Google's new AI search guide: what changed in 2026

Google published its guide to optimizing for generative AI features on May 15, 2026 and updated it in July. The document is written for visibility in AI Overviews and AI Mode, but its central instruction is to keep doing SEO: generative features retrieve pages from Google's Search index through the same core ranking and quality systems. A page must be crawlable, indexed, eligible to show a snippet, and useful before it can support a generated answer. This post stays with Google's own guide; for the wider framework, including what SEO, GEO, and AEO each mean and where they overlap, read the framework briefing.

The guide matters because it turns several disputed industry claims into official yes-or-no guidance. Google wants useful, expert-led source material; a clear technical structure; high-quality images and video; accurate Merchant Center or business data where relevant; and measurement in Search Console. It says brands do not need a separate AI version of their content, tiny artificial content chunks, a special schema type, inauthentic mentions, or an llms.txt file for Google Search.

Google describes two techniques behind its generative results. Retrieval-augmented generation grounds an answer in up-to-date pages retrieved by the core Search systems. Query fan-out runs multiple related searches to assemble the information needed for a more complex question. Neither technique removes the need for ordinary search eligibility. It expands the number and shape of queries through which a useful page might be retrieved.

For a site owner, that keeps the order of operations familiar: make the important URL accessible, return a valid response, expose the main content and links, use a consistent canonical, avoid accidental noindex rules, and give the visitor a good experience on the device they actually use. Google can render JavaScript, but a visually sophisticated site still needs its meaning and navigation to survive the rendering path.

The guide repeatedly separates original, experience-based work from material that merely restates what already exists. Google's examples favor a first-hand review, expert interpretation, distinctive evidence, and a real point of view over a generic tips article that could have been assembled by any publisher or model. This is more specific than the familiar instruction to create 'quality content': it asks what the page contributes that a summary of the current results cannot.

For a beauty brand, non-commodity material can include a documented wear test with conditions and limits, an original shade or texture library, an accountable formulator or founder explanation, a claims-safe translation comparison, customer-service questions organized into a decision guide, or first-party data from sampling and reviews. The point is not to expose confidential information. It is to publish useful evidence that comes from the brand's actual product and market experience.

Google now calls out high-quality images and video directly because generative results can surface media as well as links. That supports a broader change in search behavior: a useful demonstration, diagram, swatch, comparison, or product video is not an ornamental add-on to the article. It is another retrievable expression of the evidence, provided the media has descriptive context and follows the existing image and video SEO guidance.

The guide also gives ecommerce and local data a concrete role. Merchant Center feeds, product information, and Google Business Profiles can supply product listings and business facts in generative and traditional results. On a commerce site, visible price, availability, variants, shipping, returns, structured data, and the merchant feed should agree. Structured data remains valuable for established result features, but Google says there is no special schema required for generative search.

The mythbusting section is the clearest change for planning. Google says llms.txt has no positive or negative effect on Google visibility; there is no required 'chunking' pattern or ideal page length; content does not need to be rewritten in a special style for AI; exact long-tail variations do not each need their own page; inauthentic mentions are not a durable shortcut; and there is no special generative-search schema.

This does not mean formatting, structured data, third-party coverage, or machine-readable files are always useless. Clear sections help readers. Supported schema can create eligibility for established features. Authentic independent coverage can introduce and corroborate a business. Another system may choose to read an llms.txt file. The correction is narrower: none of those becomes a special Google AI-ranking lever simply because it is labeled for GEO or AEO.

The guide arrived inside a wider set of Search changes. Google clarified on June 15 that llms.txt neither helps nor harms Google visibility. It removed FAQ rich-result documentation because the feature stopped appearing on May 7, 2026. FAQ content can still help a customer and strengthen a page; brands should simply stop treating FAQPage markup as a route to extra result space.

Google also added a Generative AI performance report for Search and Discover, reporting for social and video platform content, and more product-related structured-data support. These updates point toward one connected information system: editorial pages, product pages, feeds, images, videos, social publishing, and business facts should support the same accurate source rather than operate as separate SEO tricks.

Google's guide introduces agentic experiences as an emerging area. A browser agent may inspect screenshots, the DOM, and the accessibility tree while comparing specifications or completing a task. Google also points to Universal Commerce Protocol as an emerging way for Search agents to do more. This is not a direction to rebuild a site around an unfinished protocol. It is a reason to keep product facts, controls, states, and policies explicit and machine-interpretable.

The same practices help people today: semantic controls, stable labels, visible prices, accessible product options, dependable inventory and shipping information, and clear confirmation states. Agent readiness is strongest when it grows out of a usable commerce experience rather than a separate machine-only layer.

Use Google's Generative AI performance report to understand discovery through its generative features, then connect that visibility to the site's own qualified sessions and outcomes. Search Console is still a sampled diagnostic surface rather than a complete business attribution system, and third-party visibility tools do not have access to Google's internal ranking or AI systems.

A useful scorecard joins technical eligibility, query and page visibility, branded demand, cited or linked mentions, engaged sessions, product discovery, leads or orders, assisted conversion, and the pages that create repeat value. The objective is not the largest possible count of impressions inside a new report. It is to learn which original sources cause the right customer to discover, trust, and choose the business.

Start with the current site, not a speculative AI microsite. Fix access and indexation; identify the customer decisions the existing pages fail to support; add original evidence and useful media; synchronize product and business facts; and distribute the work through authentic channels. Then use Search Console and commercial analytics to decide what deserves another round.