AI SEO Best Practices, Sorted by Evidence

Most AI SEO checklists mix three kinds of advice. Some items have documented mechanisms from the engines themselves. Some are contradicted by primary sources. Some have never been tested in a way that answers the question you care about. Sorting them first saves months of work.

An AI answer passes through four stages. Each one has a different owner, and most advice targets the wrong one.

1. Parametric. What the model knows with no web access, frozen at a training cutoff. Nobody changes this in a quarter.
2. Access. Whether a crawler may fetch and store your page. You and your CDN control this.
3. Retrieval. Whether your page enters the model's context for a given query. The engine controls this, and it often rewrites the user's question into its own searches first.
4. Generation. Which retrieved pages the model uses and cites. The model controls this.

Almost every GEO technique with a published effect size was measured at stage 4, on pages already sitting in the model's context. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) concluded that no reviewed technique shows a stable, longitudinal, cross-platform causal effect on discoverability. C-SEO Bench went further: it found most conversational SEO methods were largely ineffective and frequently hurt ranking.

That is why studies seem to contradict each other. They often measure different stages, as covered in why GEO studies disagree. The practical point: fix stage 2 first, because it is the only stage fully in your hands.

These practices work by removing a block. They don't make you more likely to be chosen. They decide whether you can be chosen at all.

Put load-bearing content in the HTML

Vercel's crawler study found ChatGPT and Claude crawlers download JavaScript but can't read client-side rendered content. If your pricing, product names or key claims appear only after JavaScript runs, those crawlers may see an empty page. Googlebot renders JavaScript, so this gap mostly affects non-Google engines.

Set robots.txt per bot

Rules apply to each crawler separately. OpenAI's bots page says GPTBot governs training use, OAI-SearchBot controls ChatGPT search, and "each setting is independent of the others." A blanket AI block can silently remove you from ChatGPT search. Allowing GPTBot does nothing for search.

Check the firewall, not just robots.txt

OpenAI's help center says eligibility needs two things: allow OAI-SearchBot, and confirm your host or CDN allows its traffic. A firewall can return a 403 to a crawler that robots.txt welcomes. As of October 2026, Cloudflare also blocks Training and Agent crawler categories by default for some accounts since 15 September 2026, while Search stays allowed.

Stay index and snippet eligible

Google's AI features page says a page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode, and that there are no additional technical requirements. A template-level noindex, nosnippet or data-nosnippet removes you without warning.

State specific facts plainly

A page that states a price, a number or a one-line definition gives an engine a sentence it can quote. This is the only on-page content signal with positive experimental support. It is a narrow one, so don't stretch it into a writing formula.

Schema as a citation lever

Ahrefs tracked 1,885 pages that added schema, each compared with three matched controls. It reported no major uplift on any platform. AI Overviews citations fell a statistically significant 4.6%, while AI Mode rose 2.4% and ChatGPT 2.2%. Google says structured data isn't required for generative AI search, and FAQ rich results were deprecated in May 2026.

Don't overcorrect. Google says existing markup does no harm, and schema still serves classic rich results. The accurate line: schema is not a lever for AI answers. Leave working markup in place and stop adding it in the hope of citations.

llms.txt, chunking and AI-friendly writing

Google's AI optimization guide says you don't need new machine-readable files, AI text files, markup or Markdown to appear in its generative AI features, and that Google Search ignores them. The same guide says you don't need to write in a specific way for AI, that there is no ideal page length, and that you needn't chase every long-tail variation. Chunking research measures how the retrieval operator splits documents inside its own index, a variable you don't control. See the document chunking study for the detail.

Scores that don't exist

Google says E-E-A-T "isn't a specific ranking factor." Domain Authority is a Moz metric that no engine computes. Any tool reporting either as an AI visibility input invented the number.

Citation-bait rewrites

Injected statistics, appended quotes and keyword stuffing tested null to negative in C-SEO Bench. Superlatives fare no better, as covered in whether calling your product the best helps. If these tasks sit on your roadmap, retiring unsupported GEO tasks walks through removing them.

Some practices are not refuted. They are unmeasured. Treat them as open questions, not tasks with promised returns.

• Schema for pages not yet cited. Every page in the Ahrefs study already had 100+ AI Overview citations. Whether schema helps a page enter the consideration set was not tested.
• Freshness updates. The main supporting study looked only at cited pages, with no comparison against the general web. Dates are publisher-declared, so changing them proves nothing.
• IndexNow. The protocol speeds discovery for Bing and a few other engines. The spec never mentions AI, Google doesn't consume it, and a 200 response confirms only receipt.
• Bing as a route into ChatGPT. Bing's index backs Copilot. OpenAI says ChatGPT search sometimes partners with other providers. The end-to-end effect has never been measured.
• Third-party mentions. When engines cite other sites instead of yours, being named on those sites is associated with appearing in answers. That is an association, not a guaranteed effect. Managing your brand as an entity covers this route.

Be clear about the boundary before you set targets.

In your control: rendering, robots.txt rules per bot, firewall and CDN settings, noindex and snippet tags, index status in Google and Bing, and whether your own site states facts consistently across pages.

Out of your control: the model's training cutoff, the searches an engine writes from a user's question, the order of pages in its context, and which retrieved page it cites. C-SEO Bench found position in the context list drove results far more than content edits, and no publisher can set that position.

Partly in your control: being mentioned on the third-party sites engines already cite, and organic ranking. Ahrefs found 37.9% of URLs cited in AI Overviews also appeared in the first 10 results (March 2026), so rank helps but doesn't decide.

This is why no honest vendor promises a citation or a score. You can remove what blocks you. You can't force what a model says.

Work from the gates outward. Each step is cheap to check and has a clear fix.

1. Test crawler reachability. Request a key page as OAI-SearchBot, PerplexityBot, Claude-SearchBot and Bingbot, and look for 403s or challenge pages. The AI Crawler Checker shows which ones your firewall turns away.
2. Read robots.txt per bot. Separate training bots from search bots, and make sure search bots are allowed if you want to appear in answers.
3. Diff raw HTML against the rendered page. Anything important that appears only after JavaScript runs needs server rendering.
4. Hunt for accidental suppression. Check templates for noindex, nosnippet, max-snippet:0 and data-nosnippet.
5. Confirm index status engine by engine. Check Google Search Console and Bing Webmaster Tools separately. A healthy one tells you nothing about the other.
6. Make facts consistent and specific. Put the same price, product names and one-line definitions on every page that states them.
7. Measure each engine separately. Then re-check after changes. The free checker on this site gives a snapshot of whether AI engines name you for three buyer questions.

For pacing, the 30/60/90 day AI visibility plan spreads these steps across a quarter.

• Blocking every AI bot at once. A blanket rule meant to stop training often removes ChatGPT search eligibility too.
• Trusting robots.txt alone. A permissive file means nothing if the CDN returns 403.
• Ripping out schema. Removing working markup is as unsupported as adding it for AI. Leave it.
• Rewriting pages in an AI style. Google rejects the idea, and controlled tests show such rewrites can hurt.
• Mass-publishing. In Otterly's 2026 test, Google deindexed two new sites carrying about 1,000 AI-generated posts each. It was one pair of sites, but the removal came with no manual action notice.
• Chasing one blended score. Engines behave differently. Compare the trade-offs in content rewriting vs access fixes before you pick a metric.

Should I remove the schema markup I already have? No. Google says existing structured data does no harm, and it still supports classic rich results in Google Search. The evidence only says schema is not a lever for AI answers: Ahrefs' matched test found no major citation uplift after pages added it. Keep working markup, stop adding new markup in the hope of AI citations, and spend that time on crawler access and rendering instead.

Does allowing GPTBot help my site appear in ChatGPT? No. According to OpenAI's crawler documentation, GPTBot governs whether your content may be used for training. ChatGPT search uses OAI-SearchBot, and OpenAI says each setting is independent of the others. You can block GPTBot and still appear in ChatGPT search, as long as OAI-SearchBot is allowed and your firewall lets its traffic through.

Does ranking in Google still matter for AI answers? It matters, but it doesn't decide. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the first 10 results, using its March 2026 method. The rest came from lower positions. Google says AI Overviews and AI Mode need only indexing and snippet eligibility, with no extra requirements. Ranking well helps, and many cited pages rank lower.

Will publishing lots of AI-generated content get me cited faster? The evidence points the other way for Google. In Otterly's 2026 experiment, Google removed two new sites with about 1,000 AI-generated posts each, with no manual action reported. Copilot kept citing one site because Bing kept it indexed. It was one pair of sites in a low-stakes topic, so treat it as a warning, not proof.

How do I know if a best practice is worth doing? Ask three questions. Is there a documented mechanism from the engine itself? Can you check the result mechanically, like a status code or a robots.txt rule? Was it tested on pages trying to enter answers, not pages already in the model's context? Access fixes pass all three. Most content tactics fail the third.