AI citation monitoring
Assistant-driven discovery mostly does not appear as a referrer. Someone asks a model, gets your name, and arrives by typing it - showing up as direct or branded search. Without deliberate checking you cannot tell whether GEO work did anything.
Use to find out whether assistants actually cite you, since this traffic is largely invisible in analytics and referrers.
The skill file
What you need first
- The 15-25 questions your buyers actually ask
- Access to the assistants your market uses
Method
- 01 Write the questions as a buyer would type them, not as keywords. "Is there an AI tool that writes and publishes SEO articles" rather than "ai seo tool".
- 02 Ask each question in each assistant on a fixed cadence, in a clean session with no history.
- 03 Record three things per run: whether you were named, which competitors were named, and which URL was cited if any.
- 04 Track the citation source. Being quoted from a third-party listicle is a different result from being quoted from your own page, and demands a different response.
- 05 Watch branded search and direct traffic for correlated movement. That is the closest measurable proxy for assistant-driven demand.
- 06 When a competitor is named and you are not, fetch the cited page and compare structure rather than word count - it is usually more extractable, not longer.
What this produces
A tracked table of question, assistant, named-or-not, competitors named and cited URL, run on a fixed cadence so change is visible.
Where this goes wrong
- Running queries in a session that already carries history, which biases the answer toward your brand
- Treating a single run as a result - answers vary between runs and models
- Expecting a referrer that will usually never exist
Use this skill in your own AI
The download is a plain markdown file with the name and trigger in its frontmatter. Where an assistant supports skills it can load itself, that frontmatter is what it reads to decide this one applies.
Questions about this skill
When should I monitor citations rather than read analytics?
When you want to know whether GEO work did anything, because analytics will not tell you. Assistant-driven discovery rarely arrives with a referrer: someone asks a model, hears your name, then types it, so it lands as direct or branded search. Asking the assistants directly is the only way to observe the answer you are trying to change.
What do I need before the first run?
Fifteen to twenty-five questions written as a buyer would type them, and access to the assistants your market actually uses. Start without the question list and you drift back into keyword strings, which nobody asks a model. You also need clean sessions: run it in an account that has been discussing your brand and it will name you regardless.
What do I end up with, and which column matters?
A table of question, assistant, whether you were named, which competitors were named, and the URL cited, re-run on a fixed cadence. The cited URL column is the one that drives work: being quoted from someone else's roundup and being quoted from your own page call for completely different responses. Movement over time is the finding, not any single row.
What ruins this most often?
Treating one run as a result. Answers vary between runs and between models, so a single mention proves nothing and a single absence panics a team into rewriting a page that was fine, which costs a cycle of misdirected work. Waiting for a referrer to confirm the effect is the other one; it will usually never exist.
More in AI search (GEO)
Answer extractability pass
Use when a page ranks well but is never quoted by AI assistants, and you need to make its claims...
Entity consistency check
Use when assistants describe your company inaccurately, conflate you with another brand, or stat...
llms.txt authoring
Use when you want to tell assistants which of your pages are authoritative, and be honest with y...
Quotable statistic sourcing
Use when your content makes claims no assistant will repeat because nothing in it is specific en...