Knowledge base v2.13.0 · Community edition · data through 2 October 2026
Day 1 · Wednesday 30 September 2026 · 13:10
Hosted by Gary Illyes and Cherry Prommawin. Seven talks: AI automation in agency SEO work (James Powley, SEO Director, Blue Array: self-introduction and hand-over); LLM errors on search data (Rafael Kovashikawa, AI Engineer & SEO Specialist, FUSE: hand-over 'Raphael', the only Rafael on the official speaker list); an LLM fix loop for structured data (a live demo with Google's Rich Results Test); automating monthly SEO reports (Kira Breuer, SEO Manager, Greven Medien GmbH & Co. KG: hand-over 'Kiran' in the speech-to-text and her self-description as an SEO manager at a local agency in Cologne, matched to the only Kira on the official speaker list, whose job and city agree); preparing sites for humans and AI agents (Carlos Ortega, Freelance SEO Consultant: the next speaker's back-reference 'as we were just shown by Carlos', the only Carlos on the official speaker list); GEO and its metrics (Thiago Pojda, SEO Director & Entrepreneur, SIXT SE: the host's hand-over and his self-introduction); AI search as a funnel. The other speakers were not named. Transcript from a second attendee recording and an audio recording; no slides photographed.
Google received close to 200 submissions for the community lightning talks of the Deep Dive, reviewed them over about a month, and gave each selected community speaker seven minutes on stage.
Speaker not identifiedEvidence transcript
Google's hosts said the lightning talks exist because the SEO community holds a lot of knowledge, some of it untapped, and highlighting the people who have it serves the community best.
Speaker not identifiedEvidence transcript
A community speaker's approach to AI in agency SEO work is to use it where it meaningfully elevates and enriches existing work rather than replacing it, keeping full strategic oversight and control of the output.
“Human judgment and strategy is the default, and AI simply pushes it at scale.”
Wording checked against the slide or recording
Speaker James PowleyEvidence transcript
Asking AI to produce whole SEO deliverables from scratch, such as creating content or auditing a website, is where people get stuck; break a task into its elements and decide where AI can and should play a part, a community speaker advised.
Speaker James PowleyEvidence transcript
A community speaker designs AI-assisted deliverables like a flowchart, keeping guardrails in place and keeping the prompts succinct and measured, so the team controls the process end to end.
Speaker James PowleyEvidence transcript
In an agency's competitor keyword analysis, a deliverable defined before AI, the team set the strategy buckets and the keyword scoring metrics itself and used AI only to pull in the data, never to make a judgment.
Speaker James PowleyEvidence transcript
With AI pulling in the data, a competitor keyword analysis covering 26 competitors and 79 data points per keyword went from three days of work to a few hours, a community speaker reported for his agency.
Speaker James PowleyEvidence transcript
For a content audit of a client blog with over 3,000 posts, an agency defined the parameters, scoring and weightings manually and had AI apply them, producing posts grouped by category, possible cannibalisation flags, a priority list and a list of posts to check by hand.
Speaker James PowleyEvidence transcript
The key lesson of a community talk on AI automation: AI should not decide what good looks like; people set the rules for what is good and AI applies them.
“it didn't decide what good looked like. We did that.”
Wording checked against the slide or recording
Speaker James PowleyEvidence transcript
In a post-migration analysis of over 200,000 lines of search traffic data and more than 45,000 pages, where many URLs had been merged into one, an agency set the rules for what counted as the same page and used AI to match old and new addresses and roll the traffic up per page.
Speaker James PowleyEvidence transcript
A community speaker builds SEO recommendations without AI and then uses AI to turn them into mock-ups and visuals, which help win buy-in from clients and internal teams.
Speaker James PowleyEvidence transcript
An agency built its own AI visibility tracking tool so it could control the methodology: the team wrote pages of flowchart-style prompts (if this happens, do that) and tested manually, and AI only generated the code.
Speaker James PowleyEvidence transcript
A community speaker's example of an LLM error on Search Console data: a chatbot called an average position that went from 8 to 9 an improvement, although position 1 is the top, so the move from 8 to 9 is a decline.
Speaker Rafael KovashikawaEvidence transcript
How often an LLM gives a wrong answer about data depends on the model, but how often the mistake is noticed depends on the person using it, a community speaker said.
“how often can you notice the mistake? And that depends on you.”
Wording checked against the slide or recording
Speaker Rafael KovashikawaEvidence transcript
In agency marketing-data pipelines, as in financial forecasting, one misplaced figure travels downstream and compounds into later analysis and decisions, which is why LLM errors on search data must be caught early, a community speaker warned.
Speaker Rafael KovashikawaEvidence transcript
A community speaker reported that a code refactor of his company's AI analysis product left data fetching correct but made the analysis shallow while the answers still read well, with the same model, so polished output is no proof of correct analysis.
Speaker Rafael KovashikawaEvidence transcript
The first rule for using LLMs on search data, according to a community speaker, is never to let the model check its own output.
“not let the model grade its own homework”
Wording checked against the slide or recording
Speaker Rafael KovashikawaEvidence transcript
A community speaker's first guardrail for LLM answers about data is a receipt: every figure the model gives must be traceable to the data request behind it, so the answer can be verified.
“If there is no receipt, there is no reimbursement.”
Wording checked against the slide or recording
Speaker Rafael KovashikawaEvidence transcript
In a community speaker's verification system, each claim in an LLM's data answer is marked verified, flagged (a hallucination), or skipped or not found.
Speaker Rafael KovashikawaEvidence transcript
Claims that cannot be verified go to a judge, models from other vendors that score them; a failing claim is struck through with the reason shown, and the system annotates and proposes corrections rather than rewriting the answer, a community speaker described.
Speaker Rafael KovashikawaEvidence transcript
Give an LLM your metric definitions as fixed, protected context, for example that position 1 is the top, and leave nothing for the model to guess, a community speaker advised.
Speaker Rafael KovashikawaEvidence transcript
Before using AI-generated numbers, for example in a client meeting, always ask the AI for the receipts and the reason behind every number, because AI hallucinates; LLMs present their reasoning as a forward chain, so the user has to check it backwards, a community speaker said.
Speaker Rafael KovashikawaEvidence transcript
A community demo showed a script that opens Chrome with a saved session, pastes a URL into Google's Rich Results Test, waits about 15 seconds, then screenshots and saves the result and retries on failure (the demo's recording is largely unintelligible; the tool name is a best reading).
Speaker not identifiedEvidence transcript
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.
Speaker not identifiedEvidence transcript
A community demo treated warnings in Google's structured-data test as not critical, in an example result of 22 items with two errors and two warnings.
Speaker not identifiedEvidence transcript
Used byrequirement DEV-SDA-11
A community demo matched model size to the task: a small, cheap model (Claude Haiku) chose keywords from a page's URL, title, H1, description and schema types, while a large, expensive model (Claude Opus) fixed code and wrote the summary.
Speaker not identifiedEvidence transcript
At a community speaker's agency, monthly client reports took each SEO manager one full workday every month, contained technical errors and were not read by anyone, which led the team to automate them.
Speaker Kira BreuerEvidence transcript
To automate SEO reporting with AI-assisted coding, a community speaker advised against aiming for an app or a perfect dashboard: pick one repetitive task, define the smallest useful output, ask the AI what is possible, then build and debug one step at a time.
Speaker Kira BreuerEvidence transcript
A community speaker who is not a developer automated monthly SEO reports with Python in Visual Studio Code, using Claude and ChatGPT as coding partners and Google Cloud for access to the Search Console API.
Speaker Kira BreuerEvidence transcript
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.
Speaker Kira BreuerEvidence transcript
Before writing code to pull Search Console data for one property, check which data you actually need, what the API can provide and what the API costs, a community speaker advised.
Speaker Kira BreuerEvidence transcript
Pulling Search Console data through the API required setting up a Google Cloud account to get API access, a community speaker said.
Speaker Kira BreuerEvidence transcript
A community speaker gave the AI a detailed description of the tech stack and requirements and started tiny, with one script, one client and one output, before having the AI write the code.
Speaker Kira BreuerEvidence transcript
When debugging AI-written code with AI, always read the code and understand what went wrong, so you can fix it yourself next time, a community speaker advised.
“Don't trust it blindly, and try to understand what is happening in your code”
Wording checked against the slide or recording
Speaker Kira BreuerEvidence transcript
A good AI summary in an automated SEO report needs cleaned, good-quality data, a clear definition of what is wanted and a detailed, explicit prompt, a community speaker said.
Speaker Kira BreuerEvidence transcript
The resulting monthly report showed Search Console data on one side and an AI-written summary of what happened that month on the other; an in-house AI specialist built the dashboard from the SEO manager's code.
Speaker Kira BreuerEvidence transcript
To start automating SEO work with AI, a community speaker said you need no expertise, only one painful task, a clear outcome and the patience to iterate.
“you don't need to be an expert; you just need to try.”
Wording checked against the slide or recording
Speaker Kira BreuerEvidence transcript
A community speaker framed the web as now visited by both humans and AI agents, so sites have to be prepared for both.
Speaker Carlos OrtegaEvidence transcript
A community speaker said llms.txt files are not necessary, citing a third-party study from summer 2026 (heard as Ahrefs') that looked at almost 40,000 sites over one month: 97% had no AI agent hits on the file, and those that had any got about two hits a month.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-02
A community speaker advised against publishing markdown copies of HTML pages for AI agents: the copy is a duplicate (which the speaker also called a possible source of cloaking, an uncertain word in the recordings), and the models are trained to read HTML, CSS and JavaScript.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-02
A community speaker pointed out that, according to Google's documentation, user-triggered fetchers, which fetch a page because a user asked for it, generally ignore robots.txt rules.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-05glossary term User-triggered fetchers
A community speaker said AI agents understand a page through a combination of three inputs: a screenshot, the DOM (the HTML plus the changes rendered by JavaScript) and the accessibility tree.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-HTM-04
A community speaker said layout shifts, such as a button or image popping in after the first load, annoy users and may confuse AI agents, and recommended watching Cumulative Layout Shift (CLS), the Core Web Vitals metric for visual stability.
Speaker Carlos OrtegaEvidence transcript
Used byglossary term Cumulative Layout Shift (CLS)
A community speaker said schema markup helps AI agents interpret a page: on a product page, marking up which number is the price saves the agent from guessing.
Speaker Carlos OrtegaEvidence transcript
A community speaker advised keeping headings inside the main content in a logical hierarchy (title, section subtitles, subsections) so agents can follow its structure, avoiding several H1 elements and skipped levels such as an H3 followed directly by an H5.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-HTM-04
A community speaker said a block of content without semantic HTML or landmarks is just a div whose purpose an agent cannot tell, and recommended landmark elements (header, nav, main, article for independent sections, footer) plus p and h1-h6 for text.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-HTM-04
A community speaker described ARIA (Accessible Rich Internet Applications) as a set of attributes, not a programming language, that adds accessibility information to HTML: a div used as an 'add to favourites' button can get role=button, an aria-label and aria-pressed set to true or false.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-HTM-04glossary term ARIA
A community speaker warned that no ARIA is better than bad ARIA: wrong or confusing ARIA attributes do more harm than leaving them out.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-HTM-04glossary term ARIA
A community speaker presented WebMCP (Web Model Context Protocol) as a way for a site to declare the actions it offers to AI agents, for example booking an appointment in a calendar, so an agent does not have to scrape the interface; the speaker called it faster, easier and cheaper.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-06glossary term WebMCP
A community speaker said that at the time of the event (30 September 2026) WebMCP was not supported by all browsers but could be enabled for testing through browser flags.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-06
A community speaker said WebMCP has two kinds of tools: declarative ones, mostly HTML annotations such as the fields of a contact form, and imperative ones for other actions such as booking, filtering a catalogue, adding products to a cart, getting product specs or reordering.
Speaker Carlos OrtegaEvidence transcript
Used byrequirement DEV-AIF-06glossary term WebMCP
A community speaker summed up agent readiness in three layers: a good CLS improves how agents see content, schema, landmarks and ARIA how they understand it, and WebMCP how they interact with it.
Speaker Carlos OrtegaEvidence transcript
A community speaker contrasted real audience understanding (reading logs, tracking how the brand is mentioned and perceived, letting customer complaints change pricing pages and products) with keyword-tool research that turns People Also Ask questions into H2s and blog posts.
Speaker Thiago PojdaEvidence transcript
A community speaker said the SEO industry turned everything into a metric because budgets are approved on numbers, not ideas.
“an idea doesn't get you budget approval; a number does.”
Wording checked against the slide or recording
Speaker Thiago PojdaEvidence transcript
A community speaker said not all SEO work was wrong: a site still has to be crawlable and visible before it is even considered, and much of that work is plumbing, not gaming.
“There's a lot of plumbing that's not gaming.”
Speaker Thiago PojdaEvidence transcript
A community speaker said Google's AMP project, meant to get people building fast websites and addressed directly to engineers and marketers, was smart, but the industry reduced it to a metric and a checklist ('now I have an AMP website').
Speaker Thiago PojdaEvidence transcript
A community speaker said a third-party study from early 2026 found that websites with JSON-LD were shown much more often, and that the industry reacted by adding schema markup everywhere.
Speaker Thiago PojdaEvidence transcript
A community speaker said a later controlled study, which added schema markup to pages that lacked it and kept everything else unchanged, found only a negative correlation: the markup had been a sign of teams doing everything else right, not a cause.
Speaker Thiago PojdaEvidence transcript
A community speaker said listicles do earn more mentions in AI answers, not because they are trusted more but because LLMs mainly check that the thing being talked about exists, not whether the content is honest.
Speaker Thiago PojdaEvidence transcript
A community speaker suggested that LLMs pick up some of Google's years of work on judging honesty and trust (E-E-A-T) by proxy, because they send many queries (probably fan-out queries) to search engines.
Speaker Thiago PojdaEvidence transcript
A community speaker warned that the SEO industry is overusing listicles for AI visibility the way it once overused links.
Speaker Thiago PojdaEvidence transcript
A community speaker cited a third-party study (heard as Peec AI's) finding that 43% of prompts produced fan-out queries in a language other than the one the user searched in.
Speaker Thiago PojdaEvidence transcript
A community speaker showed a Spanish-language AI prompt about places to eat in Barcelona whose fan-out queries were all in English, so the content used to answer may be written by and for a different audience than the one asking.
Speaker Thiago PojdaEvidence transcript
A community speaker argued for adopting the GEO label as the industry's chance to leave behind the bad reputation SEO built, unlike Google's view earlier the same day that the new name is not needed.
Speaker Thiago PojdaEvidence transcript
Used bystory angle A-001
A community speaker compared traditional search to walking into a store and searching the shelves yourself, and AI search to asking a sales assistant who searches, selects, summarises and presents the answer for you.
Speaker not identifiedEvidence transcript
A community speaker said most AI-visibility tracking copies rank tracking and is worse than it, because it tracks invented prompts with no known search volume.
Speaker not identifiedEvidence transcript
A community speaker proposed measuring AI search as a five-stage funnel (know, association, search, retrieval, selection) to learn whether the AI understands, trusts and finally chooses a brand, measuring the earlier stages rather than only the last.
Speaker not identifiedEvidence transcript
A community speaker said most AI-visibility measurement covers only the selection stage, whether a brand is mentioned or selected, without showing why or why not.
Speaker not identifiedEvidence transcript
A community speaker named two signals that in their view strongly influence whether an AI selects a brand: the bias already in the model's memory before it searches, and the brand's presence in the search results the model is grounded on.
Speaker not identifiedEvidence transcript
A community speaker proposed an 'AI bot allow rate' for the know stage: the share of AI training bots a site allows (three of four is 75%), with 100% as the target.
Speaker not identifiedEvidence transcript
A community speaker advised not blocking AI training bots in most cases, because a model that may not crawl a site finds it harder to represent the brand accurately in its parametric memory.
Speaker not identifiedEvidence transcript
A community speaker described the association stage as whether an AI already connects a brand with the topics that matter to it before running any search, and said problems there are mostly brand problems.
Speaker not identifiedEvidence transcript
A community speaker's 'brand association rate' asks a model, with web browsing switched off so that only its memory answers, to name ten topics it associates with the brand, or the other way round, starting from a topic.
Speaker not identifiedEvidence transcript
A community speaker advised repeating brand-association prompts over time: association in every run is good, in about half the runs shows some association with work to do, and little or none shows a clear gap.
Speaker not identifiedEvidence transcript
A community speaker proposed a 'brand co-occurrence rate' for the search stage: the share of an AI's fan-out queries that contain the brand (one in three is 33%), which also hints at the associations the model already has; problems there are mostly relevance problems.
Speaker not identifiedEvidence transcript
A community speaker said Microsoft is ahead of Google in giving site owners the grounding queries its AI uses.
Speaker not identifiedEvidence transcript
A community speaker described the retrieval stage as whether an AI requests a site's pages when grounding its answer; if it does not, the cause may be a crawling or an indexing issue.
Speaker not identifiedEvidence transcript
A community speaker described the selection stage as whether a brand is chosen in the answer, measured with mentions and similar metrics; failure there is mostly a trust or relevance issue.
Speaker not identifiedEvidence transcript
Closing the lightning talks on automation and AI, one of the session's Google hosts called AI a tool that can be used well or badly: it can even be used for content generation, which none of the talks featured, and it is fantastic for coding and analysis but less so for other tasks.
Speaker not identifiedEvidence transcript
One of the Google hosts of the AI lightning talks advised always checking what AI tools do and focusing on what the business and its users need.
Speaker not identifiedEvidence transcript
A community speaker applied Pascal's line about makers of false windows built for symmetry, whose rule is to make pleasing figures rather than to speak accurately, to today's LLMs: sometimes they are not trying to produce the most accurate figures or the most accurate interpretations of them.
Speaker Rafael KovashikawaEvidence transcript
A community demo's script used both input modes of Google's Rich Results Test: the URL mode for public pages, which Google fetches itself, and the code mode, into which the script pasted the page's HTML, for private pages and pages Google cannot fetch (best reading of a largely unintelligible recording; the second kind of page was heard as 'dead', possibly 'dev').
Speaker not identifiedEvidence transcript
Used byrequirement DEV-SDA-11
In a community demo, the structured-data problems Google's Rich Results Test reported on the test page included an empty name, a breadcrumb problem and a price of zero, with errors shown in pink and warnings in orange (best readings of a largely unintelligible recording; each item is heard in only one of the two recordings).
Speaker not identifiedEvidence transcript
Used byrequirement DEV-SDA-08
In a community demo, the small model choosing keywords (Claude Haiku) followed fixed rules, no brand terms, no navigation terms and at most two keywords per page; the Spanish example was 'zapatos de mujer' and 'comprar zapatos de mujer' (women's shoes, buy women's shoes).
Speaker not identifiedEvidence transcript
Search Console Help defines position as the position of a link on Google, where 1 is the topmost position, and says the Performance report shows the topmost position a property's link occupied, averaged across all queries.
Publisher Google Search Console Help
Search Console Help separates critical issues, which make a structured-data item invalid and stop it from appearing as a rich result, from non-critical issues, listed under 'Improve item appearance'; the Rich Results Test reports valid items that have warnings.
Publisher Google Search Console Help
Used byrequirement DEV-SDA-11
To use the Search Console API you need a Google Account with Search Console permission on the property, a project in the Google API Console and OAuth2 credentials (all Search Console APIs except the Testing Tools API need OAuth2).
Publisher Google Search Console API documentation
Google states that all use of the Search Console API is free of charge but subject to usage limits.
Publisher Google Search Console API documentation
Google's SEO Starter Guide says headings in semantic order help screen readers but do not matter to Google Search if used out of order, and that there is no ideal number of headings for a page.
Publisher Google Search Central
Used byrequirement DEV-HTM-04
Chrome's WebMCP documentation describes WebMCP as a proposed web standard that site owners can test through an origin trial starting in Chrome 149 or through a Chrome flag for local development.
Publisher Chrome for Developers
Used byrequirement DEV-AIF-06glossary term WebMCP
Google's Rich Results Test help says the tool tests either a page's full URL or a pasted code snippet (Code instead of URL); all page resources must be reachable by an anonymous user on the internet, so resources behind a firewall or a password are not available to the test unless exposed, for example through a tunnel.
Publisher Google Search Console Help
Used byrequirement DEV-SDA-11glossary term Rich Results Test
Google's merchant listing documentation lists the product name as a required property and, unlike product snippets, requires a price greater than zero for merchant listing experiences.
Publisher Google Search Central
Used byrequirement DEV-SDA-08
Author's view: the Search Console API itself costs nothing, so the cost check advised on stage before automating reports is about the API's usage limits and the paid parts of the pipeline, such as storage and AI model calls.
Author Ibrahim Anjro
Author's view: an LLM fix loop that stops when Google's test passes proves only that the JSON-LD is valid, not that its values are true; Google's generative AI content guidance asks for AI-generated structured data to be fact-checked as well as validated.
Author Ibrahim Anjro
A logical heading hierarchy is not a Google Search requirement, since Google says out-of-order headings do not matter to Search; the case for it is accessibility, where web.dev advises against skipping levels, and agents that read the accessibility tree.
Author Ibrahim Anjro
Google's AI optimisation guide says structured data is not required for generative AI search and needs no special schema.org markup, and on Day 2 Google said raw schema.org is generally not put into model context (D2-C477); use markup for rich-result eligibility and clear data, not as an AI-visibility lever.
Author Ibrahim Anjro
The study cited on stage matches Peec AI's analysis (reported in February 2026) of over 10 million ChatGPT prompts and 20 million fan-outs: 43% of the fan-out searches for non-English prompts ran in English, and nearly 78% of non-English prompt runs had at least one English fan-out (from 66% for Spanish to 94% for Turkish), so the 43% is a share of fan-out searches, not of prompts.
Author Ibrahim Anjro
The llms.txt study cited on stage matches Ahrefs' June 2026 study of 137,210 domains: 28% (about 38,000, the 'almost 40,000' sites heard on stage) published an llms.txt file and 97% of those files received no requests at all in May 2026; the files that were fetched got about 22,000 requests across some 1,100 domains, around 20 each, not the 'two hits a month' heard on stage.
Author Ibrahim Anjro
Used byrequirement DEV-AIF-02
A community speaker argued for adopting the GEO label as the industry's chance to leave behind the bad reputation SEO built, unlike Google's view earlier the same day that the new name is not needed.
Gary Illyes argued that GEO is a label invented to create a new field and is not needed. Understanding how SEO works is enough.
A community speaker argued for adopting the GEO label as the industry's chance to leave behind the bad reputation SEO built, unlike Google's view earlier the same day that the new name is not needed.
Google's answer to 'SEO is dead, long live GEO' is not to worry about the name: good SEO is good GEO and AEO.
A community speaker said llms.txt files are not necessary, citing a third-party study from summer 2026 (heard as Ahrefs') that looked at almost 40,000 sites over one month: 97% had no AI agent hits on the file, and those that had any got about two hits a month.
A community speaker said a block of content without semantic HTML or landmarks is just a div whose purpose an agent cannot tell, and recommended landmark elements (header, nav, main, article for independent sections, footer) plus p and h1-h6 for text.
Google's guide for generative AI features says browser agents may read a site through screenshots, the DOM structure and the accessibility tree, and recommends semantic HTML because it helps users such as screen reader users navigate a page.
A community speaker described ARIA (Accessible Rich Internet Applications) as a set of attributes, not a programming language, that adds accessibility information to HTML: a div used as an 'add to favourites' button can get role=button, an aria-label and aria-pressed set to true or false.
Combining schema-based interfaces with web standards such as WebMCP lets a site expose tools to AI agents, which can then operate the interfaces, for example to test them.
A community speaker presented WebMCP (Web Model Context Protocol) as a way for a site to declare the actions it offers to AI agents, for example booking an appointment in a calendar, so an agent does not have to scrape the interface; the speaker called it faster, easier and cheaper.
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.
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.
In a community lightning talk, agency founder Nik Vujic presented how his agency feeds Search Console and GA4 data to an LLM agent to support decisions for its clients.
A community speaker who is not a developer automated monthly SEO reports with Python in Visual Studio Code, using Claude and ChatGPT as coding partners and Google Cloud for access to the Search Console API.
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.
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.
Nik Vujic's slide said no agent is needed to start: pull the Search Console data, import it into any LLM and query it in conversation; his agency built its agent to have everything in one place.
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.
Google does not tell lightning-talk and poster speakers what to talk about; they bring their own ideas.
Google received close to 200 submissions for the community lightning talks of the Deep Dive, reviewed them over about a month, and gave each selected community speaker seven minutes on stage.
A community speaker said llms.txt files are not necessary, citing a third-party study from summer 2026 (heard as Ahrefs') that looked at almost 40,000 sites over one month: 97% had no AI agent hits on the file, and those that had any got about two hits a month.
Myth: optimise for AI over readers. Google's answer: optimise for people, with no need to obsess over precise keywords or AI phrasing, no need to chop content, and no need for llms.txt.
A community speaker said AI agents understand a page through a combination of three inputs: a screenshot, the DOM (the HTML plus the changes rendered by JavaScript) and the accessibility tree.
Google's guide for generative AI features says browser agents may read a site through screenshots, the DOM structure and the accessibility tree, and recommends semantic HTML because it helps users such as screen reader users navigate a page.
User-initiated fetchers, such as a translation service fetching a page a user asked to translate, are a different category from crawlers and generally do not check robots.txt, because a user asked for the fetch.
A community speaker pointed out that, according to Google's documentation, user-triggered fetchers, which fetch a page because a user asked for it, generally ignore robots.txt rules.
Google says user-triggered fetchers, which fetch a URL because a user asked for it in a Google product (for example Gemini Notebook fetching URLs users add as sources, or Google-Agent acting on a user's request), generally ignore robots.txt rules.
A community speaker pointed out that, according to Google's documentation, user-triggered fetchers, which fetch a page because a user asked for it, generally ignore robots.txt rules.
Review the filters and regexes that Search Console's AI-powered configuration proposes before trusting the numbers: Google's launch post calls the feature experimental, warns that AI can misinterpret requests, and limits it to the Search results Performance report (not Discover or News) and to configuration, not sorting or exporting.
The first rule for using LLMs on search data, according to a community speaker, is never to let the model check its own output.