69.1% of 913 US marketing agencies were named by no AI assistant at all.

Three findings

• 631 of 913 (69.1%) were never named once — not by one assistant in one answer, across every question asked about their category. These are working agencies with clients and websites. To an assistant answering a buyer, they do not exist.
• Where an agency was named, 84.6% of the time the assistants disagreed — of the 280 agencies that any assistant named, only 43 were named by all three. A single “AI visibility score” averages that disagreement away, and the average is true of no assistant your buyer actually opened.
• The answers are built out of a handful of pages — clutch.co was cited in the answers for 912 of the 913. Getting listed and reviewed on those pages is the lever, and it is not on the agency’s own website.

The assistants do not behave the same way

ChatGPT named an agency in 5.6% of its answers and Perplexity in 3% of its own — the same agencies, the same week, the same questions. Perplexity was asked the most questions by a wide margin and still named the fewest agencies per answer, which is the opposite of what more chances would predict.

What the assistants read

The sources cited most often across all 26 categories. “Agencies” is how many of the 913 had this page cited in the answers about them.

Worth noticing: thriveagency.com is an agency, not a directory, and it was cited in the answers about 835 other agencies. Whatever it publishes has become a source the assistants read when they answer about its own competitors.

Method

• The questions — buying questions for each category, of the form a person types before choosing an agency, generated per category and reused unchanged for every agency in it.
• The assistants — Claude, ChatGPT and Perplexity through their official APIs with web search on. An answer produced without a search is discarded rather than scored: it reflects training data, not what a buyer would see today.
• Counting — an agency is “named” when the assistant writes its brand name in the answer. Being cited as a source is counted separately.
• The window — 2 August 2026 to 24 August 2026. Assistants change; this is a photograph, not a law.

What this does not show

• The assistants were not asked equally. Perplexity was asked about five times as many questions per agency as the other two, because it is the cheapest to run. More questions means more chances to be named, so the disagreement figure above is inflated by our own sampling, not only by the assistants.
• One attempt per question. Assistants are not deterministic. A single ask can miss a name that a second ask would find, which pushes the “never named” number up.
• These agencies did not volunteer. They were surfaced by asking assistants and reading public directories, so the sample is of agencies findable that way — not a register of every US agency.
• No causal claim. Nothing here says that being listed somewhere makes an assistant name you. It says the assistants cite those pages when they answer.

Your own number

If your agency is one of the 913, its measurement already exists and the check is free — no signup and no card. If it is not, the same page says so plainly rather than pretending.

Computed 2026-09-10 from AEOScope’s own scans. Quote it, link it, or argue with it — if you want the underlying counts for a category, ask and I will send them.