A pattern we now see in most engagements: the team has AI mention data in one place, Search Console and analytics in another, and no defensible way to put them on the same slide. Someone senior has asked whether AI search is working, and the honest answer is currently a shrug.
The two datasets genuinely do not share a key. That is a real problem, not a tooling gap, and pretending otherwise produces reporting that falls apart the first time someone interrogates it.
Why they do not join
Classic organic reporting has a clean chain: a query produced an impression, which produced a click, which produced a session, which produced a conversion. Every step has an identifier.
AI answers break it in three places. A person asking ChatGPT about your category generates no impression in your Search Console. If they are convinced, they frequently arrive later by typing your name into Google, so the visit is recorded as branded direct or branded search. And the conversation that did the persuading leaves no trace on your property at all.
So the influence is real and the attribution is missing. Any model that claims otherwise is inferring, and it is worth being explicit about which parts are measured and which are inferred.
What can honestly be joined
Three joins hold up under scrutiny.
Cited sources to your own pages. If your visibility tool records the URLs cited in answers, and some are yours, that is a direct link between the two datasets. You can see which pages get cited, cross-reference their organic performance, and learn what the cited ones have in common. This is the most useful join available and it needs no modelling.
Branded search volume as a downstream proxy. AI-influenced discovery tends to surface later as branded demand. Branded impressions in Search Console, tracked over time and against your mention rate, is a rough but defensible indicator. Rough because a dozen other things move branded volume, so treat direction over several months as the signal and never a single month.
Query themes on both sides. Take the prompt themes where you are absent from AI answers and check the equivalent unbranded queries in Search Console. Where you are weak in both, you have a coverage problem. Where you are strong organically and absent in AI, you have a corroboration or structure problem: the material exists and is not being retrieved. That second pattern is common and it is the one worth acting on first.
Watch for the branded CTR trap
One diagnostic worth running before anything else, because it is frequently misread as an AI effect.
Take your branded queries in Search Console. Compare average position against CTR over the last two quarters. If position is flat and CTR has fallen materially, something is appearing on your brand term and taking clicks. That could be an AI Overview summarising you well enough that nobody clicks. It could equally be a competitor, or another company with a similar name being confused with yours.
Those have opposite responses, and the data looks identical until you go and look at the actual results page. Do that before you attribute the decline to AI.
What to report
Report the three realities separately rather than collapsing them into one number. It survives questioning, and it maps onto decisions:
| What you report | Where it comes from | What it tells you |
|---|---|---|
| How the market describes us | AI answers, sampled | Whether Internet Reality matches the business |
| Where we are absent, and who is there instead | Prompt set, by theme | Which questions we have no claim on |
| Which sources the systems trust | Cited URLs | The surface worth influencing |
| Branded demand over time | Search Console | Whether awareness is moving |
| Organic performance | Search Console, analytics | The measured part, unchanged |
The instinct is to build a single AI visibility KPI because a board wants one number. Resist it for a couple of quarters. A composite score built on an attribution chain that does not exist will be challenged, and when it is challenged it will not hold.
What holds is showing the distance between what the company is and what the market can find, then showing that distance closing. That is measurable, it is honest about which half is sampled and which is inferred, and it is the argument that survives the meeting.