Measuring visibility when the answer replaces the click

Every analytics setup in common use assumes the visit is the unit of success. Answer engines break that assumption in a specific way: the value is delivered without a session, so the system that measures value goes quiet while the thing it measures is working.

Which metrics stop meaning what they meant

Read impressions against clicks, segmented by intent

The useful analysis is a two-by-two: whether impressions are rising or falling, and whether clicks are. Each quadrant means something different, and only one of the four is bad news.

                 clicks up            clicks down
              ┌────────────────────┬────────────────────────┐
impressions   │ Working. Nothing   │ Being answered on the  │
up            │ to investigate.    │ results page. Check    │
              │                    │ branded search next.   │
              ├────────────────────┼────────────────────────┤
impressions   │ Fewer, better      │ Losing visibility.     │
down          │ queries. Check     │ The only quadrant that │
              │ which ones you     │ is straightforwardly   │
              │ lost first.        │ a problem.             │
              └────────────────────┴────────────────────────┘

Run this per query group, never in aggregate. A site-wide view
averages the four quadrants into a flat line that hides all of them.

Segmenting by intent is what makes the top-right quadrant readable. Informational queries losing clicks while gaining impressions is the new normal. Commercial queries doing the same thing is a conversion problem, because nobody's buying decision was resolved by a snippet.

Branded search is the best available proxy

When someone reads your answer without visiting, the trace it leaves is that they later search for you by name. Branded query volume is therefore the closest thing to a measure of uncredited visibility, and it is available in the tools you already have.

1. 01Split every query into branded and unbranded, with a consistent rule including misspellings and your product names.
2. 02Track branded impressions and clicks as a separate trend from the unbranded ones.
3. 03Watch for branded volume rising while unbranded clicks fall — that pattern is people encountering you without clicking, then coming back deliberately.
4. 04Record direct and dark-social traffic alongside it, since an answer read in an assistant often results in a typed URL rather than a referred visit.

Manual checks, because analytics covers almost none of this

Only Bing Webmaster Tools reports how often an AI assistant cites you, covering Copilot, Bing's AI summaries and a few partner integrations; Google counts AI Overviews inside Search Console without separating them. For every other assistant, periodic manual observation is the only source of data, and a structured version of it is genuinely informative.

• Keep a fixed list of ten to twenty questions your best pages answer, phrased as a person would ask an assistant
• Run the same list across the major assistants and search engines on a fixed schedule — monthly is enough
• Record, per question: whether you were cited, who else was, and whether the answer given was correct
• Keep the question wording identical between runs, or the comparison measures your phrasing rather than your visibility
• Note incorrect answers about your own offering separately; those are the ones worth acting on first

That last point is the one with immediate commercial value. An assistant confidently stating something wrong about what you do is costing you enquiries in a way no dashboard will surface, and the usual cause is that the correct information is not stated plainly anywhere on your own site.

What to put in the report

The temptation is to add metrics until something looks good. A defensible report is short and names its uncertainty.

The last row is the one that keeps the rest honest. A visitor who arrives already knowing what you do, because an answer engine told them, converts differently from one arriving cold. If enquiry volume holds and enquiry quality rises while sessions fall, the programme is working and the traffic chart is simply measuring the wrong thing.

Setting expectations before the numbers move

This is worth stating in advance, because it is unconvincing after a decline. Nobody can guarantee a snippet, a citation or a mention — placement is decided by systems that publish no criteria and change without notice. What can be committed to is a page that answers the question better than the ones currently being cited, in a form that can be extracted.

Anyone promising a fixed number of AI mentions is either misunderstanding the mechanism or misrepresenting it. The measurable commitment is the work and the observation; the placement is an outcome, reported honestly whichever way it goes.