Google says new pages take about 20 hours to be found, some never are

Gary Illyes, an analyst at Google, showed internal timing data for crawling, indexing and serving on October 2, 2026, the closing day of Search Central Live Deep Dive Europe in Barcelona. A new URL is typically discovered in about 20 hours, while the slowest cases run from weeks to never, according to a recap by John Campbell of the agency ROAST.

In Short

At a Google event in Barcelona on October 2, 2026, a Google analyst showed how long it usually takes for a web page to be found, stored in Google's index and shown in results. Most steps take hours, days or a few weeks, but a few take months, and some may never finish when Google judges a page not worth the effort. The practical result is a set of Google-supplied ranges that publishers, agencies and clients can compare with what they see on their own sites, in place of rules of thumb.

The event behind the data

In March 2025, Google outlined a multi-day Deep Dive format for its Search Central Live series, built around extended technical sessions and hands-on workshops. The first run took place in Bangkok from July 23 to 25, 2025, where Illyes told attendees that AI search shares its crawling mechanisms with traditional search. On July 6, 2026, Google named Barcelona as host of the first European edition, chosen from six shortlisted cities after a community interest form closed on July 1. Admission was free, application-based and English-only.

The three days ran from September 30 to October 2, 2026. ROAST, a marketing agency, published a recap of each, written by John Campbell, its Head of Innovation and AI: Day 1 on crawling (September 30), Day 2 on indexing (October 1) and Day 3 on serving and ranking (October 2). All three are among the documents supplied for this article, together with three community infographics, one per day, and a LinkedIn thread around them. The recaps are an attendee's account of live sessions, and none is a Google publication.

The tables Illyes put on screen

Day three was given over to serving: ranking, quality, Search Console and, in the recap's phrase, "how long things actually take". The last of those drew the most attention. According to the recap, Illyes showed "Google's own data on how long things take across crawling, indexing and serving", material it calls "brand new information". The recap adds that the times "are based on Google's internal analysis".

One gap needs stating at the outset. The recap says each process had a fastest, a typical and a slowest time, yet the tables it reproduces carry only the last two columns. Everything below rests on those.

Crawling

Figures as listed in the ROAST recap of Illyes' session.

Two rows frame the rest. Discovery of a new address is listed at about 20 hours in the typical case; the refresh of a known one at about 30 days. Both carry slowest cases measured in weeks, and both include "never" as an outcome. The robots.txt row is the tightest, at roughly a day typical and 25 hours slowest. According to the recap, crawl capacity can drop within seconds when Google backs off, for instance when a server is struggling, while the road back runs from one to three weeks.

Indexing

Indexing end to end, at about 1.5 hours typical, is the fastest headline figure on the slide, though the recap defines the term narrowly: "End to end means all the critical processes finish successfully." Site moves sit at the other extreme, with 1-3 months typical and 6 months to a year or more at the slow end, although the recap notes that "A small site move can be done in a few weeks." Canonicalisation changes and removals both take 1-3 weeks in the typical case, and the slowest cases are given in months.

Serving

The serving rows hold the longest waits. Recovery from a core update is listed at 3-6 months typical, stretching to between six months and a year when it has to wait for the next core update. Manual action removals take 1-2 weeks, or 4-6 weeks at the slow end, with dormant sites facing "much longer". The recap gives core updates a rollout of 2-4 weeks and spam updates a rollout of 1-2 days.

What the data comes with

Illyes attached one caveat, according to the recap: many of the processes are linked. A page cannot be indexed until it has been crawled, so delays stack up. The recap's own commentary observes that "never" shows up a lot and is usually tied to quality, and that "Fast technical fixes don't help if Google doesn't think the page is worth it." That second sentence is the recap author's reading rather than a quotation from Illyes, a distinction the source itself draws by placing it under a separate "Why this matters" heading.

Where the numbers meet earlier reporting

Several rows can be set beside figures PPC Land has recorded over the past year.

Take spam updates. The recap gives a rollout of 1-2 days, yet Google's 2026 spam updates have not followed a single pattern. The March update finished in 19.5 hours, June took two days and August three, while the September update, which began at 09:15 Pacific on September 24, carries a window of up to two weeks, the longest of the year. Illyes spoke eight days into that window. Whether the update had finished by then is not stated in the supplied material.

Core updates tell a similar story. The recap's range of 2-4 weeks is wider than recent history: the December 2025 core update ran 18 days, from December 11 to December 29, after earlier 2025 updates of 14 and 16 days, and the March 2026 update took 12 days and 4 hours, from March 27 to April 8. Four days into that March rollout, John Mueller said that core updates have no single deployment mechanism and that components sometimes have to be switched on one after another. A rollout length and a recovery time are different measurements, of course; the 3-6 months in the table concerns the latter.

On canonicalisation, Google updated its troubleshooting guide on July 10, 2026 to say systems may take up to two weeks to recognise a fix in a cluster of duplicate pages. That sits inside the 1-3 weeks shown on the indexing slide.

What Days 1 and 2 described

Several rows on the Day 3 slides map onto sessions from the first two days. The mechanics come from the recaps of Day 1 and Day 2, again as relayed by Campbell.

Keynote and AI Mode figures

Day 1 opened with Lino Cattaruzzi, President of Google Iberia, who restated Google's mission "to organize the world's information and make it universally accessible and useful" and said user expectations keep rising. According to the recap, he named three shifts: people turning to closed-garden platforms for information once found in search engines, a preference for personalised and more visual content, and AI meeting information needs through new products. Of his five points on how Search will evolve, two stand out. "Made for Search" content will not be successful, and long-held traffic patterns are likely to change, which Google framed as new opportunities for all sites. The second sits beside NewzDash data on more than 400 news publishers, in which Google Web Search fell from 51.10% of publisher traffic in 2023 to 27.42% in the fourth quarter of 2025 while Google Discover rose from 37.03% to 67.51%.

Illyes then gave AI Mode statistics, as listed in the recap. Queries have doubled every quarter since launch; the average AI Mode search is about three times the length of a traditional query; 1 in 6 AI Mode searches are multimodal, using voice or images; and follow-up queries have grown by more than 40% on average per month in the US. He sorted new uses into four groups. Brainstorming queries have grown 30% faster than queries overall since launch, searches starting with "which" 40% faster than AI Mode queries overall in the past six months, and planning queries 80% faster over the same period. The fourth group, learning, carried no figure in the recap. The numbers match a Google data report on AI Mode's first year, reported on May 19, 2026, which also put monthly users above one billion. The conclusion that one URL cannot answer a long, multi-part question, so that Google must build the answer from several sources, is the recap author's own.

Cherry Prommawin of Google and Illyes then covered how Search works. Search, AI Overviews and AI Mode rely on the same main crawler, Googlebot; Gemini is separate and uses a different crawler, and the recap adds "For now at least." It condenses the pair's message as "good SEO is good GEO".

Crawl budget, errors and crawlers

Prommawin closed Day 1's main sessions on crawl budget. According to the recap, Google runs one central crawling system for Search, Ads, Shopping and Images, and its main job is to fetch from the web without overloading anyone's servers. Crawl budget has two parts. The crawl rate limit, also called hostload, is shared across all of Google's crawlers for a host; connection time, time to first byte and 429 or 5xx status codes drive it, and if any of those rise, Google slows the crawl. Crawl demand is driven by the quality of the site, how often URLs change and how popular they are on the web, and where Google does not know a URL's quality or popularity it uses the aggregate for the parent path, then that path's parent, and so on. Errors, useless pages and infinite URL spaces such as calendars and parameters burn budget. The Day 3 rows for crawl capacity and crawl demand updates look like the timing counterparts of those two parts, though the recaps use different names for the first and do not say so outright.

Two problems were described as growing. DNS and CDN errors are harder to debug, and CDNs are blocking more bot traffic; a new Cloudflare setup, for instance, might end up blocking Googlebot, with the blocks showing up as HTTP errors. Soft 404s are the other: a CAPTCHA challenge page can return a 200 OK, so that Google sees a page that looks fine but has no real content. Crawler access has a commercial side too. Cloudflare had planned in July 2026 to block Googlebot, Applebot and Bingbot for sites refusing AI training, and dropped the plan, citing fewer than 1% of its sites blocking search crawlers against 17% wanting to block AI training, according to a report of September 27, 2026. Dave Smart of Tame the Bots added a robots.txt point: when a page loads files from another domain, such as a CDN or a separate subdomain, that domain's robots.txt applies, not the page's own.

Rendering, duplicates, structured data and media

Day 2 turned to indexing. Mueller went through the robots meta tags, from noindex and none to max-snippet and max-video-preview, and closed on a pairing of max-image-preview:large with max-snippet:-1, the first allowing large image previews and the second letting Google show as long a text snippet as it wants. The first of those is also the setting Day 3's Discover session named.

Erin Sparling of Google explained rendering. Google loads pages with a viewport around 10,000 pixels tall and does not scroll or click, so JavaScript that waits for a scroll, interaction or click will not run, and content that loads only on a scroll event will not be seen. That is the background to the rendering row, with seconds of work and hours in a queue. Illyes said content in the main part of a page is given more importance than navigation, footer and other boilerplate.

Mueller covered duplication. Google first clusters similar pages, picks one as the canonical, then assigns all signals to that main page. The reasons he gave were to stop duplicate content appearing in results, to leave more room to store unique content, to keep the signals from every version and to understand alternative versions such as hreflang, branding changes and migrations. The last point, the recap says, is why a site: search on an old domain can still show results after a migration. It reads as background to the canonicalisation and site move rows, at 1-3 weeks and 1-3 months typical.

On structured data, Ryan Levering of Google said Schema.org data is cleaned and filtered, then either shown in normal search results or used as context for AI Overviews and AI Mode. According to the recap, the text in the schema is passed along with the content on the page, so "it does feed Google's AI features", not as code but added to context. His view on adding it was that it is never a negative, and the open question is whether it is worth the effort. Server-side validation of structured data is coming, and he hinted at more shopping schema news for Day 3; the Day 3 recap records no schema release. Illyes added that images, structured data and video are pulled out and passed to a separate media indexing engine, which is the setting for the image and video rows.

International, the index and Google Trends

Prommawin listed common hreflang mistakes, using more than one implementation method at once and using incorrect language codes, and said localisation goes beyond translation to date formats and currency. Illyes showed how results are returned from the index: a query is split into words, each pointing to a list of pages, and Google matches the pages that appear across the lists, with vector embeddings used alongside. Omri Weisman, an Engineering Manager on Google Trends, said Trending Now data is fresh to within 10 minutes and described Gemini in Explore, which suggests search terms and compares up to eight terms at once, a feature that began rolling out on desktop on January 14, 2026. There was no mention of the Trends API going public, and the later Q&A brought no update either. Google had opened alpha testing for that API on July 24, 2025.

The rest of Day 3

Query understanding and AI search

Mueller opened Day 3 with query understanding, treating AI search and normal search separately. In normal search, the steps run from language detection, which is harder for brand terms because a brand name often is not a word in any language, through removal of stop words that do not matter, which entity recognition overrides for names such as "The Lord of the Rings", to synonym expansion. Mueller called the last a very important part of the ranking system. It works much like the OR operator, so that a search for "fried chicken place in barcelona" is rewritten with place, area, location and restaurant as alternatives, and some synonyms depend on context: "GM" means General Motors next to "car" and genetically modified next to "barley".

The AI version adds a step. The query goes to a large language model, which creates fan-out queries and sends them back to the search engine; the results ground the answer, which returns with links. Each fan-out query still passes through query understanding. They do not appear in Search Console, the recap says, because they do not come from users. Illyes then covered retrieval: documents carry signals such as language and country, quality makes the final decision, and retrieval happens before ranking.

Quality, spam and Discover

Duy Nguyen, a Search Quality Analyst at Google, said there is no single ranking system that works out quality. According to the recap, hundreds of signals exist, they differ by result type, and Google runs a large number of tests all the time. He pointed to the public Search Quality Rater Guidelines, which define quality through four things: effort, originality, talent or skill, and accuracy. Low-quality pages can be written by humans or by AI; the example of good work was a shoe review built on its own metrics, its own photos and detailed research. "A high percentage of content is spam, and that is what the spam updates target," the recap reports, adding that Google is now using AI to help catch more of it.

Eric Murillo, a Trust and Safety Analyst for Discover at Google, covered how content gets into the feed. There is no special Discover bot; everything runs on the normal index stack. Each piece of content passes an eligibility gate with three outcomes: a policy violation is filtered out, low quality is judged on E-E-A-T and filtered out, and safe and trusted content becomes eligible. Image quality then decides performance. The recap lists image recommendations of at least 1,200 pixels wide, 300,000 or more total pixels, a 16:9 ratio and the max-image-preview:large robots setting enabled, and flags generic images such as site logos, along with text-heavy images, as weak on quality. Murillo also said AI slop is increasing, that user feedback is what new spam filters are built from, and that manual actions appear in Search Console.

Illyes closed the morning with a single number: according to the recap, Google filters out 40 billion spam pages a day. He gave three reasons for continuing to update how quality is judged. Content formats have expanded beyond ten blue links; content breadth has grown; and spam keeps changing, from cloaking, doorway pages, scraped content and link spam to hacked content. Core updates, the recap says, "don't focus on websites" but target content at page level rather than domain level, which is why some pages on a site rise while others fall. Illyes also warned that scaled content is becoming more of a problem than link spam.

The 40 billion figure carries a wording difference. Google's earlier public material on SpamBrain, its spam-prevention system, frames the same number around spammy pages found each day rather than filtered out, a distinction between detection and removal that the recap does not address.

Shopping, Search Console and the closing message

Alex Jansen, a Software Engineer at Google, described conversational attributes for shopping: flexible product information, rather than thousands of new fixed attributes, meant to help users and their AI agents find detail on surfaces such as AI Mode. His example was a kettle, where buyers ask how much water it holds and whether it has a safety feature for boiling water. Product highlights were flagged as another field that helps in AI search, and the recap sums up the message as a raised bar for product data quality.

Ariel Kroszynski, Front-End Engineering Manager for Search Console, walked through the past year: Groups, AI-powered configuration for setting up reports, and Multimodal, the newest feature. On generative AI reporting, which the recap calls the hot topic, there was no update; the team is working on adding more AI search data to the reports, but gave no dates or details. A dark-mode mock-up closed the talk and, the recap notes, is not real "at least not yet". For context, Search Console's generative AI report reached all sites on August 31, 2026 with impressions, pages, countries, devices and dates but no clicks or queries, and the Multimodal filter began rolling out on September 24, 2026, also without query data.

Illyes finished the three days with three messages: use AI, use it responsibly, and "AI on Google is just SEO". According to the recap, AI features in Google Search use exactly the same processes as traditional results, and he said that "We didn't need a new acronym for mobile-first indexing or structured data, and we likely don't need one for AI on Search." The position is consistent with what Danny Sullivan and Mueller said in December 2025, when they described no special optimisation as needed for AI search.

Community sessions across the three days

Community speakers filled the gaps between Google sessions. Their numbers are speaker-reported single examples.

On Day 1, Rafael Kovashikawa of FUSE warned that AI can pull data correctly and still reach the wrong conclusion, and described having the model show its calculations before a second model checks them. Carlos Ortega Roldán named llms.txt, Markdown versions of pages, which he said create duplicate content, and robots.txt, which agents acting on a user's request can ignore, as unnecessary for agents. Agents, he said, read pages through screenshots, the DOM and the accessibility tree, and he pointed to WebMCP, which entered a Chrome 149 origin trial on May 20, 2026. Thiago Pojda of SIXT said GEO is harder to measure than SEO, that listicles are hard to argue against because they work, and that fan-out queries are not always in the language of the original prompt. Eduard Blacquière set out five stages for tracking AI performance - know, associate, search, retrieve and select - with the usual visibility metrics sitting at the last stage only.

On Day 2, Irene Cecotti of PhantomBuster generated image alt text at scale and reported that performance dropped once the descriptions went live. Martyna Ağanoğlu of Loando and Clar described merging two domains and cutting the site from 2,000 pages to 100, with no traffic drop and higher revenue. Rebecca Yu named robots.txt blocking resources that JavaScript needs as the top rendering problem.

On Day 3, Robert Wojno of Hostinger compared a generic "What is VPS hosting?" guide, with 1.6 million impressions, against a niche guide on building a dedicated server in Palworld, with 46,000; the niche piece had a click-through rate 23 times higher and converted better. Juan Seguí of acceseo added detailed FAQs to a double-pram page and saw organic clicks rise 14%. Alex Wright of iDHL showed one brand whose organic sessions fell 47% and direct sessions 2%, yet which recorded 2.7 times more enquiries per 100 search clicks, a share of search clicks up from 17% to 38%, a Google Ads click-through rate up 36% with flat position, and 83% of its AI citations from sites it does not own. Pablo Pérez of Google revisited the Messy Middle research and said UK search interest in "is [brand] legit" has climbed since 2020 and reached its highest point in 2026. Noe Rivas of seosve paired a page with a video and ranked across AI Overviews, Gemini, YouTube and image search, and a poster by John McAlpin said serving Cloudflare-created versions of pages to AI bots helped performance in ChatGPT.

The community notes against the recaps

Three infographics, one per day, are titled "AI search is just search.", "Raw HTML or it didn't happen." and "Commodity content loses." A LinkedIn comment thread around them shows John Mueller commenting, "Thanks for sharing so many notes!" Serhii Omelchenko, whose profile lists Kharkiv, Ukraine and the headline "I automate SEO.", carries LinkedIn's Author label on a reply in the thread. The screenshot does not show the original post, so the link between Omelchenko and the graphics is inferred from that label. Another commenter, the publisher of Search Engine Watch, said the graphics might be reused with due credit; Omelchenko answered "feel free". A thank-you from a Google staff member is not a verification of each point, and the graphics are an attendee's notes, not Google statements.

With all three recaps supplied, the 28 numbered points can be compared with what the recaps report. A comparison of that kind has a limit: the recaps do not cover every session or the Q&As, and the Day 1 recap says only that what was said in the Q&A "stays with the people in the room". Absence from a recap is not contradiction.

Matched in substance. The recaps support the Day 1 point that AI Overviews and AI Mode use Googlebot while Gemini is separate; the Day 2 point that main content is separated from header, sidebar and footer; the video sitemap point; and the Day 3 points on silent query rewriting, fan-out as extra searches that run through normal query understanding, quality as one signal among hundreds, and effort, originality, skill and accuracy as its definition. Quality, not authorship, is the target in both, since Nguyen said low-quality pages can be human or AI written.

Stated differently. On multimodal search, the Day 1 graphic says 1 in 6 queries is multimodal; the Day 1 recap says 1 in 6 AI Mode searches, using voice or images, and the graphic's Gen Z remark appears nowhere in the recaps. On llms.txt and Markdown copies, the graphic marks both as skips, saying Google Search does not use llms.txt and that Markdown copies carry a cloaking risk; in the recap, the statements come from Ortega Roldán, a community speaker, and the cloaking point is absent. The graphic's Day 3 claim that "Top-10 lists lose" sits against Pojda's Day 1 view that listicles work, though the two address different questions: what Google's quality framing rewards, and what performs in practice.

The sharpest difference concerns structured data. The Day 2 graphic says schema is "still worth it" but is not fed directly into AI context. The Day 3 graphic says AI Overviews ignore structured data and work from normal indexed text. The Day 3 recap has Illyes saying AI results do not need structured data because they take text and build the result from it. The Day 2 recap has Levering saying the text in schema is passed along with the page content and does feed AI features, not as code but as context. The accounts may be reconcilable, since code is not read as code while its text content may be, but a reader taking the Day 3 headline at face value would miss the Day 2 description. Coverage of Google's AI search guide, reported on May 15, 2026, said no llms.txt file is required and dismissed content chunking, and cited an Ahrefs study that found no measurable ranking difference when schema was added to pages already appearing in AI Overviews.

Not found in the recaps. Several graphic points have no counterpart in any of the three: that Google does not try to detect AI content; that crawling more does not help rankings; that a spammy past slows crawling; that Gemini training crawls favour volume while Search crawls favour quality and freshness; that many AI crawlers do not render JavaScript; the chunking myth built on a context window of around a million tokens; that AI-generated images are fine; that machine translation is not a bad signal; that each video needs a watch page; that SpamBrain catches 5x more spam; the roughly 800,000 quality tests in 2023 and roughly 5,000 changes a year; and that PageRank is barely used. The hreflang point differs in kind: the Day 2 recap covers hreflang as an aid to localisation and a source of common mistakes but says nothing on how Google assigns a language, which the graphic says it does from content and not from hreflang or URL codes. The 5x claim has no period on the graphic; Google's published April 2023 claim of five times more spam sites detected in 2022 than in 2021 did name one.

Several of the graphics' points have a record in earlier PPC Land coverage, whether or not the recaps mention them. A study reported on July 2, 2026 found llms.txt files rising 8.8-fold in twelve months to 36,120 sites by May 2026, while 97% of files received zero requests in May. Sullivan warned against breaking content into chunks for AI in January 2026. On February 5, 2026, Mueller and Microsoft's Fabrice Canel questioned separate Markdown pages for AI bots, and on May 15, 2026 Google extended its spam policies to AI Overviews and AI Mode, listing cloaking among them. And on June 1, 2026, Illyes argued in a LinkedIn post that AI agents and LLMs would have had an easy life on a web of clean HTML, a point close to the Day 2 graphic's headline.

What the numbers mean for marketing teams

For agencies and in-house teams, the practical value of the tables is that they put Google-sourced bounds on questions that clients and finance departments ask: how long until a migration settles, how long until a technical fix registers, how long until a penalty lifts. The recap says the data "gives everyone realistic timelines", with a typical site move taking 1-3 months to settle and core update recovery typically taking 3-6 months.

The quality thread changes how the slow cases read. Where the recap writes "never", the word is usually attached to a quality judgement, so the delay is no longer a queue that eventually clears. Day 1's description of crawl demand, driven by site quality, change frequency and popularity, points the same way: the inputs that decide how often Google returns are partly the inputs that decide whether a page is kept.

Measurement is the other pressure point. The generative AI report in Search Console carries no clicks or queries, fan-out queries do not appear in Search Console, and the Search Console team had no update on adding AI search data. Blacquière's five stages describe the same gap from the practitioner side: visibility captures only the last step. Wright's case, with organic sessions halving while enquiries per 100 search clicks rose 2.7 times, is a single brand, but it illustrates why the Day 3 recap urges looking beyond click data to enquiries, conversions, brand share of clicks and AI citations. Pérez's finding on "is [brand] legit" searches points the same way: trust signals form part of the journey before a click.

Finally, there is the question of provenance. The tables did not appear in a Google blog post or help page in the material supplied; they reached readers through an attendee's recap of a live session. Whether Google will publish them in documentation is not stated.

Limits of the source material

The tables are described as based on Google's internal analysis, with no methodology, sample or period attached. The 800,000 tests, 5,000 changes and 5x spam figures appear only on a community infographic, without a source there. The case-study numbers from Hostinger, acceseo, iDHL and the Day 2 speakers are the speakers' own, drawn from single examples. The Day 2 recap skips several sessions, which its author, presenting a poster, says he missed. And the recaps' author attended as a delegate: the commentary sections are his reading of the sessions, not Google's position.

Timeline

• March 17, 2025 - Google outlines a multi-day Deep Dive format for Search Central Live Asia Pacific 2025.
• July 23-25, 2025 - Deep Dive Asia Pacific runs in Bangkok; Illyes says AI search shares crawling mechanisms with traditional search.
• July 24, 2025 - Google opens alpha testing for a Trends API.
• December 11-29, 2025 - December core update rolls out over 18 days.
• December 18, 2025 - Sullivan and Mueller describe AI search as needing no special optimisation, as reported on this date.
• December 23, 2025 - NewzDash data shows Google Web Search falling to 27.42% of news publisher traffic.
• January 8, 2026 - Sullivan advises against chunking content for AI on a Google podcast.
• January 14, 2026 - Gemini suggestions begin rolling out in Google Trends.
• February 5, 2026 - Mueller and Canel question separate Markdown pages for AI bots.
• March 27 - April 8, 2026 - March core update runs 12 days and 4 hours.
• March 31, 2026 - Mueller explains why core updates deploy step by step.
• May 15, 2026 - Google's AI search guide and spam policy extension to AI Overviews and AI Mode.
• May 19, 2026 - Google's AI Mode first-year report reports queries doubling every quarter.
• May 20, 2026 - Chrome 149 opens a WebMCP origin trial.
• June 1, 2026 - Illyes posts on JavaScript and AI agents.
• July 2, 2026 - llms.txt adoption study reports 97% of files with zero requests in May.
• July 6, 2026 - Google confirms Barcelona as the Deep Dive Europe host; applications close September 5, with confirmations by September 10.
• July 10, 2026 - Google's canonical troubleshooting guide adds a two-week re-evaluation note.
• August 31, 2026 - Search Console's generative AI report reaches all sites, without clicks or queries.
• September 24, 2026 - September spam update begins at 09:15 Pacific with a window of up to two weeks; the Multimodal filter starts rolling out in Search Console.
• September 27, 2026 - Cloudflare drops a planned Googlebot block for sites refusing AI training.
• September 30, 2026 - Deep Dive Europe opens in Barcelona; Day 1 covers crawling, with a keynote from Lino Cattaruzzi and AI Mode figures from Illyes.
• October 1, 2026 - Day 2 covers indexing, with sessions on rendering, duplication, structured data, media, international SEO and Google Trends.
• October 2, 2026 - Day 3 covers serving and ranking; Illyes presents the timing data and the event closes.
• October 8, 2026 - Latest end of the September spam update window.

Related PPC Land coverage

• Google picks Barcelona over five rivals for first EU Search Deep Dive - how the host city was chosen and the application terms.
• Google search experts reassure industry on AI search at Bangkok event - Illyes on AI search sharing crawling and indexing mechanisms in July 2025.
• Google announces deeper, multi-day format for Search Central Live Asia Pacific 2025 - the origin of the Deep Dive format.
• Google's September spam update gets a two-week rollout, its longest yet - 2026 spam update durations compared.
• Sites breaking Google spam rules face lower rankings in two-week window - start time and window of the September update.
• Why Google's core updates take weeks: Mueller explains the mechanics - sequential deployment of core update components.
• Google's December 2025 core update finally wraps after 18 days - dates and duration compared with earlier 2025 updates.
• NYT and Guardian win big as Google's March 2026 core update wraps - the 12-day, 4-hour March 2026 rollout.
• Google forces publishers to wait two weeks for canonical fixes to register - the July 2026 documentation change on canonical re-evaluation.
• Google's new AI search guide challenges everything SEOs thought they knew - guidance on llms.txt, chunking, structured data and AI content.
• Google warns against breaking content into chunks for AI - Sullivan's January 2026 podcast remarks.
• Google and Bing say no: separate markdown pages for AI violate search policies - Mueller and Canel on Markdown copies.
• llms.txt adoption rises 8.8x but 97% of files get zero AI requests - adoption and request data to May 2026.
• Gary Illyes: the web's JavaScript mess is an AI agent nightmare - the June 2026 post on JavaScript and AI agents.
• Google spam policies now officially cover AI Overviews and AI Mode in Search - the May 2026 policy clarification.
• Google gives every site AI search reports, minus the clicks and queries - what Search Console's generative AI report shows.
• Google Search Console gains Lens and Circle to Search data, without queries - the Multimodal filter from September 24, 2026.
• Google tells website owners nothing has changed for AI search - Sullivan and Mueller's December 2025 position.
• Google AI Mode hits 1 billion users - how search changed in one year - the AI Mode usage figures that Day 1 repeated.
• News publishers lose half their Google search traffic in two years - NewzDash data on search and Discover shares of publisher traffic.
• Cloudflare drops planned Googlebot block for sites refusing AI training - the reversal on mixed-use crawlers.
• Chrome 149 origin trial puts WebMCP in developers' hands at last - the proposed standard for exposing site tools to AI agents.
• Google Trends now uses Gemini to suggest search terms - the January 2026 Explore change.
• Google opens alpha testing for new Trends API targeting developers and journalists - the July 2025 alpha that Day 2 did not update.

Summary

• Who: At Google, Gary Illyes, Cherry Prommawin, John Mueller, Lino Cattaruzzi, Erin Sparling, Ryan Levering, Omri Weisman, Duy Nguyen, Eric Murillo, Alex Jansen, Ariel Kroszynski and Pablo Pérez; community speakers including Dave Smart, Carlos Ortega Roldán, Thiago Pojda, Robert Wojno and Alex Wright. The account comes from John Campbell of ROAST, with community notes tied to Serhii Omelchenko.
• What: Internal Google timing data for crawling, indexing and serving, listing typical and slowest times: about 20 hours to discover a new URL, 1-3 months for a typical site move, 3-6 months to recover from a core update, with "never" recurring for quality-related cases. The first two days added crawl budget, rendering, duplication, structured data and AI Mode usage figures.
• When: The timing data was shown on October 2, 2026, the third day of an event that ran from September 30 to October 2, 2026.
• Where: Search Central Live Deep Dive Europe, Barcelona, the first European Deep Dive.
• Why: The recaps do not give Google's motive for presenting the data. The Day 3 recap describes the figures as giving realistic timelines, and records Illyes' closing view that AI features on Google run on the same processes as traditional Search.