What Is AI Visibility (LLM Visibility)? How to Check Yours

AI visibility is the share of relevant buyer questions where an AI engine names your brand, plus how it describes you. LLM visibility is the same measure under another name: large language models (LLMs) power ChatGPT, Claude, Gemini and the other engines.

You are visible on a question only when three things hold:

• Known. The model remembers you, or the engine can retrieve a page about you.
• Relevant. It connects you to this question, not just to your name.
• Accurate. What it says about your product, price and audience is correct.

SEO rankings measure your position in a list of links, AI referral traffic counts visits from AI engines, and AI visibility measures whether the answer itself names you. The three move independently: one can be strong while another sits near zero.

Ranking first does not buy you the answer. A results page has ten slots; an answer names however many brands the engine chooses, possibly not you. And as of October 2026, Google's documentation says Google AI Overviews and AI Mode may run several related searches (query fan-out) and show a wider, more diverse set of links than classic search.

Visibility is not traffic either. In Pew Research Center's browsing data for March 2025 (900 US adults), people clicked a result link in 8% of visits that showed an AI summary and 15% of visits that did not, and a link inside the summary in just 1%.

Search Console folds AI-feature traffic into its regular Web totals, so it cannot show how often answers name you. For clicks, see tracking AI referral traffic in GA4.

An engine names you when it can reach pages about you and independent sources back you. It gets there by two routes: what the model absorbed in training, and what it retrieves while it answers (the retrieval-augmented generation pattern). Retrieval is the one you can change this quarter.

As of October 2026, OpenAI runs separate crawlers for each route (GPTBot for training, OAI-SearchBot for ChatGPT search) and says sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers.

For brands, a locked door is rare. In AEOeye's crawler access study (384 well-known sites, fetched once in October 2026; per-site CSV), 1 of 193 SaaS sites blocked any AI retrieval bot (such as OAI-SearchBot, Claude-SearchBot or PerplexityBot), against 27 of 38 news publishers. It is a convenience sample, and robots.txt cannot reveal firewall or bot-protection blocking: 24 of 88 e-commerce and consumer homepages blocked or challenged our crawler.

Corroboration is the stronger half. A September 2025 University of Toronto study found AI search leans on earned media (independent media, review and comparison sites) over brand-owned pages. For software prompts, a web-enabled GPT drew 72.7% of its US citations from earned media and 26.7% from brand sites; Google's results were 45.4% earned and 43.7% brand. Independent pages outvote your own site, so make sure they all name you the same way.

On-page evidence helps: statistics, quotations and cited sources lifted visibility by 15–40% across the two metrics of the KDD 2024 generative engine optimization (GEO) benchmark; keyword stuffing did little.

Ask the engines your buyers' questions, run each one several times, and score every answer on four things: whether you are named, where, how you are described, and who else is named. Aggregate across runs and engines, because one answer is an anecdote.

• Mention rate: answers naming you ÷ answers collected. At zero, nothing else matters. Count citations (links to your domain) separately.
• Position: your average rank among named brands (1 is first), over the answers that include you.
• Accuracy and tone: the share of mentions that get your product, price and audience right. A confident wrong claim costs deals.
• Share of voice: your mentions ÷ all brand mentions, yours plus competitors', on the same questions. Fix the competitor set first.

Repeat because answers drift. In SparkToro's study (run November–December 2025 with AI-tracking startup Gumshoe.ai: 600 volunteers using their usual settings, 2,961 runs on ChatGPT, Claude and Google's AI features), two responses to the same prompt listed the same brands less than 1 time in 100 for ChatGPT and Google. One hospital appeared in 69 of 71 ChatGPT answers but was the top mention in only 25: presence was stable, order was not.

So treat position as an average, never as "we're #2 in ChatGPT." Report each engine separately, since an average hides the engine where you are absent and the Toronto study found engines differ in sensitivity to phrasing. Ask each intent two or three ways.

Good means being named in most answers to the questions closest to a purchase, ideally first and accurately, judged against your competitors rather than a universal number. In our audit records, answers were close to winner-take-most: a named, ranked brand was usually first.

AEOeye's recommendation benchmark covers 127 completed audits of 27 brands, created June 26 to September 7, 2026. Of 892 answers, 157 (17.6%) named the audited brand; of 147 ranked mentions, 71.4% were first, with an average position of 1.54. Answers named 2.9 competitors on average, and 144 of 157 mentions were positive: the common problem is invisibility, not bad press.

Read it with care: the brands are self-selected and include repeat audits, so 17.6% is not a market average, and engine samples are unequal and not comparable (four of the five engines have fewer than 10 mentions). Treat quoted industry averages the same way.

Your benchmark is the competitor the engines name instead of you. If they appear in 8 of 15 answers and you in 1, that gap is the goal.

Run a 15-minute manual test: write five buyer questions, ask each in ChatGPT, Perplexity and Google (AI Mode or AI Overviews), and log who is named. It is a smoke test, not a measurement, so repeat it on another day.

1. Write five unbranded questions a buyer would ask, such as "best [category] for [use case]" or "alternatives to [competitor]". Branded questions only test what a model knows about you; unbranded ones test whether it brings you up.
2. Ask each in a fresh chat on all three engines, with web search on.
3. Log every answer: named or not, position, competitors named, domains cited, anything wrong about you.
4. Compute mention rate (answers naming you ÷ 15), then repeat on another day. The gap between runs is your noise level.

Absent from every answer on both days? Treat yourself as invisible and start with why AI recommends your competitors and how to get recommended by AI. Named in a few? A rate takes many more runs: SparkToro suggests 60 to 100 per prompt, which is where automation beats a spreadsheet.

AEOeye's free audit does a first pass on ChatGPT: three unbranded buyer questions asked live, the full answers, the competitors named instead and starter fixes (run it). It is one engine and one snapshot; the full report ($29 one-time, as of October 2026) adds Claude, Gemini, Google AI and Perplexity. For ongoing monitoring, a dedicated tracker fits better; see the measurement guide and the free checker comparison.