “a” and you are left with just “Millionaire” – a completely different query. Google must decide per query which stop words carry meaning and which can safely be dropped.
Entity Recognition
Once tokenization is complete, Google attempts to identify entities in the search query. An entity is a uniquely identifiable concept – a person, place, product or brand. This gets complex quickly with ambiguous terms: “Bayern Munich” could refer to the football club, the city or the German state. Google must decide from the query context which entity is meant.
According to a Google patent (US9542450B1), Google uses a confidence score to evaluate the relevance of an entity within a search query. Based on this score, the system decides which entity is primary – and accordingly selects the matching document corpus from the index.
In my practice, I frequently see pages that fail to rank for ambiguous queries despite strong content. The reason: Google assigned the query to a different entity. Equipping your content with structured data, clear entity signals and consistent terminology helps Google with this assignment.
Synonym Expansion & Anti-Synonyms
After parsing comes what Gary Illyes called the “magic”: expanding the search query with synonyms. Practically: if you search for “book a holiday,” Google expands the query to “book a vacation,” “book a trip,” “plan a holiday” and further variants. This happens automatically – the user types one version, Google searches with all of them.
Importantly: synonyms are not stored in the index but added at query time. This keeps the index lean and the matching flexible.
Google also knows anti-synonyms – cases where terms appear similar on the surface but are not interchangeable. “Book” and “cancel” are related terms in an information retrieval sense, but the intent is opposite. According to Gary Illyes, Google deliberately weights such pairs differently: a page about “cancel a holiday” will rank much lower for the query “book a holiday,” even though the words are thematically close.
Pluralization & Inflections
Google also automatically injects plural and inflected forms into the parsed query. “Book holiday” additionally becomes “book holidays” and potentially “booking holiday” or “holidays” alone. This step ensures the index delivers broader matches without the user having to manually enter every variant.
Language Detection
Google determines the language of a search query primarily via the browser’s Accept-Language header, as Illyes confirmed. If the browser sends English, Google searches primarily in English documents. However: if a search query is obviously in another language, this overrides the browser setting.
Spell Correction
Spell checking is also part of query processing. Based on current analysis, Google uses methods such as Levenshtein distance to detect typos and identify the likely intended spelling. The familiar “Did you mean…” feature is the visible output of this process.
Query Understanding – From Keywords to Concepts
Pure parsing – tokenization, stop words, synonyms – is the foundation. But the real intelligence sits in Google’s AI systems, progressively integrated into query processing since 2015. Their role in query processing can be summarized as follows:
RankBrain (since 2015) was Google’s first machine learning system in search. It transforms queries into mathematical vectors and compares them with historical searches – especially relevant for the 15% of daily queries Google has never seen before, according to official documentation. In the context of query processing, RankBrain is the system that recognizes an unknown query like “how do I connect a remote” as semantically similar to “pair remote control.”
BERT (since 2019) reads words bidirectionally – capturing context from both directions simultaneously. For query processing, this means: BERT recognizes that “for” in “get medicine for someone else” changes the entire intent. Without BERT, this nuance would be invisible. Internally, BERT is called DeepRank in the ranking context, as confirmed during the 2023 DOJ trial.
Neural Matching (since 2018) operates at concept level. It maps queries and documents as vectors in a multidimensional meaning space. Example: searching for “tie my laces” also returns results about “tying shoes” – because Neural Matching recognizes the conceptual overlap even though no word is identical.
MUM (since 2021) is, according to Google’s Ranking Systems Guide, multimodal – it understands text, images and potentially video across 75+ languages. In query processing, MUM is currently used primarily for complex information queries, not for regular ranking of all search results.
From Parsed Query to Ranking
Posting Lists & Inverted Index
Once the query is parsed and understood, it is sent to the index. Google does not search through the entire index – that would be impossible with billions of documents. Instead, Google uses posting lists (also called the inverted index).
Gary Illyes explained the principle in the “Search Off the Record” podcast: a posting list is a mapping of terms to documents. The term “car” appears in documents A, B, C, D. The term “buy” appears in B, C, D, E. Google forms the intersection: B, C, D contain both terms and become candidates for the SERPs.
In practice, this is far more complex: Google’s index is distributed across thousands of machines (shards), and the query is executed in parallel across all relevant shards. Each shard returns its best matches, which are then merged and ranked globally. Communication between shards uses serialized data formats like Protocol Buffers – optimized for minimal latency with maximum data density. The posting lists themselves are weighted – they store not only whether a term appears in a document, but also where (title, H1, body, anchor text) and how often. From my work on technical SEO audits, this is the direct connection: why title tags and H1s carry so much more weight than body text lies in exactly this posting list architecture.
From Ranking to AI Mode: Query Fan-Out
The candidate documents pulled from the index then pass through a multi-stage ranking process – from coarse relevance filters to final re-ranking by DeepRank (BERT for ranking) and the user signal system NavBoost. For a detailed analysis of this ranking phase with DOJ evidence on NavBoost, goodClicks/badClicks and slicing by location/device/query type, see my article on Google’s AI ranking systems.
What has been added since 2025: Google’s AI Mode (globally available since late 2025, powered by Gemini 3) adds a new layer to the classic pipeline. For complex queries, Google performs what is called Query Fan-Out – the original query is decomposed into multiple sub-queries that are processed in parallel and then synthesized by Gemini into a combined answer. The classic query processing remains the foundation: even in AI Mode, every sub-query must be parsed, tokenized and enriched with synonyms before it hits the index.
What Query Processing Means for Your SEO Strategy
From my work at iGaming.com and across my SEO consulting projects, I have developed a rule of thumb: for every phase of query processing, there is a specific SEO lever. Here is the summary:
| Pipeline Phase | SEO Lever | Practical Example |
|---|---|---|
| Tokenization | Cover compound words in split variants too | “search engine optimization” + “search engine optimisation” |
| Stop words | Keep meaningful stop words in titles/H1 | “Who Wants to Be a Millionaire” – do not shorten to “Millionaire” |
| Entity Recognition | Structured data (Schema.org), clear entity description | Organization markup, sameAs references, Knowledge Graph optimization |
| Synonym expansion | Natural synonym variation in text | “book holiday” + “book vacation” + “plan trip” |
| RankBrain | Cover long-tail queries through topical depth | Not just “SEO tips” but covering the entire topic field |
| BERT / DeepRank | Write context-sensitively, do not underestimate small words | “for” in “get medicine for someone” changes the entire intent |
| Neural Matching | Conceptual coverage instead of pure keyword density | Topic clusters, semantically related terms, topical authority |
The core insight: keyword optimization alone addresses only the parsing phase. BERT, Neural Matching and MUM operate on a higher level – on concepts and intent. Aligning your content with semantic completeness, topical depth and clear entity signals addresses the entire pipeline.
I still regularly see pages that want to rank for a keyword but offer neither synonyms, nor related terms, nor a clear intent alignment. Such pages typically reach position 10-20 at best – they pass the parsing phase but fail at understanding. Why understanding AI-driven query processing matters more than ever is something I explored in my interview about SEO in the age of AI.
For more on how to build topical authority systematically, see my article on crawling and indexing at Google.
Infographic: Google’s Query Processing Pipeline
Frequently Asked Questions (FAQ)
What is the difference between query parsing and query understanding?
Query parsing is the technical decomposition of a search query into its components – tokenization, stop word removal, segmentation. Query understanding goes beyond that and encompasses semantic interpretation: Which entities are in the query? What is the intent? Which concepts are related? Parsing is step 1, understanding is step 2.
Does Google use BERT for every search query?
Not necessarily. Google decides per query which system or combination to use. BERT is particularly strong for queries where small words change the context. For simple single-word queries, RankBrain or classic keyword matching usually suffices.
Why does my page rank for a keyword that does not appear in the text?
This is due to synonym expansion and Neural Matching. Google expands search queries with synonyms and recognizes conceptual overlap. If your text covers “plan a trip,” it can also rank for “book a holiday” because Google establishes the semantic connection.
How does location affect query processing?
Location factors in at multiple levels: language detection (browser language), intent interpretation (local vs. global search intent) and ranking via NavBoost (click behavior varies by region). A query like “order pizza” will deliver different results in New York than in London.
Does Google store synonyms together with the index?
No. According to Gary Illyes, synonyms are not stored in the index but added at query time. This means synonym mapping can change at any time without Google having to rebuild the entire index. For SEO, this means: natural synonym variation in text helps but is no substitute for clear semantic signals.
Conclusion: Understanding Query Processing Means Better SEO
Query processing is not an abstract technical topic – it is the foundation of every SEO strategy. Every decision Google makes during parsing and understanding directly influences whether your content is classified as relevant or not.
My practical recommendation: before writing your next article, do not just ask “Which keyword am I optimizing for?” – ask “Which entities, synonyms, concepts and intent signals does my content need to cover so that Google’s entire processing pipeline classifies it as optimally relevant?”
If you want to dive deeper into the individual AI ranking systems, read my analysis of Google retiring the num=100 parameter – a practical example of how changes to query processing directly impact SEO workflows.
