01Notes · 24 August 2026
What Google Actually Says About AI Optimisation
Quick answer
Google published guidance on optimising for its generative AI features in May 2026 and updated it in July. It is short, it is public, and it names three things as unnecessary that a large part of the AI-visibility market currently sells.
The one-line summary: this is still SEO. Generative features run on Google’s core ranking systems, so the optimisation is the work you already know.
Key takeaways
- llms.txt is named as unnecessary. No special treatment. It is a file you can create in ten minutes, which is exactly why it sells.
- Content chunking for machines is unnecessary, and it makes pages worse for people.
- No AI-specific schema exists or is required. Independent research found no schema type with a consistent citation advantage.
- Spam policies now cover generative responses, which puts bought links at risk in a second place.
- Technical health matters more, not less, if generative features run on core ranking.
- Ask any proposal containing these three what it is based on. The answer tells you about the firm.
The guidance, in Google’s own words
The document is not long and it is worth reading in full rather than through a summary, including ours. The load-bearing points:
- Generative features run on core ranking systems. The same crawling, indexing and ranking that produce search results produce AI Overviews and AI Mode.
- There is no separate optimisation track. Good SEO is the optimisation.
- llms.txt receives no special treatment. Stated directly.
- Content does not need chunking for machines. Write for people.
- No AI-specific structured data exists. Mark up what is visibly true.
Sold anyway, by most of the market
Put those two lists side by side and the picture is uncomfortable.
| Tactic | How it is sold | What Google says |
|---|---|---|
| llms.txt implementation | "Tell AI models how to read your site" | Unnecessary. No special treatment |
| Content chunking | "Restructure pages so models can parse them" | Not required |
| AI-specific schema packages | "New markup for AI search" | Does not exist |
| Guaranteed AI rankings | "Get cited in ChatGPT, guaranteed" | Generative answers are non-deterministic. Nobody can guarantee this |
| High-volume link buying | "Authority at scale" | A link scheme, now in scope for AI responses too |
This is not a claim that everyone selling these is acting in bad faith. Some are repeating what they read. But a proposal containing all five is not describing work that has been checked against the primary source.
Where we think Google is incomplete
We agree with the optimisation position and we do not think it is the whole story.
Google is describing its own surfaces. ChatGPT, Perplexity and Gemini run their own retrieval over their own corpora. In 2026 research the two most similar platforms shared only 24.1% of cited pages. Advice that is correct for AI Overviews does not automatically transfer.
And nothing in the guidance addresses measurement. Search Console will not tell you whether an assistant recommends you, which competitors it names instead, or whether what it says about you is accurate. That gap is real, it is new, and it is where we think the actual work is — our method is published.
A better approach
Unglamorous, and it is what the evidence supports:
- Check crawler access first. OAI-SearchBot controls ChatGPT Search discovery; GPTBot is a training crawler. Confusing them is the most common self-inflicted problem we find.
- Answer in the first 100 words. Pages doing this were cited at 22.1% against 8.7% otherwise: the strongest structural signal measured.
- Serve query families, not keywords. Pages covering eight or more related queries reached the AI top ten 43.3% of the time against 15.5% for single-query pages.
- Fix your entity facts. Being described wrongly is more common than being absent, and cheaper to fix.
- Measure per platform. A blended score hides the finding.
What this means for you
- Read the primary source before buying a programme. It takes fifteen minutes and it disqualifies a lot of proposals.
- If a proposal contains llms.txt, chunking and AI schema, ask what it is based on.
- Do not let this convince you AI search does not matter. The optimisation is familiar; the measurement is not.
02Common questions
Questions this raises
Does Google say llms.txt helps?
It says the opposite. The guidance names llms.txt as unnecessary for its generative features, alongside content chunking and AI-specific structured data. Three things a large part of the market sells.
Is optimising for AI features different from SEO?
Google's position is that generative features run on its core ranking systems, so the optimisation is the same work. Our position is that this is right about optimisation and incomplete about measurement, which is genuinely new.
Do spam policies apply to AI answers now?
Yes, explicitly. Scaled content abuse and link schemes are in scope for generative responses, not just search results. That makes a bought link profile a liability in a second place.
Should I stop doing technical SEO then?
The reverse. If generative features run on core ranking systems, technical health matters more rather than less. The cheapest finding in most audits is still a blocked crawler.
Why do agencies still sell these tactics?
Because they are easy to explain, easy to invoice and hard for a buyer to verify. Naming a file you can create in ten minutes sounds like a deliverable.
Sources
Every figure on this page comes from one of the following. Where a claim is ours rather than a source’s, it carries an evidence label instead.
- Google Search Central — Optimizing your website for generative AI features, published 15 May 2026, updated 10 July 2026. Primary source for every position quoted in this post. Primary
- Google Search Central — spam policies, now explicitly covering generative AI responses in Search. Primary
- Rankability — State of AI Search, 1,645 citation observations across 31 topics, 2026. Source for the finding that no schema type showed a consistent independent citation advantage. Independent study
Ask an assistant
Hand this page to an assistant and ask it to summarise or challenge the argument. If it gets something wrong about us, that is itself a useful finding.
→Next step