Classic SEO is measurable because it ends in a click. GEO frequently does not: the engine answers the question, names you, and the person never visits. That is a good outcome and an invisible one, which is why AI citation tracking has to be built deliberately rather than read off a dashboard.
Why analytics cannot see most of it
Analytics records visits. When an AI engine answers a question directly, there is often no visit to record — no referrer, no session, no event. The recommendation happened entirely inside a conversation your tools never observed, so absence of traffic says nothing about absence of influence.
Referral traffic from AI tools does exist and is worth segmenting, but treat it as a floor rather than a measure. It counts the people who clicked through, not the people who were satisfied by the answer — and the second group is usually larger.
Prompt monitoring: the core method
The workable method is to ask the engines directly. Pick the questions your customers actually ask, run them across each engine on a fixed schedule, and record what comes back. That converts an invisible outcome into a repeatable measurement you can act on.
A prompt set should reflect real buying language, not your internal vocabulary. Mix unbranded questions ("what is the best tool for X"), branded ones ("is [your brand] any good"), and comparisons ("[you] vs [competitor]"). Each tests something different, and the unbranded ones are usually the uncomfortable ones.
What to record each time
For every prompt and engine, record four things: whether you were mentioned, whether you were cited with a link, what was said about you, and which competitors appeared. Those four turn a set of screenshots into something you can compare over time.
The third of those matters more than people expect. Being mentioned inaccurately is a distinct problem from being absent, and it needs a different fix — usually correcting the third-party sources an engine is leaning on rather than publishing more of your own content.
Run it on a schedule
A single check is a snapshot and tells you very little. Answers move as the underlying pages, models and sources change, so the signal is in the trend. Running the same prompt set at a fixed interval is what makes a change attributable to something you did.
Consistency beats frequency. The same prompts, the same engines, the same cadence, recorded the same way — a monthly check done identically each time is more useful than a thorough one-off.
Server logs: the other half
Your server logs show which AI crawlers are actually fetching your pages. Filtering for agents such as GPTBot, ClaudeBot and PerplexityBot tells you whether engines are reading the site at all — which is the first thing to rule out when you are never cited.
Crawler activity is a leading indicator and citations are the lagging one. If the crawlers are not arriving, no amount of content work will help until that is fixed; if they are arriving and you are still not cited, the problem is the content rather than access.
Doing it in practice
Start small and keep it honest: twenty to thirty prompts covering your main buying questions, run across the engines that matter to your audience, recorded in a spreadsheet. That is enough to see position and movement before investing in anything more elaborate.
Terrank exists to automate exactly this loop — a one-off audit measures where you stand today, and the recurring plans re-run the same prompt set so movement is visible without anyone doing it by hand.
Frequently asked questions
Why does my analytics not show AI traffic?
Analytics records visits, and an AI answer often produces no visit at all. The engine answers the question directly, so there is no referrer and no session to log — absence of traffic says nothing about absence of influence.
How often should I check AI citations?
Run the same prompt set at a fixed interval rather than ad hoc. A single check is a snapshot; the signal is in the trend. Consistency matters more than frequency — a monthly check done identically each time beats a thorough one-off.
How can I tell if AI crawlers are reading my site?
Filter your server logs for agents such as GPTBot, ClaudeBot and PerplexityBot. Crawler activity is the leading indicator: if they are not arriving, no amount of content work will help until access is fixed.