What is the Google AI optimization guide?
The Google AI optimization guide is a Search Central page, last updated on 10 July 2026, that lists the tactics Google Search ignores in AI Overviews and AI Mode. Hold the GEO proposal on your desk next to it and you can see which priced lines Google says you do not need, before you pay for them.
"From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
Google Search Central, AI optimization guide, last updated 10 July 2026
Google is an interested party, but its "Mythbusting" section is specific enough for you to test a pitch against it. The guide says nothing about tables, interactive tools or answer-first openers, so no vendor can cite it to you for those. It covers Google Search only, so your ChatGPT and Copilot questions need other evidence. Our visibility optimisation work starts by scoring the proposal.
Which GEO pitch claims does Google's guide contradict?
Nine common proposal lines run into the guide: you can ignore five, Google warns against two, and two carry risk. The points column feeds your scorer below.
| Line in your proposal | What Google's guide says | Your verdict on Google | Points |
|---|---|---|---|
| llms.txt, AI text files, Markdown copies | "Google Search ignores them" | Not needed | 1 |
| Chunking into AI-sized blocks | "There's no requirement to break your content into tiny pieces" | Not needed | 1 |
| An ideal word count for AI | "There's no ideal page length" | Not needed | 1 |
| AI-friendly, long-tail rewrites | "You don't need to write in a specific way" | Not needed | 1 |
| A dedicated AI schema layer | "no special schema.org markup you need to add" | Not needed for AI | 1 |
| Fake or bought brand mentions | seeking inauthentic mentions "isn't as helpful as it might seem" | Not helpful; spam risk is our judgement | 3 |
| A page for every query variation | Doing so "primarily to manipulate rankings or generative AI responses" breaches the scaled content abuse policy | Spam exposure if pages add no value | 3 |
| A tool reading Google's internal AI signals | "No third-party tool has access to our internal ranking or AI systems" | Contradicted | Gate |
| Guaranteed AI Overview citations | "Be wary of third-party tools that promise ranking success" | Google warns against | Gate |
Google does not name fake mentions as spam; we weight them 3 because Google says seeking them "isn't as helpful as it might seem" and they carry reputational risk. Genuine directory listings and coverage are ordinary work and score 0 in our scorer. Google also says "Indexing and serving aren't guaranteed", even for a compliant page.
On llms.txt, Google's statement closes the question for Google: the file is "completely fine" and will "neither harm nor help" your visibility. Our llms.txt post covers the logs and studies. No page we read documents it for the other engines.
Score a GEO proposal against the guide
Tick every line your proposal makes, and the scorer returns a band: CLEAN, NEGOTIATE or WALK AWAY. It is our illustrative heuristic, and it covers only the Google half.
Score = (not-needed lines × 1) + (spam-exposure lines × 3). A ticked gate line (internal signals or guaranteed citations) returns WALK AWAY directly. Otherwise 0 is CLEAN, 1 or more is NEGOTIATE, and 4 or more including a risk line is WALK AWAY. Not-needed lines alone never reach WALK AWAY.
3 points: NEGOTIATE. 3 lines are work Google says you do not need; none is spam. Ask for them to be struck or priced as optional extras, then score again.
Made-up proposal, round numbers. Suppose you run marketing at a Malaysian insurer and a 12-line GEO proposal lands on your desk. Eight lines are ordinary SEO. Three score 1 each: an llms.txt, a chunking rewrite, an "AI schema" layer. One scores 3: a programme of planted forum posts and paid listicle slots. Score 6 with a risk line, so WALK AWAY, although eight of the twelve lines are sound. Four are the ones you strike.
If a GEO proposal is waiting for your signature, we score every line against Google's guide and return a marked-up copy with the lines you should strike. Send us your proposal and we reply with your marked-up copy.
What does Google say does move AI visibility?
Our reading of Google's guide: four core asks, none new, grouped our way. They are an indexed, snippet-eligible page, inclusion in AI features, a unique viewpoint, and Merchant Center and Business Profile data.
| Core ask, our grouping | Google's wording | Your check this week |
|---|---|---|
| Your page, indexed and snippet-eligible | "indexed and eligible to be shown in Google Search with a snippet" | Run your ten most commercial URLs through URL Inspection; search your templates for nosnippet and max-snippet |
| Your site's inclusion in AI features | "included in Search generative AI features in Search Console" | Our proxy: your Search Console generative AI report appears for you |
| Your unique viewpoint | "a unique viewpoint that stands out" | Add your original evidence and your named position |
| Your Merchant Center and Business Profile | "can help your products and services to be visible in both AI responses" | Our advice: keep your feeds and profile accurate and matched to your site |
| Your structured data (separate note, rich results only) | "helps with being eligible for rich results on Google Search" | Keep your Organization, Product and Article markup current; drop the AI label |
The guide also asks for headings and paragraphs that give a clear structure, crawlable content, a good page experience, supporting images and videos, and semantic HTML where possible. If you run a regulated site, check snippet eligibility first: Google's AI features page lists nosnippet, data-nosnippet and max-snippet as controls that limit what AI features show. Hypothetical example, not a client case: an insurer that added nosnippet to protect disclaimers would have pulled those pages from AI Overviews by its own hand. Your markup serves Google's regular results.
The Search Console generative AI report defines impressions as "how many times links to your site were shown to a user in a generative AI feature on Google Search". It mentions no clicks, position or CTR; "As of August 31, 2026, we've rolled out these insights to all websites worldwide." A vendor promising you an AI traffic number from it has not read it.
Does the guide apply to ChatGPT, Perplexity and Copilot?
No. The other engines have published crawler rules and some advice but no equivalent list of tactics that do not work, so you will find thinner evidence on both sides of a GEO claim about them.
| Engine | Retrieval basis documented | Sources you can check | Published advice |
|---|---|---|---|
| Google AI Overviews and AI Mode | Core ranking and quality systems; indexed, snippet-eligible page | Search Console impressions report; preview controls you set, such as nosnippet | Mythbusting list; headings, paragraphs, images, video |
| ChatGPT search | OAI-SearchBot surfaces sites; GPTBot trains models separately | Sites you opt out are not shown in answers, may appear as links | None found |
| Perplexity | PerplexityBot surfaces and links sites; not used for model training | Numbered citations; whole-site source labels from a review | Review asks about corrections, bylines, news kept apart from ads |
| Microsoft Copilot and Copilot Chat | "may fetch information from the Bing search service" with web search on | Bing Webmaster Tools AI Performance you can open; launch post titled public preview | "Clear headings, tables, and FAQ sections"; IndexNow |
OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features"; opted-out sites "will not be shown in ChatGPT search answers, though can still appear as navigational links". PerplexityBot "is designed to surface and link websites in search results on Perplexity". Microsoft's Learn page says Microsoft Copilot and Copilot Chat "may fetch information from the Bing search service" when web search is enabled. Microsoft's own advice in the Bing post: "Clear headings, tables, and FAQ sections help surface key information".
Pick the engine you care about and what the vendor is selling you. The checker returns what the documentation supports.
Google says it does not need new machine readable files, AI text files, markup or Markdown, no special schema.org markup and no chunking into tiny pieces. Strike the line or re-price it as hygiene.
OpenAI's crawler page tells you how to get eligible (allow OAI-SearchBot) and what opting out costs you, and nothing about ranking once you are eligible. If you see a line pricing "ChatGPT ranking factors", you are being sold something OpenAI has not published.
Perplexity's crawler documentation covers eligibility. Its source-labels page adds a site-level review, says partnerships and payments do not affect a label, and asks "Does the site correct its mistakes?", "Does it say who wrote each piece?" and "Does it keep news separate from advertising and opinion?" Your corrections policy and bylines serve those; a mentions programme does not.
Independent tests, none from Google, point the same way on schema and llms.txt, and tell you retrieval comes ahead of page features.
| Source | What was tested | Result | What you should weigh |
|---|---|---|---|
| Ahrefs, 11 May 2026 | 1,885 pages adding JSON-LD against 4,000 controls, 30 days each side | "No major uplift in citations on any platform"; AI Overviews -4.6% | Pages you would expect to be cited anyway |
| Ahrefs, 15 Jun 2026 | 137,210 domains in May 2026; about 28% had an llms.txt | 97% of those files received zero traffic | Its customers skew technical; treat it as an upper bound for you |
| arXiv 2610.15033, 10 Oct 2026 | Queries replayed through engine APIs | Retrieval explains 32% of citation variation, page properties 2.8% | Preprint, not peer reviewed |
| Same preprint | Pages within one website | Pages whose titles and paragraphs match the question are cited more | Same preprint caveat |
AI Performance in Bing Webmaster Tools is "a new set of insights that shows how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations."
Krishna Madhavan, Meenaz Merchant, Fabrice Canel and Saral Nigam, product managers at Microsoft AI, in Microsoft's launch post, 10 February 2026
Google denying a tactic settles it for you on Google. OpenAI, Perplexity and Microsoft saying nothing settles nothing. For those engines, accept only claims the vendor will test on your prompt set with a before-and-after reading, priced as an experiment you can stop. Our citation measurement stack sets out the protocol.
How to buy GEO work after reading the guide
You buy the Google half as SEO with a new report, and the rest as measured experiments.
| Your contract line | Your acceptance test |
|---|---|
| Google eligibility: indexed, snippet-eligible, in AI features, no stray nosnippet | URL Inspection shows your target pages indexed, with no stray nosnippet |
| Your viewpoint, not volume | Your original evidence and your named authors; you never pay per "pages per month" |
| ChatGPT, Perplexity, Copilot: one line each | The vendor's test design, attached to your contract so you can rerun it |
| Any dashboard you are shown | Search Console impressions, Bing Webmaster Tools citations and a prompt set you own sit behind it |
| Your editorial basics: corrections policy, bylines, advertising kept apart | Live on your site; cheap, and the only lever one engine documents outright |
Score before you call: any gate line ends your conversation, and a NEGOTIATE score starts it. Impressions are a result, not a promise. For the agency question see SEO agencies and GEO consultancies; for the method, the GEO playbook. Eligibility is ordinary SEO work, which Leapbuzz runs for brands across Singapore, Malaysia, Australia, the US and Canada.
Do this today: print the mythbusting section of the guide, lay your proposal beside it, and strike every line the page names. Send the vendor what is left and ask for a new price. If the reply argues with Google's page, you have your answer.
