The AI Visibility Ladder: Retrieved, Cited, Mentioned, Recommended

"Are we visible in AI?" sounds like a yes-or-no question. It isn't. When a buyer asks an AI agent for help, your brand can show up in four different ways, each earned differently. Collapse them into one number and you end up fixing the wrong thing.

The AI Visibility Ladder splits it into four rungs: retrieved, cited, mentioned, and recommended. The framing comes from the marketingskills AI SEO reference on citations vs recommendations (MIT licensed), which credits work by Lily Ray at Amsive. This guide adds evidence labels to each rung so you can see which moves are backed by research and which are just popular.

What each rung means

1. Retrieved

An AI agent's crawler can reach your page, and your content is in the HTML it receives. This is the gate: if an engine can't fetch the page, or gets an empty JavaScript shell, nothing above this rung can happen from that page. Retrieval is mostly invisible; the answer may not show that the agent read you.

2. Cited

Your page appears as a source in the answer: a link, footnote, or numbered reference. The engine found it useful to consult. That doesn't mean it named or recommended you.

3. Mentioned

Your brand is named in the answer text. That depends on whether the model recognises your brand and links it to the category, which depends heavily on how the rest of the web talks about you.

4. Recommended

You're suggested as the pick, or put on the shortlist the buyer actually considers. This is the rung that changes buying behaviour, and it mostly reflects consensus across the web: reviews, comparisons, forums, press, and analysts. There's also a shadow rung, being recommended against: on requirements-heavy questions, AI agents sometimes name products to avoid.

What moves each rung

The labels below follow the same rules as the rest of our evidence library (see the method):

• Supported: consistent evidence that it matters, usually large observational studies or controlled tests, with some replication. Outcome data is still mostly correlational.
• Mixed: real evidence that depends on the engine or context, or sources conflict.
• Unproven: widely recommended, without credible evidence of citation lift. Fine as hygiene.

Retrieved: access and readable HTML

• Allow AI search and agent bots in robots.txt, your CDN, and your firewall. Evidence: Supported. OAI-SearchBot, PerplexityBot, Claude-SearchBot, and user-triggered fetchers like ChatGPT-User matter for live answers. Training crawlers such as GPTBot are a separate policy choice. See AI bot access.
• Put the main content in the initial HTML. Evidence: Supported. Most AI crawlers read only the raw HTML. Since late September 2026, two site owners have logged OpenAI's OAI-SearchBot rendering JavaScript on some pages (Hisashi Space, Create With Logic); OpenAI hasn't confirmed it. Don't rely on it: put the main text in the first HTML response. See server-rendered content.
• Match your title and paragraphs to the buyer's question. Evidence: Supported. The University of Toronto study What Drives AI Citations? (arXiv 2610.15033, a preprint) held websites fixed and replayed ChatGPT's own searches. Whether a page was retrieved explained about 32% of within-site citation variance. All the page features they measured together explained about 2.8%. Of those features, only a query-matching title and paragraphs held up across every robustness test. See query-match.
• Publish /llms.txt. Evidence: Unproven. Google's John Mueller has said no AI system currently uses it, and a large SE Ranking analysis found no link to citations. Our free check gives it zero weight. See llms.txt.

Cited: pages that are easy to quote

• Answer first, under question-style headings. Evidence: Mixed. A short, direct answer at the top of each section can be quoted on its own. Across sites, cited text skews to the top of the page, but that's correlational, and within a site Toronto found headings made no difference. See answer-first headings.
• Cite sources, real statistics, and named quotes. Evidence: Mixed. The Princeton GEO paper (Aggarwal et al., KDD 2024) found these rewrites raised visibility by up to about 40%. Read that number carefully: it is position-adjusted word count in a lab engine reading a fixed set of Google results, not a ChatGPT citation rate, and top-ranked pages lost visibility with one tactic. Later benchmarks found rewrites weak or inconsistent. Source claims because it improves the page, and never invent numbers.
• Show real publish and updated dates. Evidence: Mixed. Ahrefs found AI-cited URLs skew fresher than organic results. That finding is correlational, and older pages still win when they are the best match.
• FAQ, HowTo, and other schema. Evidence: Unproven. Schema is hygiene, not a citation boost. Ahrefs' before-and-after study of 1,885 pages found no meaningful citation lift. When an engine says schema helps it understand a page, that is not the same as citing it.

Mentioned: how the web talks about you

• Branded mentions across the web, including video. Evidence: Mixed. Ahrefs' study of 75,000 brands found branded web mentions and YouTube mentions correlated more strongly with AI visibility than site size did. That is a correlation in a vendor dataset, not a recipe, and among top domains Ahrefs found the link strong for AI Overviews but weak for ChatGPT.
• Earned media and PR. Evidence: Supported. Toronto's earlier GEO study found AI search leans toward earned, third-party sources over brand-owned content.
• Community and forum presence. Evidence: Mixed. Heavily cited on some engines, much less on others, and the mix shifts.
• Organization and sameAs entity markup. Evidence: Unproven. It helps engines identify you; there's no controlled evidence it lifts mentions.

Recommended: third-party consensus

• Inclusion in independent comparisons and "best of" lists. Evidence: Supported. Comparison and list pages are a large share of what AI agents cite for recommendation questions. The lists that count are ones you don't write.
• Reviews, analysts, and the wider record. These are the inputs a model weighs when it builds a shortlist. Our library doesn't grade them yet, so they carry no label here. The practical test from the marketingskills reference is a good one: if a model ignored everything on your domain, would the rest of the web still put you on the shortlist?
• Self-ranked "best X" guides. Evidence: Mixed. Your guide can earn the citation, but if you aren't yet the consensus pick, the engine may use your research to recommend competitors.
• Organic rankings and backlinks. Evidence: Mixed. Ahrefs updated its own finding that 76% of AI Overview citations come from Google's top 10: the newer figure, from March 2026 across about 4 million URLs, is about 38%. Engines fan out into many sub-queries.

Why measurement is shaky

Every rung above the first is hard to measure. As SEO consultant Daniel Foley Carter puts it: visibility in which AI?

• Answers vary by phrasing. Buyer questions are long and rarely repeat word for word. Toronto found that ChatGPT's own background searches overlapped only about 2.7% between runs of the same question on the same day.
• Answers vary by engine. ChatGPT, Perplexity, Claude, Gemini, and Google's AI features retrieve differently and favour different source types. A blended average can hide opposite behaviour.
• Answers vary by day, not just by week. A study of repeated, same-prompt answers (arXiv 2609.22655, a preprint) saw each engine's cited URL set change by about 67% on average from one day to the next, and engines shared under 1% of cited URLs with each other. It covered only 15 prompts, and its author sells AI-visibility tooling. Profound, a company that sells AI visibility tracking, reports 40–60% of cited domains changing month to month (vendor research).
• Analytics are thin. There's no Search Console for ChatGPT, and buyers influenced by an AI answer often arrive via branded search, so referral reports undercount.

So one answer is an anecdote, not a measurement. Track clusters of real buyer questions, run each several times per engine, and record the rate ("mentioned 3 of 5 runs") rather than a single yes or no. Be wary of any single "AI visibility score" that doesn't tell you which engines, prompts, and dates it covers.

What you can check yourself for free

The bottom two rungs live on your own pages, so you can check them directly. Our free AI visibility check covers retrieved and cited:

• whether robots.txt allows AI search and agent bots, and whether a firewall blocks them when we fetch as OAI-SearchBot, PerplexityBot, and Claude-User
• whether the main content is in the raw HTML
• how well your title, H1, and paragraphs match buyer questions, including your own if you add them
• lower-weight signals like dates, sources, and answer-first structure

It gives no overall score and no citation forecast. Every check links to the evidence behind it.

For the top two rungs, ask a few real buyer questions in two or three AI agents, several times each, and note who gets named. Our guide to ChatGPT brand mentions walks through how.

Where outside help matters

Mentioned and recommended depend on the wider web: third-party sources, reviews, independent comparisons, community discussion, and a clear, consistent entity record of the kind Wikipedia-style sources hold. No on-page tweak fixes those. It takes knowing which sources the engines cite in your category, earning a place in them, and measuring honestly over months.

That's the work Uncanny Labs does with clients. If your pages clear the first two rungs and you still aren't being named, book a call. It starts with a 20-minute call about your category, and no one, us included, can promise you citations.

A short checklist

• AI search and agent bots are allowed in robots.txt, the CDN, and the firewall
• Main content is in the raw HTML response
• Priority pages answer a real buyer question in the title and opening paragraphs
• Claims link to real sources; no invented numbers (Mixed)
• Dates are visible and reflect real updates (Mixed)
• Schema and llms.txt are treated as hygiene, not levers (Unproven)
• You know which third-party sources AI agents cite in your category
• Buyer questions are re-run across engines on a schedule, and you record rates rather than single answers

Check your site

See whether AI agents can retrieve your page and match it to buyer questions, with the evidence behind each fix: run the free check.