Knowledge base v2.13.0 · Community edition · data through 2 October 2026
Topic · Serving and ranking
Google rewrites pretty much every query before looking it up: it detects the language (falling back on location and browser settings for brand-only queries), drops stop words, recognises entities and expands the query with synonyms, so a short query can become a much longer internal one. Its synonyms are learned from search behaviour rather than catalogued by linguists, so words people use interchangeably count as synonyms, while 'siblings' such as Canon and Nikon are kept apart, and spellings with and without diacritics are generally treated as the same. Google's How Search Works page confirms the synonym and spelling systems; stop words, siblings and the rewriting detail were said at the event and are not in Google's docs. Fan-out queries from the AI features pass through the same query understanding, which also predicts which result types a query gets, using historical interaction data for ambiguous queries such as 'orange'. Query processing deliberately mirrors indexing, as Day 2 said of word segmentation, and the diacritics point extends Day 2's statement that Google usually understands words with or without accents; Google's summary slide added that typos and plurals need no attention because Google rewrites them automatically. Day 1's second recording put this at the start of serving: Cherry Prommawin said serving begins by interpreting the query, cleaning it up, detecting its language and expanding it; Gary Illyes contrasted 1990s searchers who had to 'speak machine' with today's natural-language queries, and said MUM helps interpret the context of query words. On Day 2 the non-Latin-script talk added that users may type the same Persian query in their own script or in Latin letters with the same intent, as Google's 2023 post on multilingual searches describes for Hindi.
Things in this topic 7
Counts are claims that name the thing. All things
What to do
Gary Illyes said that in the mid-1990s searchers had to speak machine, typing queries like 'weather Barcelona rain', whereas today machines understand how people naturally speak.
Speaker Gary IllyesIn Day 1, 11:35 · What's new in the world of SearchEvidence transcript
Serving starts with interpreting the query, which includes cleaning it up, detecting its language and expanding it.
Speaker Cherry PrommawinIn Day 1, 11:45 · How Search works and where's AI?Evidence transcript
Google said MUM helps it understand the context of the words in a query: a search for hiking shoes that mentions Mount Everest rather than Kilimanjaro gets results suited to that climb.
Speaker Gary IllyesIn Day 1, 11:45 · How Search works and where's AI?Evidence transcript
Users of non-Latin-script languages do not always search in their own script: the same Persian query may be typed in Persian script or in Latin letters, with the same intent and the same expected results.
Speaker a second community speakerIn Day 2, 14:30 · Lightning session G: InternationalisationEvidence transcript
Used byrequirement DEV-INT-11glossary term Transliterated queries
Google's post on multilingual searches (8 September 2023) says that, because of typing difficulty on some keyboards, a person in India might search in Hindi using Latin rather than Devanagari characters and want and receive Hindi results written either way.
Publisher Search Central blog (8 September 2023)Annotates Day 2, 14:30 · Lightning session G: Internationalisation
Used byrequirement DEV-INT-11glossary term Transliterated queries
Google's opening slide for serving day showed serving as the third stage after crawling and indexing, and drew Google's serving infrastructure as query understanding and retrieval leading into the index, then ranking and search features leading back to the user, for the example query 'Where to eat jamon'.
Speaker GoogleIn Day 3, 10:15 · Welcome to serving and ranking day!Evidence slide photo
Google's summary slide on query understanding told site owners not to worry about typos and plurals, which Google rewrites automatically.
“Don't worry about typos & plurals!”
Wording checked against the slide or recording
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence slide photo, transcript
Google's summary slide on query understanding said that synonyms are expanded automatically.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence slide photo, transcript
Google's summary slide on query understanding said Google's synonyms are not always language-based; on stage this was explained as words people use interchangeably counting as synonyms even when they are not synonyms linguistically.
“Google's synonyms aren't always language-based”
Wording checked against the slide or recording
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence slide photo, transcript
Google's summary slide on query understanding concluded that all this query expansion gives site owners many opportunities for their content to be found and shown.
“There are many opportunities to find and show your content.”
Wording checked against the slide or recording
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence slide photo, transcript
Google's query understanding talk set out to show how the changes Google makes to users' queries relate to what a website can provide.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
A Google speaker said that a messy query, made of a hotel location copied from a website plus a question about Italian food, was figured out by Google's systems, which came up with answers for it.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google placed query understanding in the traditional part of Search, the part where answers are looked up in an index and then presented and ranked.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google's first step in understanding almost any query is to detect its language, which tells Google roughly what content the user wants: a query in German suggests German content, a query in English English content.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-07glossary term Query understanding
Query language detection works poorly when someone searches only for a brand name, such as Facebook or Google, because the query does not show which language the user wants results in.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-03
For brand-only queries Google falls back on other information, such as the user's location and browser settings, to work out the language of the results.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-03
Google named Thai as a language that makes query understanding more complex because it does not separate words with spaces; the speaker added, hedging with 'apparently', that Thai uses spaces to separate sentences.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
After detecting the query language and separating the words, Google removes words it thinks matter little to the query, such as 'a' and 'of', known as stop words.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary terms Query understanding, Stop words
Google's query processing deliberately mirrors indexing: a query is transformed into something that can be matched against the index, and stop word removal is part of that transformation.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
For some searches the stop words are important, and Google then tries to recognise the whole phrase, stop words included, as an entity; in indexing, such words are indexed together.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Stop words
Google said the difference between words and entities can be seen in Google Trends, where a term can be searched as words or as an entity (Trends calls these a search term and a topic).
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
When a query names an entity, Google treats it as a request for that entity rather than as a collection of separate words.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Query understanding
In ranking, Google can match a query's words or its entity, and, the speaker said with a 'probably', mixes both to some degree.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google expands queries with synonyms in traditional Search; the speaker compared this to query fan-out but said it is part of traditional Search, not of the AI features.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
In Google's example, the query word 'photograph' could be expanded to 'image', 'picture' or 'photo', but one German candidate had to be dropped because the similar German word means 'photographer', so expansions are language-specific.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
According to Google's ranking teams, the synonym system is one of the most important parts of how Google handles a query.
“they say that the synonym system is one of the most important parts”
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Synonyms matter because users often phrase a query differently from the way content is indexed; synonym expansion makes the words Google looks up findable in its index.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
For technical terms, Google's synonym swapping can return either a technical page or a simplified page for the same query.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Site owners do not need to list all synonyms of a term on a page, because Google already knows the synonyms and looks for them on pages too.
“you don't have to do that SEO meme of, like, list all the synonyms on your page”
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-11
Google adds alternative words to a query automatically, which the speaker compared to a user joining words with the OR search operator.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Synonyms and siblings
Google suggested a test: a search that lists a term's synonyms joined with OR will probably return results very similar to the plain query, because Google adds the synonyms itself (part of the sentence is unclear in the recording).
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Googlers write queries in brackets, for example [spicy cheese store near me], so when reporting a search problem to a Googler, putting the query in brackets shows that it is a query someone searched.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google rewrites pretty much every query, through synonym expansion and in other ways, to understand better what the query means.
“pretty much all queries are rewritten”
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Query understanding
Internally, a short query can turn into a much longer rewritten query, because Google adds entities, synonyms and other information before looking it up.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
In Google's rewrite example, [fried chicken place in Barcelona] keeps 'fried' and 'chicken' as two words, may add an entity for fried chicken, and replaces 'place' with alternatives such as 'area', 'location' or 'restaurant'.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Some alternatives in a rewritten query make no sense, such as 'fried chicken area in Barcelona', which is harmless because few indexed pages match them.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
In a rewritten query, a place name such as Barcelona can stay a word or be swapped for an entity, possibly a more specific location.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
For a query containing 'near me', Google understands that the user wants results near their location, not pages containing the words 'near me', and rewrites the query accordingly.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Some synonyms are contextual and depend on the rest of the query: 'GM' probably means General Motors in [GM car], general manager in [GM restaurants] and genetically modified in [GM barley].
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google builds its synonyms as automatically as possible, behind the scenes, rather than having linguists catalogue every language and all possible synonyms.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google's synonyms need not be synonyms linguistically: words people use interchangeably are treated as synonyms, because the aim is to find the right content in the index.
“we don't need to be technically accurate”
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Synonyms and siblings
When Google highlights a word in its results that looks like the wrong synonym, the reason is probably that many people use the two words interchangeably.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Besides synonyms, Google detects 'siblings', words in the same category that are related but not interchangeable, such as Canon and Nikon.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Synonyms and siblings
Because Canon and Nikon are siblings rather than synonyms, a search for [Canon camera] should not show Nikon cameras.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Synonyms and siblings
Google learns synonyms and siblings from search behaviour: words people search with in the same way become synonyms, while frequent comparison queries mark words as not interchangeable (the end of the sentence is unclear in the recording).
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google does not prefer English words on pages in other languages; the idea that Google, as an American company, wants to see everything in English is wrong.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Where Google recognises an English term and its local-language equivalent as synonyms, a page does not need to contain both.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-11
To check whether Google treats two terms as the same, search for each: [Iberico ham] and [jamón ibérico] both brought up the same entity, so Google understands them as the same thing.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
When Google already matches a term's variants, there is no need to add them to pages artificially; mention another name only where visitors might not understand otherwise.
“I wouldn't artificially just stuff those variations in there.”
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-11
At the time of the talk, a search for [Spanish cured ham] brought up mainly a Wikipedia page, showing that Google does not automatically equate that phrase with Iberico ham.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
When Google does not treat a phrase as a synonym, such as 'Spanish cured ham' for Iberico ham, a page that wants to cover that phrase should use it in its text.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google generally treats spellings with and without diacritics as synonyms behind the scenes, for example a German 'ü' written as 'ü', as 'ue' or left out.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byrequirement DEV-INT-11
Google sometimes gets diacritic variants wrong, so it is worth searching to see whether Google understands a variant; if it does, pick one spelling and use it.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Spelling variants need no artificial focus: users are often fine finding one version of a word on one page and another version elsewhere.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google advised double-checking any word you are unsure about by searching for it on Google and looking at what comes up.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google's generative AI features in Search build on the traditional ways of searching, so query understanding also flows into AI Overviews and AI Mode.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Google treats fan-out queries generated by the LLM the same way as queries typed by users, so understanding how normal queries work explains fan-out queries too.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Every fan-out query goes through Google's query understanding, including synonym expansion.
Speaker John MuellerIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Query fan-out
At retrieval, Google splits the query into words, applies query understanding and expansion, and matches the words and their expansions against the posting lists.
Speaker Gary IllyesIn Day 3, 10:25 · Making sense of users' queriesEvidence transcript
Used byglossary term Retrieval
Which kinds of results Google shows for a query is decided by query understanding, which tries to predict the intent behind the query.
Speaker Gary IllyesIn Day 3, 13:25 · How Search results are bornEvidence transcript
When the query itself names the wanted result type, as 'pictures' does in 'orange pictures' (or a query asking for videos or news), Google can read the intent directly from that word.
Speaker Gary IllyesIn Day 3, 13:25 · How Search results are bornEvidence transcript
For an ambiguous query such as 'orange', Google uses historical data: it looks at how users reacted to different kinds of results for previous and similar queries and shows the result types users prefer.
Speaker Gary IllyesIn Day 3, 13:25 · How Search results are bornEvidence transcript
Google handles an image used as a search query much like a text query interpreted as an embedding: the image is broken down into vectors (embeddings) that are then searched for in the index.
Speaker Gary IllyesIn Day 3, 13:25 · How Search results are bornEvidence transcript
Google can still detect intent for image queries, even though they are searched as embeddings, the speaker said.
Speaker Gary IllyesIn Day 3, 13:25 · How Search results are bornEvidence transcript
Behind the visible attributes, small pieces of product data such as brand and UPC or other barcodes are very important for Google Shopping's query understanding, retrieval and ranking.
Speaker Alex JansenIn Day 3, 13:40 · Shopping on Search: Beyond the blue linksEvidence transcript
Google's How Search Works page says understanding a query ranges from recognising and correcting spelling mistakes to a synonym system that finds relevant documents without the exact words searched, such as 'adjust laptop brightness' for 'change laptop brightness'.
Publisher Google Search (How Search Works)Annotates Day 3, 10:25 · Making sense of users' queries
Google's How Search Works page says the language models it builds to match a query's few words to the most useful content took over five years to develop and significantly improve results in over 30% of searches across languages.
Publisher Google Search (How Search Works)Annotates Day 3, 10:25 · Making sense of users' queries
Google's SEO Starter Guide advises anticipating the different words readers search with (some search for 'charcuterie', others for 'cheese board') but not worrying about every variation, because Google's language matching systems relate pages to queries without the exact terms.
Publisher Google Search CentralAnnotates Day 3, 10:25 · Making sense of users' queries
Google's Search Help says the language of search results is chosen from the language of the query, the user's Google language setting, the device's languages and the user's location, which helps when a word such as 'taxi' is the same in several languages.
Publisher Google Search HelpAnnotates Day 3, 10:25 · Making sense of users' queries
A 2006 Search Central blog post says Google considers pages with and without accents for a query word (México and Mexico), and that which accented characters count as equivalent depends on the searcher's interface language.
Publisher Search Central blog (1 September 2006)Annotates Day 3, 10:25 · Making sense of users' queries
Google's How Search Works pages say Google uses aggregated and anonymized interaction data to assess whether search results are relevant to queries, turning that data into signals for its machine-learned systems.
Publisher Google Search (How Search Works)Annotates Day 3, 13:25 · How Search results are born
Write each page in the words its audience uses and drop synonym lists added for search engines: Google adds synonyms at query time, and blocks of keyword variants read as keyword stuffing, which Google's spam policies prohibit.
Author Ibrahim AnjroAnnotates Day 3, 10:25 · Making sense of users' queries
Used byrequirement DEV-INT-11
Before writing, search for each variant of a key term: if the variants bring up the same entity or near-identical results, use the term your audience uses; if one variant returns unrelated results, write that variant on the page.
Author Ibrahim AnjroAnnotates Day 3, 10:25 · Making sense of users' queries
For audiences that type the same words in two scripts or spellings (Hindi in Devanagari and in Latin letters, German with and without umlauts), check which forms appear in Search Console's queries and use those forms in headings and key text.
Author Ibrahim AnjroAnnotates Day 3, 10:25 · Making sense of users' queries
Used byrequirement DEV-INT-11
Google said MUM helps it understand the context of the words in a query: a search for hiking shoes that mentions Mount Everest rather than Kilimanjaro gets results suited to that climb.
MUM (Multitask Unified Model) understands information across text, images, audio and video, and processes information in more than 75 languages.
Users of non-Latin-script languages do not always search in their own script: the same Persian query may be typed in Persian script or in Latin letters, with the same intent and the same expected results.
Google usually understands non-English words typed 'in English' (probably meaning romanised, Latin-letter spellings, which the speaker did not spell out); the speaker said this belongs to query interpretation, a Day 3 (serving) subject.
Google's first step in understanding almost any query is to detect its language, which tells Google roughly what content the user wants: a query in German suggests German content, a query in English English content.
In ranking, country and language signals help Google serve users the right content for their country and language.
Query language detection works poorly when someone searches only for a brand name, such as Facebook or Google, because the query does not show which language the user wants results in.
Serving starts with interpreting the query, which includes cleaning it up, detecting its language and expanding it.
Google named Thai as a language that makes query understanding more complex because it does not separate words with spaces; the speaker added, hedging with 'apparently', that Thai uses spaces to separate sentences.
Text in languages written without spaces, such as Thai and Chinese, would end up in the index as long strings that might never be searched for, so Google segments it into words with statistical models built from other web content in that language.
Google's query processing deliberately mirrors indexing: a query is transformed into something that can be matched against the index, and stop word removal is part of that transformation.
For languages written without spaces, such as Thai and Chinese, Google uses exactly the same word segmentation when indexing a page as when interpreting the user's query, because otherwise the query could not be matched against the index.
Google's query processing deliberately mirrors indexing: a query is transformed into something that can be matched against the index, and stop word removal is part of that transformation.
A search query is broken into words with the same segmenter or tokenizer that Google used to build the index.
Google said the difference between words and entities can be seen in Google Trends, where a term can be searched as words or as an entity (Trends calls these a search term and a topic).
The Google Trends Explore page, which Google called the heart of Trends, shows search interest in a query or topic and how it changes over time.
Google generally treats spellings with and without diacritics as synonyms behind the scenes, for example a German 'ü' written as 'ü', as 'ue' or left out.
Google usually understands a query word whether it is written with or without diacritics (accents).
Google treats fan-out queries generated by the LLM the same way as queries typed by users, so understanding how normal queries work explains fan-out queries too.
Google said the Gemini model lets Search understand the user's intent, and query fan-out then adds further queries to the first one to enrich the quality of the answer.
Google treats fan-out queries generated by the LLM the same way as queries typed by users, so understanding how normal queries work explains fan-out queries too.
Google said that when AI Overviews or AI Mode run a query fan-out, the generated queries are sent to Google's Search index and documents come back with their snippets, which then feed the AI-generated answer (part of this passage is unclear in the recording).
At retrieval, Google splits the query into words, applies query understanding and expansion, and matches the words and their expansions against the posting lists.
In posting-list retrieval, the posting lists of the query's words are intersected, which yields an unranked list of candidate URLs.
At retrieval, Google splits the query into words, applies query understanding and expansion, and matches the words and their expansions against the posting lists.
At retrieval, Google looks up the posting lists of the query words that are actually important rather than of every word in the query.
Which kinds of results Google shows for a query is decided by query understanding, which tries to predict the intent behind the query.
Google placed query understanding in the traditional part of Search, the part where answers are looked up in an index and then presented and ranked.
Google handles an image used as a search query much like a text query interpreted as an embedding: the image is broken down into vectors (embeddings) that are then searched for in the index.
Besides posting lists, Google can retrieve documents through vector embeddings: parts of documents are associated with embeddings, which form a vector space used for retrieval.
Google's generative AI features in Search build on the traditional ways of searching, so query understanding also flows into AI Overviews and AI Mode.
Three reasons were given: generative AI features are built directly on the core ranking systems, query fan-out expands the original query to find related information, and generative AI features highlight content indexed by Google Search.
Google handles an image used as a search query much like a text query interpreted as an embedding: the image is broken down into vectors (embeddings) that are then searched for in the index.
For visual search, Google breaks an image down into vectors, sends the vectors to the index and returns results based on them.
Google's serving diagram shows the query passing through query understanding and retrieval to the index, then back through ranking and search features to the user, with the Search Features step highlighted (example query 'Where to eat orange').
Google's opening slide for serving day showed serving as the third stage after crawling and indexing, and drew Google's serving infrastructure as query understanding and retrieval leading into the index, then ranking and search features leading back to the user, for the example query 'Where to eat jamon'.
Rests on 10 claims, 8 of them in this topic
Make hreflang complete: a self-reference, return links and fully qualified URLs
Rests on 12 claims, 2 of them in this topic
Make each page's language unambiguous by translating the template with the main content
Rests on 12 claims, 1 of them in this topic