Most “AI SEO” advice is a grab bag: add schema, write FAQs, publish an llms.txt, get on Reddit, fix your headings. Some of it works, some of it is measurably useless, and almost none of it comes with a way to decide what to do first.
The evidence sorts cleanly into three layers. One decides whether AI can read you at all. One decides how much of the answer you win once you're read. And one — the layer most teams work last — decides whether you're in the answer's source material to begin with. Here's each layer, what the data says it's worth, and the order that actually makes sense.
The three layers of AI search visibility
What each layer decides — and how much of the outcome it controls
Work them top to bottom for impact, bottom to top for sequencing: confirm the gate is open (an afternoon), make your pages quotable (a sprint), then put your effort where the ceiling is — off-site.
The strongest single fact in AI visibility research: when AirOps analyzed 21,311 brand mentions in ChatGPT, Claude, and Perplexity answers to commercial queries, 85% came from third-party domains — and roughly 90% of those came from listicles, comparisons, and reviews.1 A brand was 6.5× more likely to be mentioned through someone else's page than its own. Ahrefs' 17-million-citation analysis agrees from the ChatGPT side: for recommendation-style queries, blog listicles alone supply 43.8% of citations.2
It goes deeper than retrieval. Semrush found only 34.5% of ChatGPT queries even run a live web search3— for the other two-thirds, the model answers from memory, and its memory of your category is the sum of everything ever published about it. Off-site coverage isn't just today's source material; it's next year's training data.
Where to earn it varies by engine — Wikipedia and Reddit dominate ChatGPT's citations, Reddit and YouTube lead on Google's AI results and Perplexity4— and by category, with review platforms (G2, Capterra, TripAdvisor, Clutch) owning commercial queries in their verticals. Yext's 17M-citation study found listings and directories make up 54.5% of distinct cited URLs.5 The pattern is consistent: the answer is assembled from pages about your category that you don't control — so your job is to be present in as many of them as honestly possible.
Once a page of yours is retrieved, its construction determines how much of the answer it wins. This is the only layer with peer-reviewed experimental numbers: the Princeton GEO study (KDD 2024) rewrote sources with nine different tactics across 10,000 queries and measured how much visibility each gained inside generative answers.6
Measured visibility lifts from on-page tactics
Change in position-adjusted word count inside generative answers · GEO benchmark, 10,000 queries · Aggarwal et al., KDD 2024
The winning tactics all make content easier to quote and easier to trust. The one classic-SEO tactic tested — keyword stuffing — reduced visibility. Lower-ranked sites gained the most: rank-5 sources jumped up to +115% while rank-1 incumbents lost ~30%.6
Ahrefs' citation analysis adds three structural rules worth building into every page:2
llms.txt: Ahrefs checked 137,000 sites with one — 97% of the files received zero AI-bot requests, and Google's John Mueller says no AI system currently uses it.7 Schema markup: Google states plainly that no special structured data is needed for AI features, and independent tests found citation rates “barely moved” after adding it.8,9 Neither hurts. Neither is the work.
The technical layer matters differently than the other two: it's binary. Vercel analyzed over 500 million GPTBot fetches and found that no major AI crawler executes JavaScript— GPTBot, ClaudeBot, PerplexityBot, and Bytespider all read raw HTML only (Gemini, which rides Google's rendering infrastructure, is the sole exception).10 If your product pages render client-side, they are literally blank to most AI systems. No optimization above this layer matters until that's fixed.
The second gate is bot access, and the details are easy to get wrong. OpenAI runs three separate crawlers: GPTBot (training), OAI-SearchBot (ChatGPT search indexing), and ChatGPT-User (live fetches when a user asks). Blocking GPTBot in robots.txt does not remove you from ChatGPT search — blocking OAI-SearchBot does. Google-Extended controls Gemini training but not AI Overviews, which use standard Googlebot.11 Plenty of sites are invisible to AI search because of a robots.txt line someone added in 2023 to make a statement about training data.
Once the gate is open, it's open. There is no “more crawlable.” Spend an afternoon verifying it, then move up the stack.
The layers also explain why AI visibility feels unfair to good products with quiet marketing: the model can only recommend what the record supports. Build the record, structure the proof, keep the gate open — in that order of effort, and the reverse order of urgency.
Beacon shows where you stand in AI answers today, which sources decide your category, and the exact actions — layer by layer — to climb.