What Google really says about AI-optimised websites

Google has published an official guide for the first time on how websites should be optimised for AI features – AI Overviews, AI Mode, Agentic Search. Not from SEO agencies, not from tool vendors: from Google itself. This post summarises what is in it, what it means – and what you can ignore.

How an AI Overview is created

Before you optimise, you need to understand what is happening under the bonnet. AI Mode does not work like classic search. Google describes the mechanism explicitly – and it explains why classic SEO is not becoming irrelevant, but more central than ever.

Step 1: Query fan-out. When someone asks a question, the model breaks it down into several parallel sub-queries. One concrete example from the guide: Someone searches for “how to fix a lawn that’s full of weeds”. The system turns that into “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn” at the same time – and runs all these searches in parallel. Not one search, but many.

Step 2: RAG – retrieval from the search index. The sub-queries hit the same index Google uses for classic search. RAG (Retrieval-Augmented Generation) is the technical term for this: the model grounds its answer in real, current web pages instead of hallucinating from training data. Google also calls this “grounding”. What matters is the index – if you are not indexed, you do not exist for AI Mode.

Step 3: Extraction and synthesis. The model reviews the pages it finds for specific passages that match the individual sub-question. It then synthesises an answer from them – and links prominently to the source pages. These source links are clickable and visible. If you are cited, you get traffic.

That is the core: AI Overviews are not standalone AI answers, but syntheses from indexed web pages. If you get into the index and provide the best answer to a sub-question, you will be cited. The ranking signal is the same as always – only the competition for visible placements has become tighter.

What you should do

1. Keep the technical base clean

Google is clear on this: the technical requirements for AI features are the same as for classic SEO. Crawlability, indexing, page experience.

• Pages must be crawlable (no noindex, no robots block for Googlebot)

• Load times matter – Core Web Vitals remain relevant

• HTTPS is mandatory

• Mobile first: AI Mode is often used on mobile

Nothing new – but if you were sloppy here, AI features will punish you twice.

2. Unique content with a clear point of view

This is the core recommendation. Google calls it “non-commodity content”. What it means: content you cannot get everywhere.

• First-hand experience, data, case studies

• Clear opinions instead of neutral summaries

• Specific expertise that is not on Wikipedia

• Concrete recommendations for action instead of vague advice

Generic content may be recognised as a source by AI systems – but it will not be cited. Those who add something original will be cited.

3. Structured, clear writing

AI systems extract information from text. The clearer the structure, the easier the extraction. Google recommends using the H1/H2/H3 hierarchy consistently, making concrete statements instead of filler, putting definitions at the start of sections, and adding context to facts (not “this rose by 30 per cent”, but “revenue rose by 30 per cent in 2025 compared with 2024”).

4. Include images and video

One point many overlook: AI Overviews and AI Mode can embed images and videos directly – not just text links. These are separate entry points. If you have relevant, high-quality images with clean alt text and videos with structured data, you can appear in AI features even if the text content is not cited. Google recommends supporting text content with image and video material where it makes sense, and following existing image and video SEO best practices.

5. For shops and local businesses: Merchant Center and Business Profile

AI answers can embed product listings and local business information directly. If you have a Merchant Center feed or maintain a Google Business Profile, you are automatically in the race for these placements. Google also mentions “Business Agent” here – a conversational experience in Google Search, where customers can chat directly with a brand. Relevant for e-commerce and local service providers, not for pure content sites.

What you can ignore

llms.txt is not needed. Google does not use this file for AI features. If you spend time maintaining an llms.txt, you are wasting it.

No special chunking. There is no special document structure for AI crawling. Google processes normal HTML content.

No “AEO” or “GEO” as separate disciplines. Google says clearly: there is no separate optimisation strategy for AI features. If you do good SEO, you are already set up well.

No artificial mentions. Some agencies recommend “placing” brands on third-party websites so that AI mentions them more often. Google explicitly says this does not work and breaches the guidelines.

No special structured data for AI. Schema.org remains useful for existing use cases. New schema types “for AI” are not needed.

Do not rewrite content for AI. Google says explicitly: AI systems understand synonyms and general meanings. You do not need to force long-tail variants, cover every possible fan-out query, or write in a special “AI style”. If you write content for real people, you are already set up correctly.

AI-generated content is not banned – but not without review. If you use AI tools to create content, you must ensure the output meets Google Search Essentials and spam policies. AI content itself does not break the guidelines. Mass-generated, low-quality AI content does.

Agentic experiences: the next step

AI agents do not just navigate search results – they carry out actions on websites. Reservations. Form entries. Product searches. Bookings.

For web apps and tools, that means: the DOM structure must be clean. AI agents read the accessibility tree, not the visual layout. Missing labels, unclear button text and missing aria attributes make an app unusable for agents.

UCP (Universal Control Protocol) is a new open protocol proposed jointly by Anthropic and Google. It allows agents to interact with web apps in a structured way – similar to how APIs provide machine-readable endpoints.

That is not relevant for everyone – but for SaaS products, booking platforms and e-commerce tools it will matter in 12–24 months.

What remains

Google’s recommendations are refreshingly simple – and that is the real message. No secret. No new protocols. No expensive specialist tools. If you have been doing good SEO for years – technically clean, strong on content, with your own point of view – you are ready for AI features.

The only real difference from classic SEO: the competition for citations in AI answers is tougher than the competition for rank 3 in search results. Because AI answers only name a handful of sources, the sources really do have to be good.

Source: Google Search Central – AI Optimization Guide (as of 15 May 2026)