Structured data and media

Topics in this area 8

• 87 claims · 11 sessions

  Images

  Images were stressed as important on both days: one in six AI Mode searches is multimodal (Day 1), and on Day 2 Google said images and videos drive a large amount of publisher traffic and that an extracted image can appear in web search, Google Images, Discover and AI features. Google finds images through img elements in the HTML (a picture element counts only through the img inside it) and through image sitemaps, cannot index an image without a src, and does not extract CSS background images; to keep images out of results, its documentation names a robots.txt disallow or a noindex X-Robots-Tag header. Alt text should describe the image for someone who cannot see it and can make the image rank for what it names; Google's image guide calls alt the most important attribute for image metadata, while Gary Illyes ranked it below src and said Google uses only three img attributes (said at the event, not in Google's docs). Day 2 confirmed part of Day 1's slide of image signals (resolution, colour and associated text; the list itself is not in Google's docs): the surrounding text is critical context for understanding and ranking an image, which Google's image guide supports. Google supports nearly all popular formats and recommends modern ones such as WebP with small files; Gary Illyes said the AVIF format currently has ingestion hiccups (heard in two independent recordings), although Google's image guide lists AVIF as supported. Day 3 added Discover and Search Console: Google's Discover slides tie click-through to image quality and recommend images at least 1,200 px wide, over 300,000 pixels and 16x9, enabled by max-image-preview:large, with no logos or text-heavy images. Search Console's multimodal search type, rolling out from 24 September 2026, reports searches with Lens, Circle to Search, uploaded images and Chrome's image search, without a query dimension. Google said image results among web results come from its image index, that an image query is searched as embeddings, and that images are indexed in hours to two days on average while changing the image shown with a text result takes 1-2 weeks (said at the event, not in Google's docs). The second recordings added more. On Day 1 Google said the search box is multimodal, that Lens was its answer to young users uploading pictures and expecting answers about them, and that Lens is used by over 1.5 billion people (Google's May 2025 post says every month); for visual search Google breaks an image into vectors and returns results from the index based on them, and Gary Illyes said searching by image or video was not possible before transformers (both said at the event). On Day 2 he said the media indexer attaches each image and video to the URL of the page that hosts it, that Googlebot-Image rules or a disallowed file location keep images out, that noimageindex also stops videos because their thumbnails are images (not in Google's docs), and that max-image-preview:large can make content do surprisingly well in Discover. On AI-generated images Google said the choice is up to the site owner, that Google shows them when users look for them, and that diffusion models are poor at text inside images, so sites should check them for hallucinations and regenerate where needed; Google's image metadata guide supports IPTC digital source types and C2PA details for AI-made images. In Lightning session F a community speaker showed LLM-rewritten alt text going wrong: for a cat anxiety medicine the model described only a cat and a vet, dropping the page's intent, and the site's image performance in Search Console fell; the safer script adds the page title to the image description and is tested on a few URLs first. Author’s view: visual search predates transformers in Google's own history (Google Images 2001, Lens 2017), so the remark is about today's multimodal AI search; and do not set noimageindex on video watch pages.

  Day 1Day 2Day 3

• 24 claims · 5 sessions

  Accessibility and ARIA

  Attendees were told on Day 1 to check ARIA and accessibility, and Google's guide for generative AI features says browser agents may read a page through screenshots, the DOM and the accessibility tree, and recommends semantic HTML because it also helps screen reader users. Day 2 added that Gemini in Chrome relies heavily on the screenshot it takes of a page (the Gemini in Chrome detail was said at the event). Day 2 also tied accessibility to images and translation: alt text should describe an image for someone who cannot see it, and Google's image guide calls alt the most important attribute for image metadata; John Mueller said he dislikes notranslate because automatic translation is a form of accessibility, and that the meta tag named google also turns off Chrome's translation (not in Google Search's current docs; Chromium's Translate design document describes the same opt-out). A community speaker said embedding applications in iframes brings accessibility, layout and navigation problems. Author’s view: Google has not called ARIA a ranking signal; the case for accessible markup is that assistive tools, browser agents and Google's image understanding read the same page, even if some agents lean on screenshots. A second recording of Day 1 captured a community talk on preparing sites for agents: agents combine a screenshot, the DOM and the accessibility tree, so a div with no semantic HTML or landmarks has no purpose an agent can tell; the speaker recommended landmark elements (header, nav, main, article, footer), a logical heading order and ARIA attributes such as role=button, aria-label and aria-pressed on custom controls, warning that no ARIA is better than bad ARIA. Google's SEO Starter Guide says semantic heading order helps screen readers but does not matter to Search. On Day 2 a community speaker showed LLM-written alt text that described only what a model could see, a cat and a vet, and dropped the page's meaning, and advised testing on a few URLs and giving the model the page's title.

  Day 1Day 2

• 47 claims · 8 sessions

  Why structured data still matters

  Day 2's structured data talk argued that markup still matters even though models can read pages, giving four reasons (said at the event, not in Google's docs): precision, since LLM extraction falls short of the very high accuracy, around 99.9% as the speaker put it, that Google needs and mishandles nested schemas such as sale pricing; extra content, such as full ISO dates or stable identifiers; efficiency, since parsing markup is nearly free while even Google cannot run complex models on every page, so Google will always prefer it; and focus, since markup points at the main item instead of, say, prices of related products. Google said the same cleaned, quality-checked data feeds both classic results and AI Overviews and AI Mode, and that raw schema.org markup is generally not turned into text and put into a model's context (the data is first sorted, quality-checked and indexed), partly because that would be an abuse vector (not in Google's docs). This extends Day 1's documented point that no special markup is needed: Google's generative AI guide lists overfocusing on structured data among things you don't need to do but still recommends it for rich result eligibility (features such as review stars and recipe filters), where Google located its proven value, more qualified traffic; a May 2018 case study reported 2.7 times more search traffic and 1.5 times longer sessions for Rakuten Recipes. Google's guidelines still say not to mark up information that users cannot see, even if it is accurate. Author’s view: the 'extra content' argument therefore covers the machine-precise form of visible facts, such as a full ISO date for a shown event date, not hidden content, and dropping markup because AI reads the page anyway is a mistake for prices and offers. Day 3 drew the line for AI features: AI Overviews and AI Mode are standard search features, not rich results, and need no structured data, as Google's AI features guide says; Google added that if AI Mode one day showcases elements such as recipes, those parts would need it. Author’s view: markup is therefore not a condition for appearing in AI features, but data Google takes from it can still reach them. Google Shopping said AI systems need rich data with light structure rather than deep nesting, and that some structure is much better than none (said at the event); it repeated that web markup is an efficient, unambiguous way to share product data. Day 1's lightning talks, captured by a second recording, added the community debate: one speaker said marking up which number is the price saves an AI agent from guessing, while another said a study linking JSON-LD to AI visibility was followed by a controlled study that found no benefit, the markup being a sign of teams doing everything else right. A second recording of Day 2 confirmed that a full ISO 8601 date helps where Google might not detect a visible date's local time zone, which matters for event extraction. Author’s view: use markup for rich-result eligibility and exact data, not as an AI-visibility lever.

  Day 1Day 2Day 3

• 61 claims · 5 sessions

  Implementing structured data

  Google accepts JSON-LD, microdata and RDFa and interprets them identically downstream; it recommends JSON-LD because a single block leads to fewer mistakes, and said microdata can save payload by not duplicating page content (not in Google's docs). Pick from the Search gallery only the types relevant to the page that Google actually consumes, since Google uses subsets of schema.org and very generic types say little; extra markup is never penalised, but describing every semantic detail is probably not worth the effort. Test a few pages in the Rich Results Test before templating, check that CMS plug-ins emit good markup and that several plug-ins do not emit the same type (one of the most common problems, not in Google's docs), and use unique identifiers such as an organisation's homepage URL; irrelevant markup can be filtered or, in egregious cases, earn a manual action that removes rich result eligibility without affecting web ranking. Google announced, with no date and not yet in its docs, downloadable SHACL validation rules for each feature guide so markup can be checked in a site's publishing pipeline before release; they will arrive gradually and will not replace Search Console reports. Google's documentation also recommends the same structured data on all duplicates of a page, not only the canonical, and says AI-generated markup must be fact-checked and validated before publishing. Day 3 showed worked examples: Google's Article slide reproduced the NewsArticle JSON-LD example from its documentation (headline, three image ratios, datePublished and dateModified with time zone offsets, Person authors with name and url), Article markup is highly recommended but not required for Top Stories, and WebSite markup on the home page lets a site indicate its preferred site name. Basic Product markup separates product data from offer data, and an offer can reference shipping and return policies defined once under Organization through @id, as Google's merchant listing documentation recommends. Google Shopping showed product Q&A marked up as Question items linked through subjectOf, at the cost of bigger markup, and Google said it usually picks up structured data changes within hours to one or two weeks (both said at the event, not in Google's docs). Day 1's lightning talks, captured by a second recording, showed an automated variant: a script that runs pages through the Rich Results Test, using its URL mode for public pages and its code mode, with the page's HTML pasted in, for pages Google cannot fetch (best reading of a largely unintelligible recording; Google's help documents both modes and says resources behind a firewall or password are not available to a URL test), and an LLM fix loop that gives a large model the markup, the reported errors and Google's documentation, rewrites the JSON-LD and retests, at most three times; the demo treated warnings as non-critical, matching Search Console Help, which separates critical issues that make an item invalid from non-critical ones. The problems reported on the demo's test page included an empty name, a breadcrumb problem and a price of zero (best readings of the recordings), and Google's merchant listing documentation lists the product name as required and requires a price greater than zero for merchant listing experiences. A second recording of Day 2 confirmed that a full ISO 8601 date in markup helps where Google might not detect a visible date's local time zone, as for event extraction. Author’s view: a fix loop that stops when Google's test passes proves only that the markup is valid, not that its values are true, so fact-check AI-generated markup as well.

  Day 1Day 2Day 3

• 17 claims · 2 sessions

  Schema.org vocabulary

  Schema.org is the shared vocabulary that Google, Bing and Yahoo! announced in June 2011 so that every consumer reads markup the same way, and it is an open collaboration anyone can join. Google uses it as its markup language but mostly consumes only the subsets its developer documentation declares for its features, and very generic types say too little to be useful. Google said (not in its docs) that schema.org began publishing bucketed usage statistics for every type and property in 2026, also on GitHub, to help break a chicken-and-egg problem: Google often needs to see adoption before it builds a feature, which extends Day 1's advice to request features publicly and in volume. Schema.org also added RDF lists and sets for ordered values, and in the Google speaker's own tests LLMs asked to write schema.org markup often invent properties, get nested schemas wrong and duplicate content. Author’s view: markup for types Google does not use yet is a bet on future features, so check a type's usage bucket first. Day 3 tied schema.org to shopping: Google said product markup is a source of billions of product offers next to merchant feeds, and that it works to make what its shopping feeds and the Universal Commerce Protocol can express expressible in schema.org too. Schema.org version 30.1 (16 September 2026) added retail properties such as specification for Product and valueGroup for PropertyValue, but on 3 October 2026 Google's documentation did not cover them and Merchant Center help mapped only item_group_title to a schema.org property, so reading the new conversational attributes from markup rests on what was said at the event.

  Day 2Day 3

• 36 claims · 3 sessions

  Product and merchant markup

  Google said shopping is where structured data's precision matters most: LLM extraction gets complex pricing such as sale prices wrong, and markup stops its systems from picking up prices of related products (said at the event, not in Google's docs). In 2025 Google added structured data support for merchant loyalty programs and shipping policies, set at organisation level with product-level details. On stage, validity dates for sale prices were presented as new support added in the months before the event, so merchants need not rush to remove a sale price when a sale ends; Google's documentation updates log calls the July 2026 change a clarification, a Sale duration section explaining validFrom with validThrough or priceValidUntil, aligned with Merchant Center's sale_price_effective_date. Google's guide to pausing an online business recommends staying online and updating Product markup to show current availability. Day 3 brought the promised shopping detail: product rich results and Google Images product annotations come from product structured data without Merchant Center, while the Shopping tab needs a Merchant Center feed; product snippets draw on both, Google's documentation recommends feed and markup together for the widest eligibility, and Merchant Center can use the markup to fix mismatched feed data. Basic product markup has product data (name, description, brand, images) and offer data (URL, condition, availability, price), and an offer can reference shipping and return policies defined once under Organization through @id, as Google's merchant listing documentation recommends. Google Shopping said most of its new conversational attributes can also be read from schema.org markup, and schema.org added properties such as specification and valueGroup in September 2026, but Google's documentation did not cover this markup on 3 October 2026, so the feed is the documented route. 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; a Day 1 community demo's test page was flagged for an empty name and a price of zero.

  Day 1Day 2Day 3

• 46 claims · 7 sessions

  Video in Search

  Video was covered on Day 2: Google's feature extraction takes each video and the data around it from the page and passes them to its media indexing engine, and videos can appear in the main results, video-specific tabs and Discover, with features such as key moments and previews. Google's slide listed seven key factors (high-quality content, a dedicated watch page, compelling titles and descriptions, relevant thumbnails, video markup, fast-loading pages and sitemap inclusion), and its video SEO guide adds that a video must be embedded on an indexed watch page to be eligible for video features, though a non-watch page can still show as a text or Google Images result. Google strongly suggested describing video metadata with JSON-LD structured data on top of placing the video in an HTML element, and warned that paying for a generative AI subscription does not make AI-generated videos succeed. The max-video-preview rule caps preview length in seconds (0 allows only a static image, -1 sets no limit), though John Mueller said he thinks -1 will not bring hour-long previews. The audience figures shown for Southeast Asia (over 40% of shoppers rely on videos for purchase decisions; over 219 million people use YouTube daily, a scope heard in two independent recordings but unverified; over 150 streaming apps) were said at the event and are not in Google's docs; on Day 1 a slide had listed language and text from speech as video ranking signals. Day 3 added that video results in web search take their thumbnails and attribution from the page or its markup, and Google said its deep analysis of a video, needed to actually understand it, can take months (said at the event, not in Google's docs). A second recording covered the rest of the Day 2 video talk. Gary Illyes said a video without a dedicated watch page, or without a video container in the HTML, is not indexed, and that a video placed below the fold is not indexed either; Google's video documentation only requires the player to be present at load, at a size and position Google can determine. He said a wrong thumbnail makes an extremely high share of viewers drop out at the start, that the 'fast-loading' factor really means the video itself must load fast (a CDN helps), that Google supports the popular formats and recommends MP4 with standard codecs, and that descriptive text around a video helps Google rank and retrieve it. Video sitemaps are not critical but good to have, because Google ingests them much more often than it can process HTML pages and they tell it which URLs to check; hosting on YouTube or Vimeo is worth considering because they solved video search years ago, and the more people talk about a site's videos the more likely Google is to surface them (most of this said at the event, not in Google's docs). Googlebot-Video rules or a disallowed file location keep videos out, a disallowed video is not shown by its bare URL, and noimageindex also blocks videos through their thumbnails. Author’s view: put the main video of a watch page in the first viewport without a click-to-play placeholder, and do not set noimageindex on watch pages. On Day 1 the opening keynote, in an audio recording, said YouTube has surpassed Netflix as the number one streaming service, in the US and around the world. Author’s view: YouTube's published figure is US streaming watch time measured by Nielsen, so cite the lead that way.

  Day 1Day 2Day 3

• 57 claims · 2 sessions

  Shopping data: feeds, markup and Storebot

  Product rich results and Google Images product annotations come from product structured data without Merchant Center, but the Shopping tab needs a Merchant Center feed (said at the event; Google's docs recommend feed and markup together for the widest eligibility). Because AI Mode answers detailed follow-up questions, Google added a few flexible Merchant Center attributes in 2026 instead of hundreds of narrow ones: question and answer, document link, related product, item group title, variant option and popularity rank; it now also stresses product highlights and key-value product details, and its help marks most of these as intended for conversational experiences such as AI Mode. Data quality matters more for AI and agents ('garbage in, garbage out'). Google Shopping crawls with Storebot-Google, which recrawls more often for prices and availability, validates feeds and crawls cart and checkout pages. Google said most new attributes can also be used from schema.org markup, but its Merchant Center help lists a schema.org property only for item_group_title. Google announced the conversational attributes in January 2026 and opened them to retailers globally at Google Marketing Live on 20 May 2026; in testing with lululemon, brand-submitted attributes were used in 50% of relevant AI Mode product recommendations (Google's September 2026 post). Google's help sets the limits: up to 30 question-and-answer pairs per product, up to five PDF document links, and a popularity rank that is the merchant's own 0-100 ranking against its inventory, not a sales figure across shops. This builds on Day 2's merchant markup news, such as sale-price validity dates aligned with Merchant Center.

  Day 3