AI Search Metrics Worth Tracking, And The Ones That Are Vanity

The measurement problem in generative search is not that there is no data. It is that the familiar numbers still populate, still move, and no longer mean what they used to.

Average position improves when your clicks fall. CTR collapses on pages that are working perfectly. Impressions rise on pages nobody will ever visit. Every one of those is now a normal, healthy pattern, which makes a dashboard of them actively misleading.

Here is a smaller set that still says something.

Track: Citation Presence, Checked By Hand

Once a month, ask each major assistant the five questions your buyers actually ask. Not “what is the best CRM”, but the phrasing a real person uses when they are two weeks from a decision.

Log four things per answer: whether you were mentioned, whether you were cited with a link, which sources it did cite, and how it described you. That last column is the valuable one, because it tells you whether your positioning survived compression or whether the model substituted a generic category description.

Twenty minutes a month. It is manual, it does not scale, and it currently beats every tool on the market, because tools report a proprietary score derived from sampled prompts while you are measuring the actual questions that matter to your business. The mechanics of who gets picked are in how AI assistants choose which sources to cite.

Track: Branded Search And Direct Traffic

If assistants are describing you to people who never click, the effect shows up later as someone typing your name. Branded query volume and direct sessions are the closest available proxy for unattributed influence.

Watch them as a trend over quarters, not weeks. They are noisy and they respond to everything you do, which is a limitation, not a reason to ignore them.

Track: Clicks And Conversions On Decision Pages Only

Split your pages into two groups and never average across them.

Definitional pages exist to be quoted. Falling clicks there is the expected outcome and not a problem to solve, as the zero-click research and Pew’s finding that users click a result 8% of the time when an AI summary appears against 15% without one both make clear.

Decision pages exist to be visited: comparisons, pricing, case studies, downloads. Falling clicks there is a real signal. If you measure both together, the first group masks the second and you will find out too late.

Track: Impressions Without Clicks, Deliberately

In Search Console this pattern used to mean a poor title or the wrong intent. It now often means your page is being used as a source, since an appearance in a generative response counts as an impression while the answer removes the reason to click. The counting rules are worth knowing precisely, and they are covered in what the AI performance reports do and do not tell you.

So the metric is not the ratio. It is the ratio segmented by page type. High impressions and near-zero clicks on a glossary entry is success. The same profile on a pricing page is a problem.

Do Not Track: Any Vendor Visibility Score

No assistant vendor publishes a ranking signal. Every “AI visibility score” is therefore a proxy that someone invented, computed over a prompt set someone else chose, and it will move when the vendor changes its sampling.

They are not useless for spotting large directional changes. They are useless as targets, and the moment a score becomes a goal somebody will start writing to it.

Do Not Track: Average Position Across Everything

It was already a weak aggregate. It is now actively perverse, because it can improve for the same reason your revenue falls.

Do Not Track: Word Count, Publishing Cadence, Or Keyword Density

These are inputs that were never outcomes, and generative retrieval has made them worse inputs than they were. Volume no longer buys you coverage when an answer draws on three to ten sources.

Reporting It Without Causing A Panic

The hard part is not the measurement. It is that a report full of declining familiar metrics reads as failure to anyone who has not been following the change.

Before:

Organic sessions declined 22% year over year. Average CTR fell from 3.1% to 1.9%. Impressions increased 40%. Average position improved from 14.2 to 11.6.

After:

Organic sessions are down 22%, almost entirely on glossary and definition pages that assistants now answer directly. Clicks and signups on our five commercial pages are flat year over year. Impressions are up 40% because appearing inside an AI answer counts as an impression, so the CTR decline is arithmetic rather than a quality signal. The number I am watching is branded search, which is up 8% this quarter.

Both are true. Only one lets the reader decide anything. The general move is covered in how to write a status update executives read.

A Wrivio Context for this could say:

Rewrite this metrics summary so each figure is followed by what it means for the business. Keep every number, percentage, date, and page name exactly as written. Do not add causes, forecasts, or recommendations that are not stated in the original. Where the original gives no explanation for a change, say the cause is not established rather than supplying one.

Press Ctrl+Shift+Space, paste the summary, and check the diff. The clause about not supplying causes is the one that saves you, because an explanatory rewrite will confidently attribute a decline to a competitor or an update you never mentioned.

The Uncomfortable Summary

You have less measurement than you had in 2020, and the honest response is fewer metrics reported with more context, not more metrics reported with the same confidence. A dashboard that pretends the old certainty still exists is worse than a paragraph that admits what is unknown.

Common Questions

Is rank tracking dead?

Not dead, but much weaker. Overlap between AI-cited sources and the organic top ten has fallen substantially since 2025, so a rank position no longer predicts whether you appear in the answer.

What is the single best metric to start with?

The monthly hand-check of how assistants describe you on your five key buying questions. It is the only measurement that captures framing, which is what a summary can most easily get wrong.

Should I buy an AI visibility tool?

Only for directional monitoring at scale, and only if the vendor publishes its prompt set and methodology. Never as a target to optimise against.

How do I know whether a decline is my fault?

Segment by page intent first. If definitional pages fell and commercial pages held, the market changed. If commercial pages fell, something on your side did.

Download Wrivio for Windows to turn a page of raw metrics into a report your leadership can act on, without a rewrite inventing the cause.