Nobody can measure GEO the way they measured search engine optimisation, and the reason is structural rather than a gap in the tooling. The old metrics assumed a click. Position was worth knowing because position predicted traffic, and traffic predicted revenue. When an answer is delivered without a visit, that chain breaks at the first link and every number downstream of it keeps reporting as though nothing changed.

That is the dangerous part. Rankings look fine. Sessions look survivable. The dashboard is green while the thing you actually care about is going to zero, and nothing in a standard report is shaped to tell you.

The signature to watch is impressions rising while clicks stay flat, measured per query rather than sitewide. In an aggregate it is invisible, because a few pages growing normally will mask it. On this site, over the 28 days to early August 2026, 531 queries sat in the top ten and produced a 0.56% click-through rate, with 452 of them returning no clicks at all. Position was never the problem.

How to Measure GEO Without a Rank Tracker

Start from what a retrieval system does, not from what a rank tracker reports. Three things are worth separating, and most reporting collapses them into one.

Retrieval. Whether your page is a candidate at all. Impressions are the closest available proxy, and they still work.

Citation. Whether the system chose your page as the thing worth quoting or naming. This is the number that matters now and the one no tool reports directly.

Referral. Whether anyone arrived. Still worth tracking, and no longer the same question as the first two.

A page can be strong on the first and absent from the second indefinitely. Reporting that only shows the third concludes the content failed, when in fact it was retrieved, used, and never credited.

What Search Console Actually Gives You

More than most people use, and less than you want.

Per-query impressions against clicks is the core signal. Export it, sort by impressions, and look at the click column. Queries with high impressions and no clicks are where an answer is being delivered without you. That list is your working set.

The generative AI performance reporting shows where you appear in AI surfaces, which is the closest thing to a citation metric that exists in a first-party tool. The Search Console performance report documents what each column counts, which matters here because impressions and clicks are defined differently than most people assume. Treat it as directional rather than exact, and watch the trend rather than the absolute.

Position is now a diagnostic, not a goal. Position four with no clicks tells you something specific and useful: you are relevant and not chosen. That is a content and authority problem, and it is a different problem from ranking twentieth.

Segment commercial queries from informational ones before drawing conclusions. A lost click on a research query costs little; a lost click on a buying query is the whole business. Our guide to why content ranks but never gets cited covers what causes the gap.

Testing Directly Is Unscientific and Worth Doing

Ask the assistants the questions your pages target, and see whether you are named.

This has obvious problems. Results vary between sessions, personalisation interferes, and you cannot sample enough to call it data. It is still more informative than any rank tracker, because it answers the actual question rather than a proxy for it.

Do it with structure so it is repeatable. Write down twenty questions your best pages should answer, run them across the assistants your buyers use, and record whether you appear, whether a competitor appears, and what source the answer credits. Repeat monthly. The absolute numbers mean little; the direction over three months means a great deal.

The competitor column is the most useful one. Being absent tells you nothing about why. Seeing who was chosen instead tells you whether it is authority, specificity or format, and those need different fixes.

Metrics That Survive the Change

Branded search volume. If people encounter you inside answers and later search your name, that shows up here. It is the cleanest evidence of un-clicked exposure doing something, and it is a first-party number you already have.

Direct traffic and returning visitors. Blunt, noisy, and the only trace some of this leaves.

Enquiries per query cluster, not per session. Attribution is degrading and will keep degrading. Tracking which topics produce conversations is more durable than tracking which session produced a form fill.

Share of answer. For a fixed question set, how often you are named. Manual, slow, and the closest thing to a real GEO metric available. The full method is in our generative engine optimisation guide .

Reporting This Honestly

The hardest part is not the measurement, it is telling somebody that traffic fell and the work was still correct.

Report the split explicitly: retrieved, cited, clicked. Showing that impressions grew while clicks did not is a finding, not an excuse, provided you also say what you intend to do about it. Pair every citation claim with the question it came from, because “we appear in AI answers” is unfalsifiable and “we are named for these six questions and not these fourteen” is a plan.

And be honest about what cannot be known. There is no complete citation dataset, no reliable share-of-voice number, and anyone selling one is selling a model. Saying so protects you when the model turns out to be wrong.

Mecanik reports this way as part of our SEO audit work, separating commercial from incidental traffic and reporting citation visibility next to rankings rather than reporting a position and calling it a result.

Related reading: Content Refresh: When Updating Beats Publishing , Entity SEO: Teaching Search Engines What You Are , Does llms.txt Do Anything Yet? The Evidence and Content Pruning: When Deleting Pages Raises Traffic .

Frequently Asked Questions

How do you measure GEO performance? Separate three things that ordinary reporting collapses into one: retrieval, meaning whether you were a candidate, for which impressions remain a good proxy; citation, meaning whether the system chose your page as the thing worth quoting; and referral, meaning whether anyone actually arrived. A page can be strong on the first and absent from the second indefinitely.

What does rising impressions with flat clicks mean? That an answer is being delivered above the results without you. It is the clearest signature of the shift, and it is invisible in a sitewide aggregate because normally growing pages mask it. Look per query. On this site, over the 28 days to early August 2026, 452 of 531 top-ten queries returned no clicks at all.

Can Search Console show AI citations? Partially. Its generative AI performance reporting shows where you appear in AI surfaces, which is the closest citation metric available in a first-party tool, but treat it as directional rather than exact and watch the trend rather than the absolute value. Per-query impressions against clicks remains the more reliable core signal.

Is manually asking AI assistants a valid test? It is unscientific and worth doing anyway, because it answers the real question rather than a proxy. Make it repeatable: fix twenty questions your best pages should answer, run them monthly, and record whether you appear, whether a competitor appears, and which source was credited. The competitor column is the most useful part.

Which metrics still work after the shift? Branded search volume, which captures un-clicked exposure turning into later interest; direct and returning traffic, blunt but sometimes the only trace; enquiries tracked per query cluster rather than per session, since attribution keeps degrading; and share of answer across a fixed question set, which is manual, slow and the closest thing to a real GEO metric.