How to Learn AI SEO From Primary Sources

Every AI answer passes through up to four stages, and each one has a different owner. Most bad advice comes from mixing them up.

• Parametric: what the model knows with no web access, frozen at a training cutoff. Nobody changes this quickly.
• Access: whether a crawler may fetch and store your page. You and your CDN control this.
• Retrieval: whether your page enters the model's context for a given query. The engine controls this, mostly through its own search index.
• Generation: which retrieved pages get used and cited. The model controls this.

Retrieval is discretionary. As of October 2026, ChatGPT decides per query whether to search, search in Claude.ai is a user toggle, and Google's AI Overviews often don't trigger at all. The same question can get a grounded answer one day and an answer from memory the next. If you want the vocabulary first, read what an AI citation is and why some ChatGPT answers have no sources.

Next move: label every claim you read with its stage. A tactic tested at the generation stage tells you nothing about whether you get retrieved.

The evidence is strong about access and weak about almost everything else. The levers that hold up are subtractive: they remove blocks rather than add a boost.

In your control

• Rendering. Vercel's crawler study (December 2024, independently re-confirmed in June 2026) found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Content that appears only after JavaScript runs can be invisible to them.
• Per-bot robots.txt rules. OpenAI documents GPTBot (training) and OAI-SearchBot (ChatGPT search) separately and says each setting is independent of the others.
• Network reachability. A firewall or CDN can return a 403 to a crawler that robots.txt allows. OpenAI's help center says eligibility needs both the robots.txt permission and a host or CDN that lets its published searchbot IP addresses through.
• Index and snippet eligibility. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode. noindex, nosnippet, max-snippet and data-nosnippet keep content out.
• Plain, fetchable facts. Stating a price or a one-line definition on a page that crawlers can fetch gives an engine a sentence it can quote.

Not in your control

• Training cutoffs and what the model already believes about you.
• The queries the engine actually runs. ChatGPT rewrites your prompt into its own search queries, and Google may fan out across subtopics.
• Which retrieved pages get cited.
• What third-party sites say about you. Being named on the sites engines already cite is associated with appearing in answers, but nothing promises that effect.

Next move: spend your first learning hours on the controllable list, because you can verify every item on it yourself.

Read Google first. It's the largest channel and its docs are the most explicit about what you don't need. Then read each AI crawler page side by side and write down every bot's job in one line.

Several of these pages carry no date. Record the day you read each one, so you know what to re-check later. For a fuller syllabus built from the same documents, use the free AI search course.

Papers come after the docs because most of them test something narrower than their headlines suggest. Ask one question of every study: was the content already inside the model's context, or did a live engine have to find it?

1. GEO (Aggarwal et al., 2023). The origin of the widely repeated 40% visibility figure. The authors built their own engine and supplied the sources in context. The paper reports a 30 to 40% band across three methods, and none of it touches organic discoverability.
2. C-SEO Bench (NeurIPS 2025). Found most conversational SEO methods "largely ineffective" and often harmful to a document's ranking. Its traditional SEO arm worked by moving the document to position one in the context window, which no publisher can do.
3. What Gets Cited (252,000-trial factorial). Found topical relevance and list position are the biggest drivers of being cited first, and formatting-only edits have little impact. It used two candidate documents per trial and never called a search engine.
4. The 2026 critical survey of 45 GEO studies (arXiv 2607.14035). Concludes that already-retrieved content can change its citation or use, but no technique shows a stable, cross-platform effect on discoverability.

Two industry studies are worth reading with their limits in view. Ahrefs tracked 1,885 pages adding schema and found no major uplift on any platform, with AI Overview citations falling 4.6%. Every page in that study already had 100+ AI Overview citations, so it doesn't test whether schema helps a page get considered at all. Ahrefs also found that 37.9% of URLs cited in AI Overviews appeared in the first 10 results (March 2026).

Next move: for each paper, write one line naming the stage it measured. That line protects you from most of what gets sold as GEO.

Old blog posts and course modules keep repeating rules the primary sources have retired or never said. When a source contradicts a current first-party doc, the source is wrong, not new. Look for these signs:

• It relies on a feature Google removed. Google deprecated FAQ rich results in May 2026 and removed the FAQPage documentation in June 2026. Google also says structured data isn't required for generative AI search.
• It treats a snapshot as a trend. Ahrefs changed its parsing method between its two top-10 studies, so comparing the old and new figures isn't a measured decline. Crawl-to-referral ratios from Cloudflare's July 2025 snapshot have since moved by roughly an order of magnitude.
• It reports a score no engine computes. Google says E-E-A-T "isn't a specific ranking factor." Domain Authority is a Moz metric, and Google publishes no site-level authority score. A Google "AI rank" has nothing to check against, because Search Console's AI report shows impressions only.
• It conflates bots. "Allow GPTBot so ChatGPT can find you" mixes training with search. OAI-SearchBot governs ChatGPT search.
• It recommends a file no engine documents reading. Google says Search ignores llms.txt, and no other engine documents using it.
• It ignores API changes. Perplexity announced the retirement of Sonar Chat Completions for 2026-09-27, and web search is opt-in on its Agent API. Tests built on old code may be measuring answers with no web access.
• It has no date and no primary link. Treat it as a lead to verify, not a fact.

Next move: when you meet a claim, find the first-party page it should rest on, and read the date on that page.

Reading sticks when you check each document against a real site. Work through this in order:

1. Read the four-stage model and the two Google docs. Write down what Google says you don't need.
2. Read the OpenAI, Anthropic, Perplexity, Bing and Meta crawler pages. Build a one-page table: bot name, job, what blocking it does.
3. Open your own robots.txt and resolve the rule for each bot. The robots.txt Checker for AI Crawlers does this per bot.
4. Fetch one key page as raw HTML with JavaScript off, then compare it with the rendered page. List any price, product name or claim that only appears after rendering.
5. Open Google Search Console and Bing Webmaster Tools separately. In Otterly's 2026 experiment, Google deindexed both test sites while Bing kept them and Copilot kept citing them. One healthy index tells you nothing about the other.
6. Ask a few buyer questions on each engine and note whether each answer showed sources. Then learn whether you can see which sources ChatGPT retrieved but didn't cite.
7. Only then read the papers, with your own findings beside them.

• Treating a course as a source. A good course sends you to the docs. Use it as a map, and check its claims (what a good AI SEO course covers). A certificate shows you finished a course, not that its content was right (are AI SEO certifications worth it).
• Treating a citation as proof of use. OpenAI's API returns retrieved sources separately from citations, and the retrieved list is often longer. Citation counts are a noisy proxy for influence.
• Testing only prompts you wrote. Engines rewrite prompts into their own queries. Bing Webmaster Tools' AI Performance report is the one place that shows the grounding queries the AI actually ran.
• Blending engines into one number. Otterly's experiment logged 3,934 citations for one test site, and 2,718 of them came from Microsoft Copilot. A single blended score would have hidden that split.
• Overcorrecting. Ripping out existing schema isn't supported either. Google says existing markup does no harm. It just isn't a lever for AI answers.
• Trying manipulation. Hidden text, prompt injection and cloaking for AI crawlers exist, and research on LLM search agents shows injected content can sway answers once retrieved. It's illegitimate, it breaks spam policies, and it puts your brand at risk. Learn it only to defend against it.