You measure AI citation three ways: by running a fixed set of prompts on a schedule and recording whether you appear, by filtering your server logs for AI crawler user agents, and by tracking referral traffic from assistant domains. No single one of them is sufficient. Together they give you a usable picture.
The reason this needs explaining at all is that the obvious tool doesn't work. Google Analytics measures visits. When an assistant reads your pricing page and tells someone "expect to pay around six thousand for a build like that," nobody visits anything. There is no request, no session, no event. The most valuable thing your website did that month is invisible to the software you installed to measure your website.
So you have to go and look.
Method 1: build a prompt panel
This is the one that matters most, and it's unglamorous: write down the questions a real customer would ask, ask them, and write down what comes back.
Twenty to thirty prompts is the right size. Fewer and you're reading noise; more and you won't keep doing it. Mix three kinds:
- Category questions — "who does SEO for small businesses in California?" These are the ones where you're competing to be named at all.
- Problem questions — "my website doesn't show up on Google, what do I do?" These test whether your content gets pulled in as an explanation, which is often how you first appear.
- Brand questions — "what is Becky Wexlin Creative?" These test whether the model knows you exist and describes you accurately. You'd be surprised how often the answer is confidently wrong.
Run them across the assistants your customers actually use — ChatGPT, Claude, Perplexity, Gemini, and Google's AI Overviews — and record three things per prompt: were you mentioned, were you linked, and was what it said about you correct.
Do this monthly. Not weekly — model behaviour moves over weeks, and testing more often just gives you variance you'll misread as progress. Keep it in a spreadsheet with a row per prompt and a column per month. After three months you have a trend line, and the trend line is the whole point.
Method 2: read your server logs
AI crawlers identify themselves. They arrive with user-agent strings you can filter for, and they leave a record of exactly which of your pages they fetched and how often.
The ones worth watching:
GPTBotandOAI-SearchBot— OpenAIClaudeBotandClaude-SearchBot— AnthropicPerplexityBot— PerplexityGoogle-Extended— GeminiApplebot-Extended— Apple IntelligenceCCBot— Common Crawl, which feeds many models indirectly
What you're looking for isn't the raw count. It's which pages. If the crawlers are hitting your homepage and nothing else, your internal linking or sitemap isn't doing its job. If they're fetching your service pages and your best articles repeatedly, you're in the consideration set.
Be clear about what this tells you, though: crawling is eligibility, not success. A page can be fetched every week for a year and never once be quoted. Treat crawler logs as a leading indicator — they tell you the door is open, not that anyone walked through it.
Method 3: track AI referrals
When an assistant does link you and someone clicks, that arrives in analytics like any other referral. Set up a segment for the assistant domains and watch it as its own channel, because otherwise it hides inside "Referral" alongside everything else.
Two things to expect. First, the volume will look disappointing — it's a small fraction of the times you're actually mentioned, because most people read the answer and never click. Second, the quality will be unusually good. Someone who clicks through from an AI answer has already been told you're a reasonable fit for a problem they described in their own words. They arrive further along than almost any other channel.
Which means the right way to read this number is not as a traffic source. It's as a sample. If AI referrals are rising month over month, your citation rate is rising too — you're just seeing the visible tip of it.
What does good look like?
For a small independent business, here's roughly the progression I'd expect over the first six months of doing this deliberately:
- Month one: brand questions answered correctly. If an assistant can't accurately say what you do when asked directly, nothing else will work.
- Months two to three: problem questions start pulling your content in as explanation, usually without naming you.
- Months four to six: you begin appearing in category questions — the competitive ones — for narrow, specific phrasings first.
Broad category questions ("best web designer") are the last to move and may never move, the same way they're the hardest keywords in ordinary search. That's fine. The specific ones convert better anyway.
The one number to watch
If you only track one thing, track the percentage of your prompt panel where you're mentioned at all. Not ranked, not linked — mentioned.
It's a single figure, it's comparable month to month, it doesn't depend on any tool continuing to exist, and it moves when your work is landing. Everything else in this article is diagnosis for when that number isn't moving.
If you want the underlying mechanics of why a page gets picked up in the first place, I wrote about that in what answer engine optimization actually is. And if you'd rather just see how your own site currently reads to a model, there's a page here written specifically for that.