What an AI visibility audit should check, and what it cannot see

Search “ai visibility audit” today and the top results agree on the outline. Yotpo’s guide, last updated 20 October 2026, runs it in four stages: pick queries, prompt the engines, map the citations, act. OutreachBloom’s checklist suggests 20 to 50 prompts, each run three times. Both are sound. Neither shows a finished audit, so you cannot see what the checks catch.

This page does that. On 25 September 2026 we sent twenty beginner crypto wallet questions to ChatGPT, Perplexity, Gemini, Claude and Google AI Mode and logged all 100 answers. Each section below takes one check an audit should run and shows what it found in that log. If you want the twenty-minute version you can run by hand, that is the AI visibility checker. This is the full audit.

What is an AI visibility audit?

An AI visibility audit measures how AI answer engines treat your brand when a buyer asks a category question: whether you are named, where you sit in the answer, and which pages the engine relied on. It is a snapshot taken under fixed conditions, so it can be repeated and compared.

An AEO audit and a GEO audit are the same exercise under different labels, one from answer engine optimization and one from generative engine optimization. The difference that matters is with an SEO audit. An SEO audit inspects your own site. An AI visibility audit inspects the answers, and most of what shapes them sits elsewhere: in our April to August 2026 citation study, 92.22% of citations landed off the audited brand’s own domain (39,948 citation events across 72 audited crypto brands).

So the useful output of an audit is rarely a list of fixes to your website. It is a list of other people’s pages, ranked by how often the engines read them before answering your buyer.

Which prompts should an AI visibility audit use?

About twenty questions written the way a buyer types them, none of which names a brand, locked before the first run. A prompt that names a brand seeds the answer, and a seeded answer cannot be counted toward a market figure.

We learned that on our own first draft. Version one of the wallet prompt set had 11 brand-neutral questions, 6 comparisons and 3 brand checks. The nine that named a brand pulled that brand into the answer, so only 55 of the 100 observation points could be used for market figures. Version two made all twenty brand-neutral, and all 100 points counted.

Which engines should an audit cover, and why all five?

ChatGPT, Perplexity, Gemini, Claude and Google AI Mode, because they disagree with each other often enough that one engine’s answer cannot stand in for the market. An audit on ChatGPT alone reports ChatGPT’s opinion.

In the wallet audit, 23 different wallets took the first slot across the 100 answers. Three or more engines agreed on the first pick for 7 of the 20 prompts. Only one prompt was unanimous. The leaders also differed by engine: Gemini led with Zengo on 6 prompts, ChatGPT and Google AI Mode most often led with MetaMask, and Claude most often led with Trust Wallet. ChatGPT never named Zengo or Exodus on any of the twenty prompts.

For a wallet brand, that means “how visible are we in AI” has five answers. A brand that tested only ChatGPT would not see Zengo as a competitor at all. A brand that tested only Gemini would think Zengo owned the category.

Should an audit count mentions or first picks?

Both, reported separately. A mention rate on its own flatters brands that appear in long lists. The first pick is the name a buyer is most likely to act on, and it moves very differently.

Ledger was named in 56 of the 100 wallet answers, more than any other wallet, and was the first pick in 7. MetaMask was named in 48 and picked first in 18. Zengo was named in 35 and picked first in 13. Ranked by mentions, Ledger leads the category. Ranked by first picks, it is fourth equal.

An audit should also log what each answer says about you. Fees, supported chains, custody model and licences are facts an engine can get wrong, and both of the top-ranking guides include accuracy and sentiment as a check. A wrong fact in an answer usually traces back to a cited page that states it, which makes it one of the cheaper findings to fix.

What should an audit check about sources?

Every URL each engine cited, rolled up to the domain and kept per engine, because the sources are the pages the engine read before it chose a name. They are the closest thing an audit produces to a to-do list.

The engines source very differently, which is why a pooled source list misleads. Across the same twenty wallet prompts, Perplexity returned 578 citation slots and ChatGPT returned 127, spread over just 22 domains. ChatGPT leaned on product documentation: trezor.io (19 slots), support.metamask.io (18) and ledger.com (14) made up 51 of its 127. ledger.com was cited 49 times across four of the five engines, which is part of why Ledger is named so often.

The wider data says the same thing at scale. In our April to August 2026 citation study (39,948 citation events, 72 audited crypto brands), a brand that was the subject of a page averaged 6.22 citations per brand, against 3.78 for a brand ranked inside someone else’s list and 3.39 for pages where it was absent. An audit that finds the pages engines read, then sorts them by how they treat you, has found where the work is.

What can an AI visibility audit not see?

Answers given without sources, and answers whose sources the engine does not expose. A good audit reports those cells as gaps. Counting them as zero citations makes an engine look like it ignores the web when it simply did not show its reading.

Two gaps showed up in the wallet log. Claude came back with web search off on 6 of the 20 prompts, despite search being requested, and answered from training with no sources at all. Google AI Mode exposed a list of source links on 1 of the 20 prompts, six links in total, and named sites inside the prose or not at all on the other nineteen. Its first picks and mentions are as usable as any other engine’s. Its sourcing column is not, and an audit should say so in the report.

The third limit is sampling, and it applies to our own audit. AI answers are non-deterministic: the same prompt can return a different list minutes later. OutreachBloom recommends running each prompt three times across at least two days. Our audit sends each prompt once per engine, so any single cell can flip on a re-run. We read patterns across the 100 cells and do not act on one cell. If you need a single prompt to be reliable, run it several times.

How is an AI visibility audit different from a checker or tracking?

A checker is a quick look, an audit is a measured baseline, and tracking is the same audit repeated on a schedule. Start with a checker if you have never looked. Run an audit before you spend money on AI visibility work. Track once the work starts, so you can tell movement from noise.

Tools that automate some or all of this are compared in the best AI visibility tools, tested and scored. citeOS is one of them and is our product, so read that page with the disclosure in mind.

What should you do with the audit results?

Rank the cited domains by how many engines and prompts they appear in, then work down that list. The top of it is the small set of pages that shape answers across the board, and those are worth more than any change to your own homepage.

Per-engine playbooks for the second and third steps are in how to rank on ChatGPT and how to rank on Perplexity.

How was this audit run?

What else do people ask about AI visibility audits?

Disclosure. Published by Emergence Media, which operates citeOS. Contact: Sagar Saxena on X. citeOS is our product and is mentioned on this page. The wallet figures describe what five AI engines said on one day and say nothing about which wallet is safest. AI answers are non-deterministic, so any single answer you see may differ. Nothing here is investment advice.