Citation vs mention: how to measure whether AI recommends you
Here is a claim worth testing. GEO is not working, because the AI referral traffic in your analytics is flat. It is the most common read, and it is wrong. The reason is one number: click-through from the body of an AI answer runs near one percent, so referral volume was never going to show the effect. What follows is what that number proves, what it does not, and the measure that survives once you stop trusting the dashboard.
Two things worth measuring
Start by separating two outcomes. A citation is your link appearing in the sources under an answer. A mention is the engine saying your name inside the recommendation. A page can cite you while the model never names you, which looks like progress in a backlink tool and does nothing for the buyer. The mention is what wins the deal, so track it in its own column.
What analytics can and cannot see
Your analytics will show some direct AI referrals. In the acquisition report, look for sources like the major assistant domains. Those are real, and worth watching month over month. What they will not show is the far larger number of buyers who saw a recommendation, did not click the citation, and arrived later through a branded search or a direct visit. Click-through from an AI answer sits near one percent, so referral traffic is a floor, not the full effect. Read literally, it undercounts your visibility by roughly two orders of magnitude, which is exactly how a working programme gets mistaken for a failing one.
| Signal | How to read it | Cadence |
|---|---|---|
| AI referral traffic | A visible floor, not the whole effect. Watch the trend | Monthly |
| Mentions for key prompts | The real scoreboard. Are you named, and in what position | Monthly |
| Citations for key prompts | Which of your pages the engine trusts as a source | Monthly |
| Share of voice | How often you appear versus named competitors | Monthly |
Run the manual audit
The reliable measure is a repeatable check by hand. Take your key decision prompts, run them in a clean session on each engine, and record for every one whether you were named, in what position, whether you were cited by link, and which competitors appeared. Do it in a temporary or incognito window so your own history does not skew the result. Save the answers in a spreadsheet.
That sheet does double duty. Paste it back into an assistant and ask it to summarise your share of voice, the recurring sources, and the sentiment of how you are described. You get a competitive read on your AI visibility for the cost of an afternoon.
Same prompts, same engines, every month.The value is in the comparison. If you change the questions each time, you cannot tell whether last month's content actually moved anything. Freeze the set and run it on a schedule.
The bottom line
So is the flat-traffic read ever right? Only if referral clicks were the goal, and they are not. The honest verdict is that referral traffic is a floor worth watching and a poor scoreboard. Judge the work by whether the engine names you for the prompts that matter, in what position, and how often against your competitors. Those three move months before the traffic does, and they are what the buyer actually sees. Anyone reading GEO off the analytics dashboard alone is measuring the one percent and missing the rest.
Frequently asked questions
Why does so little AI traffic show up in Google Analytics?
Because click-through from an AI answer is low, often around one percent, and because a recommendation can influence a buyer who then arrives through a branded search or direct visit rather than a tracked referral. The influence is real even when the click is not attributed. Judging GEO by referral volume alone badly understates it.
What is the difference between a citation and a mention?
A citation is your URL appearing in the sources an engine lists. A mention is the engine naming your brand in the answer itself. Both are useful, but they are not the same. You can be cited as a source without ever being named as a recommendation, and being named is what actually moves a buyer. Track the two separately.
How do we estimate real AI recommendation volume?
Take the AI referral clicks you can see in analytics and treat them as the visible tip of a much larger number, because typical click-through from an AI answer is about one percent. The clicks are a directional signal, not the whole picture. The more reliable measure is a repeated manual audit of whether you are named for your key prompts.
How often should we run the measurement?
Monthly, using the same prompts on the same engines in clean sessions. Consistency is the whole point. A fixed set run on a fixed cadence lets you see whether your position is improving, holding, or slipping, and ties any change back to the content and outreach you shipped in between.
See how you show up today.
The free Snapshot checks how your brand appears across ChatGPT, Perplexity, Claude and Google AI in your market, then hands you a ranked list of fixes. No card, no commitment.