Tracking AI Overviews and Where Your Citation Sits

The tracking problem is that presence is unstable. The same query, checked twice within an hour from the same location, may produce a generated summary once and not at all the second time. That makes any single observation nearly worthless, and it makes “does this keyword have an AI overview?” the wrong shape of question.

The right shape is a rate: over the last thirty checks, how often did a summary appear, and how often were we cited in it?

Why presence is unstable

Generated summaries are not a fixed layout element attached to a query. Whether one appears depends on the query, on how the system currently rates its own confidence, on the surface and device, and on ongoing changes to when the feature triggers at all. Coverage has also shifted repeatedly since these summaries started appearing, in both directions.

Practically, that means:

• Two trackers can honestly disagree about whether a keyword has a summary, because they checked at different moments.
• A summary disappearing is not necessarily a loss you caused, and its appearing is not necessarily a gain.
• A single screenshot proves nothing, which is awkward, because a screenshot is exactly what a stakeholder will bring you.

If you take one thing from this: record every check, not just the latest state. The series is the data.

What is actually worth measuring

Three things, in decreasing order of reliability.

Trigger rate. The proportion of checks in which a summary appeared for a given keyword. This is stable enough to trend if you check consistently, and it tells you whether a keyword is in feature-heavy territory at all.

Citation presence. Whether your domain appears among the sources when a summary does appear. Report it as a proportion of the checks where a summary appeared, not as a proportion of all checks, or a falling trigger rate will look like a citation loss.

Position within the summary’s sources. The least stable of the three. Source ordering moves around, and the list length varies. Worth recording, not worth alerting on.

Notably absent from that list: any click or traffic figure attributable to the summary. Whether a summary produced a visit is not something you can observe from outside, and any confident number in that shape is a model, not a measurement.

Keeping it separate from organic position

The mistake to avoid is folding summary presence into your organic position metric. They’re separate measurements with different volatility, and blending them produces a number that jumps for two unrelated reasons.

Keep organic position as the primary series. Track summary presence as a secondary attribute on the same keyword, the same way you’d track whether a local pack is present. The relationship between them is genuinely interesting — being cited without ranking well organically happens, and so does the reverse — and you can only see that relationship if the two are recorded separately.

This is the same discipline as the rest of feature tracking, described in what position one means on a crowded result page.

Interpreting a change in traffic

When a keyword’s clicks fall and a summary has appeared, the honest position is that you have a correlation and a plausible mechanism, not a demonstrated cause. Other things also changed: layout, ad load, seasonality, competitor pages.

What you can do is narrow it. Check whether clicks fell on that keyword specifically or across the cluster. Check whether impressions held — if impressions are flat and clicks fell, something on the page changed rather than your ranking. Check whether the effect coincides with your own release notes. The diagnostic order is the same as for any drop — how to diagnose a ranking drop.

What this does not change

The reason a page gets cited is not a new discipline. It’s the same set of properties that made a page a good organic result: it answers a specific question directly, it’s clearly structured, and the claim a summary would want to quote is stated plainly rather than buried in a paragraph of setup.

Nobody outside the systems themselves knows the selection rules, and treating speculation about them as technique is how a lot of confident, wrong advice gets written. Write pages that are easy to quote accurately, and measure the citation rate rather than theorising about it.

What to actually do

1. Check on a fixed schedule and keep every observation. Trigger rate needs a series; a current-state field is not enough.
2. Report citation presence as a share of summary-present checks, not of all checks.
3. Never merge summary presence into organic position. Separate fields, separate charts.
4. Don’t alert on it. Volatility is high enough that alerts will fire constantly and mean nothing — rank alerts that don’t cry wolf.
5. Refuse to attribute traffic to it numerically. Describe the effect qualitatively and say what you can’t know.