A Claude Code Skill for AI Visibility: Measure Whether Engines Name You at All

Someone who wants what you sell no longer types four words into Google and scans ten blue links. They ask an assistant a whole question — "best project management app for a five-person team" — and read one synthesized paragraph naming three or four brands. Either you are in that paragraph or you do not exist for that buyer, and your rank tracker cannot see the answer at all. An AI-visibility skill for Claude Code is a fixed, repeatable measurement of that surface. This page is about the one AgentSource ships for the job — the AI Visibility Audit, $9 one-time — and the captured run you can read before you buy it.

Being named and being cited are two different scores

The single most common mistake here is treating a mention as the outcome. The skill scores them apart, and the field report contains the case that shows why.

On the "Tasklet vs Asana" prompt, ChatGPT named the brand first and described it favourably: a lighter, cheaper tool aimed at small teams. It linked it zero times. A mention counter records a win. A citation counter records nothing. Both are correct, and only the second one tells you the engine is describing you from memory rather than reading your page — which is a completely different problem with a completely different fix.

The same run shows the reverse split across engines on identical prompts:

Engine	Captured	Mention rate	Citation rate
ChatGPT	3/3	1/3 (33%)	0/3 (0%)
Perplexity	2/3	2/2 (100%)	2/2 (100%)
Google AI Overview	2/3	0/2 (0%)	0/2 (0%)
Gemini / Claude / Copilot	—	NOT CAPTURED	—

Three engines, three unrelated verdicts on the same brand. Checking one of them and generalizing is how a brand concludes it is fine.

One query is one sample, not a measurement

Ask an engine the same question twice and you get different brands and different links. So a screenshot is not evidence, and the skill treats a single run as noise by construction: each prompt is fired N times per engine and mention is reported as a frequency with its spread visible.

The published field report is, deliberately, the weak version of its own method. It was run at N=1, on 3 prompts, covering 2 of the 4 required intent families, and its header says so before you reach a single number — every rate is labelled a snapshot rather than a validated frequency, and the reader is told not to treat the blended figure as a trend point. That is the honest shape of a pilot, published rather than rounded away.

The prompt set is the actual product

Most of the thinking is not in the scoring, it is in what you ask. The skill freezes a buyer-intent prompt set across four families and keeps them cohorted so the next run is comparable:

• Category / "best" — the discovery question, where you are competing against everyone.
• Problem-first — the buyer who describes a symptom and has no vendor in mind yet.
• Comparison — "you versus a named rival", the only family where the model is guaranteed to have your name in front of it.
• Brand-defense — what the engine says when someone asks about you, which is where a hallucinated or hostile claim shows up.

The field report ran only the first and third, and it draws the obvious conclusion out loud rather than letting the gap pass: sentiment came back positive with nothing negative anywhere, but the family built to catch bad claims was not run — so that clean sheet is absence of signal, not a clean bill of health.

It also surfaces the pattern those families exist to expose. Share of voice was 9% on the broad "best PM app" prompt, 0% on "affordable alternatives", and 50% on the head-to-head comparison. Read together, that is a brand the model only knows when you put its name in the question — which is a far more useful sentence than any single percentage.

An engine you didn't query is "not captured", never zero

This is the rule the whole thing rests on, and it is boring in exactly the right way. A zero looks like a finding. It drags a blended average down, it reads as "we checked and you were absent", and it is indistinguishable in a table from a real result. So an engine with no key and no paste is reported not captured this run, and the report says which ones.

The field report turns that rule on itself three separate times:

• Perplexity is labelled ASSISTED (manual paste) rather than its usual automated bucket, because no API call actually happened that run. The label follows what happened, not what normally happens.
• The page-citability grade is deferred, not run — the sample brand had no live URL, and grading a page nobody fetched would be inventing a score. No grade is given.
• ChatGPT's and Google's 0% citation readings are flagged unconfirmed, because a missing link could equally be real product behaviour or a lossy paste, and the next capture is told to check for citation chips explicitly.

A report willing to write "we don't know" in three places is the only kind whose other numbers are worth anything.

Fix your page, or go where the engines are already reading

The action list splits in two, and the second half is the one people miss: repair your own page, or earn placement on the third-party page the engines keep quoting. The audit inventories the sources cited in the answers, so you can see whose page the model is actually building its paragraph from. Frequently that is a roundup or review site you do not own, and getting into it beats another round of on-page edits.

Notably, in the captured run every link Perplexity returned was brand-owned — the vendors' own domains, no third-party listicle or review site anywhere. That is itself a finding about which lever was available.

Where it sits next to the SEO skill

This measures a surface. Product SEO Site Builder ($9) builds one — a guide page per long-tail keyword, gated on the live SERP so you are not writing into a query you cannot win. They are complementary and they are not the same job: the audit tells you the engines describe you without linking you; the builder is one of the ways you give them something worth linking. Buying the builder first and the audit never is how you end up with pages and no idea whether anything reads them.

What $9 buys, and what it doesn't

The AI Visibility Audit is one-time, not a subscription, and the captured field report is free to read in full before you decide. What you get is the method — the frozen prompt families, the two-mode engine coverage with every mode labelled, mention scored apart from citation, share of voice against a competitor set you name, the cited-source inventory, the six-axis page grade scored only from what is actually on the page, and a re-run diff so the audit becomes a cadence instead of an artefact.

What it does not do is make an engine like you. It also cannot cover a surface you give it no way to reach — no key and no paste means not captured, and it will say so on the page rather than quietly averaging you down. If you would rather have a number that always fills every cell, this is the wrong tool. Read the field report — including the three places it admits what it could not measure — and decide from that.