How Does AI SEO Work? The Four Stages of AI Search

Classic SEO has one big question: where does your page rank? AI search splits that into four questions, asked in order. A failure at an early stage makes every later stage irrelevant.

The four stages come from how an answer engine builds a reply:

1. Parametric. What the model already knows from training, with no web access.
2. Access. Whether a crawler may fetch and store your page at all.
3. Retrieval. Whether your page enters the model's context for a given query.
4. Generation. Which of the retrieved pages the model uses and cites.

Most confusion about AI SEO comes from mixing these up. A tactic that changes stage 4 gets sold as if it fixes stage 2. Keep the stages apart and the advice sorts itself into what works, what is unproven, and what nobody controls.

Each stage has a different owner. This table is the map for the rest of the page.

Parametric

A model's training data stops at a cutoff. OpenAI's model docs list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Content you publish after a cutoff cannot appear in an answer given without search. Changing that takes a new model generation.

Access

This is the publisher's stage. Crawlers are separate by purpose. OpenAI's crawler page says GPTBot governs training, OAI-SearchBot governs ChatGPT search, and "each setting is independent of the others." Access also depends on rendering: Vercel found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so content that only appears after JavaScript runs may be invisible to them. Googlebot does render.

Retrieval

At answer time, the engine decides whether to search at all, rewrites the user's prompt into its own queries, and pulls candidates. OpenAI says ChatGPT search rewrites queries into one or more targeted queries. Google may fan out across subtopics. You can't see those queries, and you can't set your position in the candidate list.

Generation

The model reads the retrieved pages and writes. It decides what to quote and which sources to cite. Citation is not the same as use: OpenAI's API returns a retrieved list that is often larger than the cited list.

Studies are easiest to run at the last stage. A researcher puts a few documents in front of a model, edits one, and counts how often it gets cited. No live index is queried. That makes the experiment clean, and it also means the results say nothing about whether a page gets retrieved in the first place.

The critical survey of 45 GEO studies (arXiv, July 2026) puts it plainly: already-retrieved content can causally alter its citation or use, "but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior."

The widely quoted "up to 40% more visibility" figure comes from the original GEO paper, which used an engine the authors built with sources already loaded into context. C-SEO Bench (NeurIPS 2025) tested conversational SEO methods and found most "not only largely ineffective but also frequently have a negative impact." Ahrefs' matched-control study found AI Overview citations fell 4.6% after pages added schema.

The gap between these setups and real engines is the subject of our comparison of lab GEO studies and live engines. The short version: a generation-stage result is conditional on retrieval having already succeeded. If your page never reaches the context, no rewrite matters.

Be honest about the line. It decides where your hours go.

In your control

• Rendering. Put the facts that matter (prices, product names, definitions) in the HTML before JavaScript runs.
• robots.txt, per bot. Allow the search crawlers you want, such as OAI-SearchBot, PerplexityBot and Claude-SearchBot. Blocking GPTBot does not remove you from ChatGPT search, and allowing it does not get you cited.
• Network reachability. A firewall or CDN can block a crawler that robots.txt allows. From September 15, 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and existing free-tier customers, while Search stays allowed. A site can lose access without anyone editing it.
• Index and snippet eligibility. Google's AI features guide 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, with "no additional technical requirements." Check for stray noindex, nosnippet, max-snippet:0 and data-nosnippet.
• Stated facts. A specific fact on a fetchable page gives an engine a sentence it can quote.

Partly in your control

• Topical relevance. The survey and the 252,000-trial factorial both name topical relevance and list position as the most reproducible drivers. You shape relevance by covering the subject well, not by stuffing keywords.
• Being named elsewhere. When engines cite other sites, appearing on those sites is associated with being in the answer. It is an association, not a promised effect.

Not in your control

• The model's cutoff and training mix.
• Whether the engine searches for a given prompt.
• The queries it actually runs and your position in its candidate list.
• Which retrieved source it decides to cite.

Work from the first gate forward. Each step is cheap compared with rewriting content, and each one removes a block that would cancel everything after it.

1. Check crawler reachability. Request a key page as each AI crawler and confirm you get a 200, not a 403 or a challenge page. The AI Crawler Checker does this per bot.
2. Read your robots.txt per agent. Separate training bots from search bots. A blanket AI block often removes ChatGPT search eligibility by accident.
3. Diff raw HTML against the rendered page. List every price, plan name or claim that exists only after JavaScript runs. Server-render those.
4. Check index status engine by engine. Look at Google Search Console and Bing Webmaster Tools separately. One being healthy says nothing about the other.
5. Remove accidental suppression. Search templates for snippet directives someone added to hide a cookie banner.
6. Fix contradictions on your own site. If the pricing page, footer and about page disagree, an engine can repeat the wrong one.
7. Then work on relevance and third-party mentions. Cover the topics your buyers ask about, and get named accurately on the sites engines already cite.

For how this overlaps with classic search work, see SEO for AI search. Ranking still correlates with AI Overview citations: Ahrefs found 37.9% of URLs cited in AI Overviews also appeared in the top 10 results (March 2026). Whether you need a Google ranking for ChatGPT is a separate question, because ChatGPT runs its own retrieval.

Most wasted AI SEO effort comes from treating a later stage as if it were an earlier one.

• Adding schema or FAQ schema to get cited. Google says structured data isn't required for generative AI search, and FAQ rich results were deprecated in May 2026. Existing markup does no harm. It just isn't a lever for AI answers.
• Publishing an llms.txt. Google says Search ignores it, and no engine documents reading it.
• Rewriting in an "AI-friendly" style or chunking pages. Google's optimization guide rejects both and says there is no ideal page length.
• Chasing an E-E-A-T score or domain authority. Google says E-E-A-T isn't a specific ranking factor. Domain Authority is a Moz metric that no engine computes.
• Tracking an AI rank. Google's AI report exposes impressions only, with no position data to check a rank against.
• Assuming one engine speaks for all. Retrieval differs per engine, so what works on one AI engine may not carry to the others.
• Using hidden text or prompt injection. Injected instructions can sway an agent once the page is retrieved, which is why it is a real risk when third parties do it to your brand. It is manipulation, and engines and spam policies treat it as such. Never use it.

A single blended visibility number averages answers given with and without live search, so it can't tell you which stage broke. You need two views: what the engines say, and whether their crawlers can reach you.

Start with answers. The free checker on this site runs up to three buyer questions and shows whether engines name you, which tells you if there is a gap worth chasing. Then trace the gap backward through the stages. If an engine describes you with old facts and no citations, the answer may be parametric. If it cites competitors for a topic you cover, check access and retrieval before touching the copy.

If you want a broader view on whether the field holds up at all, read whether GEO is legit or hype. The honest position: access is checkable and fixable, relevance is real but slow, and nobody controls what the model finally says.

What is the difference between AI SEO and traditional SEO? Traditional SEO aims for a ranked position in a results list. AI SEO aims to have your content fetched, retrieved and used in a generated answer. The access work overlaps heavily: indexing, crawlability and snippet eligibility matter for both. The difference is that AI engines add retrieval and generation stages you can't see, rewrite the user's query, and expose no position data, so there is no rank to track.

Does allowing GPTBot help my brand appear in ChatGPT? Not for search. OpenAI's crawler documentation says GPTBot governs training use, while OAI-SearchBot controls whether your site can appear in ChatGPT search answers. The settings are independent. Allowing GPTBot might put content into a future model's training data, but nothing published measures that effect. If you want ChatGPT search eligibility, allow OAI-SearchBot and make sure your CDN lets it through.

Why does ChatGPT describe my company with outdated facts? The answer may be coming from parametric knowledge, with no web search. As of October 2026, OpenAI lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Anything newer can only reach the answer through retrieval. Make sure current facts sit in server-rendered HTML on a page OAI-SearchBot can reach.

Can I control which sources an AI engine cites? No. Citation happens at the generation stage, inside the model. You can make sure your page is reachable, indexed, rendered and topically relevant, and you can state specific facts clearly so there is a sentence to quote. Lab studies show rewriting already-retrieved pages can shift citations, but the best-designed ones found most such tactics ineffective or harmful.

How long before an access fix shows up in AI answers? Nobody publishes a timeline. A fix removes a block, which makes your page eligible again, but the engine still has to recrawl, re-index and decide to retrieve it for a relevant query. Retrieval is discretionary per query on every platform. Recheck crawler responses right after the fix, then watch answers per engine over the following weeks rather than expecting a set date.