Day 3 of Google Search Central Live Deep Dive Europe 2026 had a session that I had been looking forward to: Google walking through what happens to a query before any ranking takes place.
We spend a lot of time thinking about how Google understands our pages. This session was about the other side: how Google understands what the user typed. These are my notes, along with the slides I managed to capture from my seat.
Two things were set up at the start. First, when Googlers talk about queries, they write them in square brackets, like [swiss cheese store near me]. I'm following the same convention in this post. Second, and more important: almost all queries are internally rewritten before they're processed, based on what's assumed to be meant by the user.
So the query you see in Search Console is rarely the query that Google actually runs. The rest of the session was about what that rewriting looks like.
Detecting the language
The first step is working out what language the query is in. Everything after this depends on it, as stopwords, synonyms and spelling are all language-specific.
The interesting part is when the query itself gives nothing away. The example given was [facebook]. It's the same word whether you are searching in English, Spanish or German, so Google can't detect the language from the query alone and needs an extra level of signals to decide.
This matters more than it sounds. A lot of navigational and brand queries are language-neutral, and those are exactly the queries where international sites tend to see the wrong language version showing up.
Cleaning up stopwords
Next, Google removes the words that don't carry meaning. In [a nice photograph of fried chicken], "a" and "of" are dropped. The German version on the slide, [ein schönes Foto von Backhähnchen], loses "von" in the same way.
This is also why language detection has to come first. You can't know which words are stopwords until you know which language you are reading.
Recognising the entity
Stopword removal has an obvious problem, and the next slide showed it. In [The Lord of the rings], "the", "of" and "the" are all stopwords. Strip them out and you are left with [lord rings], which is not what anyone is looking for.
So Google needs to recognise that the whole phrase is an entity and treat it as one unit, rather than a string of separate words.
Tip: check entities in Google Trends
A practical tip from the session: use Google Trends to see whether something is recognised as an entity. When you type a term, Trends will offer it either as a plain search term or as a topic (for example a film, a company or a food). If your brand or product shows up as a topic, that is a good sign Google treats it as an entity.
Expansion: Google's synonyms system
Once the query is cleaned up, it gets expanded. [photograph of fried chicken] also becomes image, picture and photo of fried chicken. In German, Foto becomes Bild and Fotos.
The reason given is simple: user vocabulary is not the same as document vocabulary. People don't search with the words that site owners write with, so the system tries to bridge the gap by automatically adding alternative words.
Two lines on that slide stood out to me:
- It is "one of Google Search's most important ranking components".
- It is similar to using "OR", but the synonyms are usually less important than the original terms.
The second point is the one I'd keep in mind. Google can match your page through a synonym, but the word the user actually typed still carries more weight.
What a rewritten query looks like
The example rewrite on the slide took [fried chicken place in barcelona] and turned it into this:
[fried chicken place in barcelona]
=>
[Fried
Chicken
(place OR area OR location OR restaurant)
in
barcelona]Notice that "restaurant" is in the list. Nobody would call "place" and "restaurant" synonyms in a dictionary sense, but in this query that is clearly what the user means.
Synonyms depend on the other words in the query
Synonyms are contextual. The same two letters expand differently depending on what sits next to them:
| Query | "gm" is expanded to |
|---|---|
| [gm car] | general motors |
| [gm restaurant] | general manager |
| [gm barley] | genetically modified |
Not the same as actual English synonyms
Google was clear that its "synonyms" are not linguistic synonyms. They are designed for good search results, they are hidden behind the scenes, and it is unimportant whether a human reader would consider them synonyms at all.
For me, this is a good reminder to stop treating synonym coverage as a thesaurus exercise. What matters is whether two terms lead to the same results being useful, not whether they mean the same thing on paper.
Siblings: similar roles, but not interchangeable
This was my favourite part of the session. Some words serve similar roles but should never be swapped: Canon and Nikon, United and Continental, Beatles and Stones, Godzilla and King Kong. Google calls these siblings.
How does Google learn them? From the way people compare things. It looks for [X vs Y] queries in the logs and uses them as a negative signal for X ⇒ Y and Y ⇒ X synonyms. If people are comparing two things, they are obviously not the same thing.
It's a neat way of looking at comparison queries. If people regularly search for your brand vs a competitor, Google has a clear signal that someone searching for one of you does not want the other.
Non-English words, variations and diacritics
Two audience-style questions were answered with the same example. Does Google understand non-English words typed in English, such as Jamon Iberico, Iberian Ham or Spanish Cured Ham? And how does it deal with diacritics, Jamón Ibérico vs jamon iberico?
The answer was two screenshots side by side. Searching [iberico ham] and searching [jamon iberico] both return the same "Jamón ibérico" food entity at the top of the results, with the accents that neither query included.
The guidance for handling non-English product names, variations and diacritics came down to three points:
- Focus on users. Write in your users' language.
- Don't overthink everything. Google tries to match users' queries with documents.
- Check terms that are not well known. Explain or elaborate on them where it helps.
The third point is the practical one. For a well-known term like jamón ibérico, Google has it covered. For a niche product name or a local term that hasn't made it into the mainstream, I wouldn't assume the same, and a short explanation on the page costs nothing.
Query fan-out does even more expansion
The session then moved on to AI features. The diagram for how LLM features with grounding generally work: the query goes to both the search engine and the LLM, the LLM generates fan-out queries that go back to the search engine, and the results come back to the LLM to produce answers with links.
What the slide said about query fan-out:
- It expands the initial query with distinct additional queries for better coverage.
- Each new query goes through query understanding. It can't be skipped.
- Fan-out queries are not recorded in Search Console, as they are not from users.
- You can see similar things in some AI systems, such as the Gemini app or some APIs.
- Every system does this differently.
Two takeaways for me here. One, fan-out is not a separate system that bypasses everything above. Every fan-out query still gets its language detected, its stopwords removed and its synonyms expanded. Two, we now have it confirmed that Search Console will not show these queries, so anyone trying to reverse-engineer fan-out from GSC data is looking in the wrong place.
After the query: language, country and quality
Towards the end there was a simplified view of what happens once the query is understood. Each candidate URL carries a language, a country and a quality level, and arrows showed some of those values being pushed up or down for a given query.
My reading of the slide: for this example query, English pages and pages for GB move up, pages in other languages move down, and high quality moves up regardless. That ties straight back to Step 1. If the detected language of the query decides which pages are promoted, getting it right is not a small detail.
Google's own summary
The closing slide summarised query understanding in six points:
- Don't worry about typos and plurals!
- Synonyms are expanded.
- Google's synonyms aren't always language-based.
- Some languages don't use spaces!?!
- Query fan-out does even more expansion.
- There are many opportunities to find and show your content.
My take
None of this is a new ranking trick, and I think that was the point. Most of what SEOs still worry about at the keyword level, such as singular vs plural, accents vs no accents, or one variation per page, is handled before ranking even starts.
What I'm taking back to my own work:
- Write with the words your users actually use. Synonyms help you get matched, but the original term is still weighted more.
- Stop creating separate pages for close variations of the same term.
- Check in Google Trends whether your brand and products are recognised as entities.
- Explain niche or local terms on the page instead of assuming Google knows them.
- Don't expect fan-out queries in Search Console. They are not there.
The last line of the summary slide is the one I'd end on too: with rewriting, synonyms and fan-out, there are more ways than ever for a query to reach your content.
As Google now cracking down on rank tracking tool, it might be harder for SEO to do keyword research with only third-party data over time. So I would start looking internally to see how your target audience are actually searching (as matching terms still carry more weight). There are a lot of customer data that we can leverage to help us create content worth finding.
