Building an AI SEO Strategy on Evidence

Every AI answer passes through up to four stages. Your strategy should treat them separately, because your influence differs sharply at each one.

1. Parametric. What the model knows with no web access, frozen at its training cutoff. Nothing you publish this quarter reaches it.
2. Access. Whether a crawler is allowed to fetch your page, can reach it, and can read it.
3. Retrieval. Whether your page enters the model's context for a given question. The engine rewrites the user's question into its own search queries, so you never see the real input.
4. Generation. Which retrieved pages the model uses and cites.

Retrieval is also discretionary. As of October 2026, ChatGPT decides per question whether to search, Claude's web search is a user toggle, and Google AI Overviews often don't trigger at all. So any single answer you look at may come from live search or from frozen training data.

This matters for strategy because almost all published GEO experiments test stage 4 only. A July 2026 critical survey of 45 GEO studies concluded that already-retrieved content can change its own citation, but no reviewed technique shows a stable, cross-platform causal effect on being discovered in the first place. If you want the background on why studies conflict, see why GEO studies disagree.

Sort every proposed tactic by the stage it acts on. Most of the market's advice sits in the bottom row.

Two things cut across the table. Brand prominence is strongly associated with being cited, but you can't change it quickly. Third-party mentions are the route into answers when engines cite other sites instead of yours, and that is partly in your hands.

The practical rule: spend first where control is high and results are checkable. Treat everything else as something to observe, not promise.

Access levers work by removing blocks, not by adding signals. They're also the only ones where the engine operators document the cause and effect.

Set robots.txt per bot

Rules are per crawler and independent. OpenAI's crawler documentation says GPTBot governs training use, while OAI-SearchBot controls whether you appear in ChatGPT search answers. Blocking one does not block the other:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

The common error is a blanket AI block that silently removes search eligibility. Check PerplexityBot, Claude-SearchBot and Meta-WebIndexer too.

Check the firewall, not just the file

A firewall or CDN can return a 403 to a crawler that robots.txt allows. OpenAI's help documentation says eligibility needs both the robots.txt permission and a host or CDN that allows its published crawler IPs. Cloudflare changed its defaults on 2026-09-15 to block training and agent crawlers on ad-displaying pages for new domains and free-tier accounts, so some sites lost access without editing anything. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

Put the important content in the HTML

Vercel found that ChatGPT and Claude crawlers fetch JavaScript files (11.50% and 23.84% of their requests) but don't execute them. Prices, product names and key claims that appear only after JavaScript runs may be invisible to them. Googlebot renders JavaScript, so this gap mostly affects non-Google engines.

Audit suppression controls and index status

noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers. Templates often apply them by accident. Google states that a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode, with no extra requirements. Check Google Search Console and Bing Webmaster Tools separately, since Bing's index backs Microsoft Copilot.

AI platforms get specific facts wrong about specific brands. Wrong prices, outdated product names and confused ownership are the complaints you'll hear from sales first.

The one positive experimental signal for on-page content is simple: stating a specific fact (a price, a number, a one-line definition) on a fetchable page gives the engine a sentence it can quote. Do this:

• Put pricing, plan limits, founding year, location and core product names in plain text in the HTML.
• Make every page agree. If the pricing page, about page and footer state a fact three different ways, you've supplied the contradiction yourself.
• Retire old pages that state outdated facts, or update them.
• Record each wrong answer you find, with the engine and date, then track whether it changes. An AI error log and correction tracker keeps this from becoming anecdote.

Be plain about the limit. A fix on your site reaches answers that use live retrieval. Answers drawn from training data won't change until the model is retrained.

There is no shared AI index. Google states that AI Overviews and AI Mode may use different models and techniques, and ChatGPT, Perplexity, Claude and Copilot each retrieve differently. A single blended score averages things that don't belong together.

Otterly's 2026 experiment shows why. Google deindexed both of its test sites, and Google's AI surfaces almost stopped citing them. Bing kept them indexed, and Microsoft Copilot produced 2,718 of the 3,934 citations Otterly tracked for one site. ChatGPT cited them almost never. One blended number would have hidden all of that.

Good measurement practice:

• Score each engine separately and note whether the answer showed sources (a sign that live search ran).
• Test logged out with memory off, so results are repeatable.
• Use buyer questions, but remember engines rewrite them. Bing Webmaster Tools' AI Performance report is the only place that shows the queries the AI actually ran.
• Treat Google Search Console's AI report as impressions only. As of October 2026 it has no clicks or position.

For a starting snapshot, the free checker on this site shows whether AI engines name you for up to three buyer questions. It's a point-in-time read, not continuous tracking.

When an engine answers your buyer's question by citing review sites, directories, comparison articles or publications, your own page isn't the route into that answer. Being named on those sites is.

The evidence here is an association, not a promised effect. Citations concentrate heavily on a small set of well-known sources, and brand size drives much of it. Still, it's the practical route when your site isn't the one being cited.

How to work it:

1. For each buyer question, list the domains each engine cites.
2. Check whether you're mentioned on those pages, and whether the facts about you are correct.
3. Correct wrong listings first. A wrong price on a cited directory is an accuracy problem you can fix.
4. Pursue genuine inclusion on the pages that matter, through outreach, partnerships and being worth mentioning.

This is not the same as building a link profile for its own sake. For how links fit in, see whether backlinks still matter for AI search.

Work in this order, because each step depends on the one before it.

1. Access audit. Per-bot robots.txt, firewall responses, rendering, suppression tags, index status in Google and Bing.
2. Fact inventory. List the facts buyers ask about and make each one consistent and present in HTML.
3. Baseline per engine. Record how each engine answers your buyer questions, with dates and whether sources appeared.
4. Third-party map. Find the cited domains and fix or earn mentions there.
5. Re-check on a schedule. Crawler access can break without you touching anything, and engines change models often.

Goals you can actually verify

Set goals on inputs, because outputs belong to the engines. Good goals look like this:

• Every retrieval crawler you want returns a 200 on key pages, with no firewall challenges.
• Pricing and product facts appear in raw HTML on every relevant page.
• Known wrong facts are logged, the source is fixed, and each one is re-checked monthly per engine.
• Accurate mentions exist on the top cited third-party pages for your core questions.

Report per-engine presence as an observation, not a target with a number attached. Nobody can honestly promise a citation count or an AI ranking, and Google exposes no position data to check one against. If you need to justify the budget, a business case for AI visibility work built on these inputs holds up better than one built on promised outputs.

Most failed AI SEO strategies spend effort on claims the primary sources contradict.

• Adding schema to get cited. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm. It just isn't a lever for AI answers.
• Publishing llms.txt. Google says Search ignores it, and no other engine documents reading it.
• Chasing an E-E-A-T score or domain authority. Google says E-E-A-T isn't a specific ranking factor, and domain authority is a third-party metric no engine computes.
• Rewriting or chunking for AI. Google says you don't need to write in a special way for AI, and chunking happens inside the engine. See who controls document chunks.
• Allowing GPTBot to appear in ChatGPT search. GPTBot is for training. OAI-SearchBot governs search.
• Mass-publishing AI content. Otterly's thousand-post test sites were removed from Google with no manual action notice.
• Hidden text or prompt injection aimed at AI crawlers. It exists, research shows some agents are vulnerable to it, and it's manipulation that puts your reputation and index status at risk. Don't do it.
• Buying outcomes. Be wary of packages that guarantee citations or share of voice. Read whether share of voice packages are worth buying and the evidence on any proven way to increase AI citations first.