AI systems already inside Search

Day 1: Crawling 19

Shown on screen 5

Said on stage 9

What Google's documentation says 3

Analysis by the author 2

Day 2: Indexing 13

Shown on screen 1

Said on stage 11

What Google's documentation says 1

Day 3: Serving: Ranking, Search Console, and Performance 18

Shown on screen 3

Said on stage 14

Analysis by the author 1

Across days and sessions 30

1. Stage D1-C199 Day 1 · What's new in the world of Search

   Gary Illyes said searching by image or by video, as Search offers it today, was not possible before transformers were invented.

   extends
   Stage D1-C161 Day 1 · Welcome and opening keynotes

   Google said it created the Transformer architecture (the paper 'Attention Is All You Need', the T in ChatGPT) and published it to move the industry forward, competitors included.

2. Stage D1-C205 Day 1 · How Search works and where's AI?

   Signals calculated for a page during indexing are stored in the index and used both to decide whether the page gets indexed and, later, for ranking.

   extends
   Slide D1-C037 Day 1 · How Search works and where's AI?

   For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

3. Stage D1-C209 Day 1 · How Search works and where's AI?

   Google launched the 'Did you mean' feature around 2001-2002 using a statistical model, which Gary Illyes counted as AI because it is a form of machine learning.

   extends
   Slide D1-C041 Day 1 · How Search works and where's AI?

   Statistical models have been used at Google for over 20 years, for catching spam and originally for the 'Did you mean' feature.

4. Stage D1-C211 Day 1 · How Search works and where's AI?

   Index selection runs after signals are collected and duplicates are dropped, and decides what goes into Google's index, which is big but not limitless.

   extends
   Slide D1-C037 Day 1 · How Search works and where's AI?

   For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

5. Stage D1-C217 Day 1 · How Search works and where's AI?

   Google began talking publicly about its use of AI in Search around 2015-2016, and RankBrain was the first such system it publicised widely.

   extends
   Slide D1-C043 Day 1 · How Search works and where's AI?

   RankBrain is used in serving to interpret the intent behind queries, especially new or unusual ones, and match them to relevant results.

6. Stage D1-C218 Day 1 · How Search works and where's AI?

   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.

   extends
   Slide D1-C044 Day 1 · How Search works and where's AI?

   MUM (Multitask Unified Model) understands information across text, images, audio and video, and processes information in more than 75 languages.

7. Stage D2-C338 Day 2 · Understanding what's on a page

   Google detects soft 404s with a language model, described as something like BERT, that is trained to understand the structure and layout of a page as well as its language, instead of reading the page as one flat wall of text.

   extends
   Slide D1-C042 Day 1 · How Search works and where's AI?

   BERT is used in indexing to understand each word in the context of the whole sentence rather than one word at a time.

8. Stage D2-C344 Day 2 · Handling web duplication

   Google deduplicates pages because many sites have very many pages and Google's index does not have room for everything.

   extends
   Slide D1-C037 Day 1 · How Search works and where's AI?

   For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

9. Docs D2-C462 Day 2 · What is Structured Data and why we need it on the internet.

   Google's guidance on generative AI content warns that AI output can contain hallucinations and says AI-generated metadata, including structured data and image alt text, should be fact-checked before publishing and the markup validated.

   extends
   Stage D1-C250 Day 1 · Lightning session A: Automation and AI

   In a community demo, a fix loop gave a large LLM (Claude Opus) the current markup, the errors Google's test reported and Google's documentation, had it write new JSON-LD and re-ran the test, with at most three attempts; the demo's page passed on the second.

10. Stage D2-C648 Day 2 · Calculating (some) signals

    Among the many signals Google calculates during indexing, the ones singled out as having large effects on search results were country, language, freshness, SafeSearch and spam.

    extends
    Slide D1-C037 Day 1 · How Search works and where's AI?

    For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

11. Stage D2-C668 Day 2 · Calculating (some) signals

    Google uses more and more AI to detect spam, and Google's testing shows that this AI-based detection is highly accurate.

    extends
    Slide D1-C041 Day 1 · How Search works and where's AI?

    Statistical models have been used at Google for over 20 years, for catching spam and originally for the 'Did you mean' feature.

12. Stage D2-C669 Day 2 · Calculating (some) signals

    SpamBrain, Google's AI-based spam detection system, is nowadays built on Gemini and fine-tuned specifically for finding spam.

    extends
    Slide D1-C039 Day 1 · How Search works and where's AI?

    Gemini is not part of Search, but it uses crawlers for data, shares some technologies such as tokenization and deduping, and grounds on the Search index.

13. Stage D2-C669 Day 2 · Calculating (some) signals

    SpamBrain, Google's AI-based spam detection system, is nowadays built on Gemini and fine-tuned specifically for finding spam.

    extends
    Slide D1-C041 Day 1 · How Search works and where's AI?

    Statistical models have been used at Google for over 20 years, for catching spam and originally for the 'Did you mean' feature.

14. Stage D2-C684 Day 2 · Deciding what goes in the index?

    Index selection is a predictive AI system that relies heavily on machine learning.

    extends
    Slide D1-C037 Day 1 · How Search works and where's AI?

    For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

15. Stage D2-C722 Day 2 · How does the index look like?

    Each document in Google's index has pretty much all the signals calculated for it attached, for example quality signals plus the page's country and language, according to an illustration the speaker called an approximation of the real structure.

    extends
    Slide D1-C037 Day 1 · How Search works and where's AI?

    For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

16. Stage D2-C741 Day 2 · How does the index look like?

    In embedding-based retrieval, the distance between the embeddings of documents and the embedding of the user's query decides which documents are returned.

    extends
    Docs D1-C129 Day 1 · How Search works and where's AI?

    Google's guide says creating separate content for every variation of how people might search, including fan-out queries, primarily to manipulate rankings or AI responses violates its scaled content abuse policy. It adds that its AI systems can understand a page's relevance even without an exact match to the query.

17. Stage D2-C947 Day 2 · Lightning session F: Media

    In a community speaker's case, an LLM asked to write alt text for a product image of a veterinary anxiety medicine for cats described only what it could see, a cat and a veterinarian, and dropped the stress, anxiety and medical-treatment intent of the page.

    extends
    Docs D2-C462 Day 2 · What is Structured Data and why we need it on the internet.

    Google's guidance on generative AI content warns that AI output can contain hallucinations and says AI-generated metadata, including structured data and image alt text, should be fact-checked before publishing and the markup validated.

18. Stage D2-C952 Day 2 · Lightning session F: Media

    A community speaker warned that articles still recommend rewriting alt text with AI, called AI a tool, and said that adopting such new AI implementations without applying existing SEO knowledge can damage ranking performance.

    extends
    Docs D2-C462 Day 2 · What is Structured Data and why we need it on the internet.

    Google's guidance on generative AI content warns that AI output can contain hallucinations and says AI-generated metadata, including structured data and image alt text, should be fact-checked before publishing and the markup validated.

19. Stage D3-C077 Day 3 · Making sense of users' queries

    The first condition for retrieving a document is that the query's words, or its concepts in the case of vectors or embeddings, are in the document or related to it.

    extends
    Stage D2-C740 Day 2 · How does the index look like?

    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.

20. Analysis D3-C135 Day 3 · How Google thinks about Quality

    The quality talk's slide listed MUM among Google's ranking systems, while Google's ranking systems guide says MUM is not currently used for general ranking in Search; read the slide as a list of systems Google runs, not as proof that each one ranks every query.

    extends
    Docs D1-C130 Day 1 · How Search works and where's AI?

    Google's ranking systems guide says MUM is not currently used for general ranking in Search, only for specific applications such as COVID-19 vaccine searches and featured snippet callouts.

21. Stage D3-C205 Day 3 · How Google thinks about Quality

    Google's quality talk said AI has fundamentally changed how Google builds spam updates, letting it evaluate many more candidates and drastically increasing its velocity, so it launches faster with more impact.

    extends
    Stage D2-C668 Day 2 · Calculating (some) signals

    Google uses more and more AI to detect spam, and Google's testing shows that this AI-based detection is highly accurate.

22. Stage D3-C206 Day 3 · How Google thinks about Quality

    Google's quality talk said AI lets Google catch new types of spam and catch more of it.

    extends
    Stage D2-C668 Day 2 · Calculating (some) signals

    Google uses more and more AI to detect spam, and Google's testing shows that this AI-based detection is highly accurate.

23. Stage D3-C311 Day 3 · How Search results are born

    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.

    extends
    Stage D2-C740 Day 2 · How does the index look like?

    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.

24. Slide D3-C477 Day 3 · Lightning session L: Understanding SERPs and your users

    The community workflow's third step feeds the exported Search Console dataset into the agency's LLM agent, which queries the dataset and returns the evidence rather than a sample.

    extends
    Stage D1-C256 Day 1 · Lightning session A: Automation and AI

    An agency's automated monthly SEO report follows four simple steps: the data comes in, a script processes it, AI summarises it and the result goes onto a dashboard.

25. Stage D3-C680 Day 3 · Wrapping all up: AI, Search, and making sense of everything.

    Gary Illyes said machine learning is about 50 years old and that Google has been using it for 'probably 30 years', starting with statistical models that made the 'Did you mean' feature possible.

    extends
    Slide D1-C041 Day 1 · How Search works and where's AI?

    Statistical models have been used at Google for over 20 years, for catching spam and originally for the 'Did you mean' feature.

26. Stage D3-C684 Day 3 · Wrapping all up: AI, Search, and making sense of everything.

    Predictive language models have been used a lot in Search, and BERT is one: technically, in the purest sense, an LLM.

    extends
    Slide D1-C042 Day 1 · How Search works and where's AI?

    BERT is used in indexing to understand each word in the context of the whole sentence rather than one word at a time.

27. Stage D3-C686 Day 3 · Wrapping all up: AI, Search, and making sense of everything.

    Hallucinations can happen with any AI model, and with current training methods there is no way to get rid of them.

    extends
    Docs D2-C462 Day 2 · What is Structured Data and why we need it on the internet.

    Google's guidance on generative AI content warns that AI output can contain hallucinations and says AI-generated metadata, including structured data and image alt text, should be fact-checked before publishing and the markup validated.

28. Stage D3-C688 Day 3 · Wrapping all up: AI, Search, and making sense of everything.

    Generative models, including image diffusion models, make things up when they lack information or because of issues in their training.

    extends
    Stage D2-C939 Day 2 · Using images to your advantage and Engaging Search users with videos

    The diffusion models that generate images were built to generate images, not text, so they are typically poor at rendering text inside an image.

29. Stage D2-C680 Day 2 · Deciding what goes in the index?

    Google's index is immense but finite, so Google cannot index every URL it finds on a web with a practically infinite number of URLs.

    repeats
    Slide D1-C037 Day 1 · How Search works and where's AI?

    For classic Search, crawling means Googlebot, scheduling and robots.txt, with AI used in parts such as scheduling. Indexing is one big but not limitless index that calculates signals and understands more than words, with AI such as BERT. Serving uses hundreds of signals tailored to the moment, with AI such as RankBrain.

30. Stage D3-C680 Day 3 · Wrapping all up: AI, Search, and making sense of everything.

    Gary Illyes said machine learning is about 50 years old and that Google has been using it for 'probably 30 years', starting with statistical models that made the 'Did you mean' feature possible.

    repeats
    Stage D1-C209 Day 1 · How Search works and where's AI?

    Google launched the 'Did you mean' feature around 2001-2002 using a statistical model, which Gary Illyes counted as AI because it is a form of machine learning.

Built on these claims 3

Developer requirements 3

• MayDocumentedDEV-IMG-08

  Keep the IPTC and C2PA metadata that marks AI-generated images

  Rests on 4 claims, 1 of them in this topic

• MustDocumentedDEV-SDA-11

  Validate markup in the Rich Results Test before it reaches templates, and monitor it after release

  Rests on 8 claims, 1 of them in this topic

• ShouldDocumentedDEV-SPM-05

  Publish AI-generated text, images and markup only after human review, and show only real authors

  Rests on 11 claims, 1 of them in this topic

Sources 12

• Google I/O 2026: Sundar Pichai’s opening keynote Google blog (19 May 2026) · checked 4 October 2026
• Transformer: A Novel Neural Network Architecture for Language Understanding Google Research blog (31 August 2017) · checked 3 October 2026
• A new era of intelligence with Gemini 3 Google blog (18 November 2025) · checked 4 October 2026
• 25 biggest moments in Search, from helpful images to AI Google blog (6 September 2023) · checked 3 October 2026
• In-depth guide to how Google Search works Google Search Central · checked 3 October 2026
• A guide to Google Search ranking systems Google Search Central · checked 3 October 2026
• MUM: A new AI milestone for understanding information Google blog (18 May 2021) · checked 2 October 2026
• How AI powers great search results Google blog (3 February 2022) · checked 2 October 2026
• Google Search's guidance on using generative AI content on your website Google Search Central · checked 3 October 2026
• Google Search spam updates and your site Google Search Central · checked 3 October 2026
• How we fought Search spam on Google in 2021 Search Central blog (21 April 2022) · checked 3 October 2026
• Detecting spam to bring you relevant and reliable results Google Search (How Search Works) · checked 3 October 2026