AI Overview Rank Tracking

Three questions, and only two of them have answers

AI Overview rank tracking is the attempt to apply the logic of keyword rank tracking — a keyword list, a repeated measurement, a position number, a trend line — to Google's AI Overviews, the generated summary block that appears above or among the classic results on a Google search page. In practice it means answering three questions per keyword on a schedule.

• Presence. Does this keyword trigger an AI Overview at all?
• Inclusion. Is my domain cited as a source inside it?
• Placement. Where does my citation sit within the block, and where does the block sit relative to the classic results?

The first two are well defined and can be sampled, with care about how many samples it takes. The third is where the analogy to rank tracking stops working, and it does not stop working because the measurement is difficult. It stops working because Google has documented that the thing being measured does not exist in the form practitioners assume.

One position for the whole block

Google's counting documentation contains a single sentence that governs this entire subject: "An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position." There is no first, second or third place inside an AI Overview as far as Google's own reporting is concerned. Every citation in the block shares one position number.

A tracker can still report that your link appeared third in the visible source list, and that observation is real in the sense that a scraper saw it. What it is not is a rank. It is not a quantity Google publishes, not a quantity Google says means anything, and not a quantity any Google surface will confirm or contradict. Improving it is therefore optimization against a target set by a scraper's parsing order.

The reason the metaphor spread is worth naming, because it explains why the idea will keep coming back. Rank tracking is the industry's native way of thinking about search, the source list has a visible order on screen, and "we moved from citation four to citation one" is a sentence that survives a client meeting. None of that makes the number a measurement. What would settle the question is a Google-published per-citation position or a documented ordering principle, and as of 29 August 2026 neither exists.

The honest formulation is that citation is binary per generation. You were in that answer or you were not, and it only becomes a rate — a proportion of generations in which you appeared — when you sample the same keyword repeatedly and say how many times you did.

Presence rates are the best-evidenced numbers here, and the most abused

A presence rate is the share of tracked keywords that returned an AI Overview at all. It is a results-page parsing problem rather than an inference, which is why the established rank trackers added it quickly and why presence figures rest on millions of keywords rather than on curated prompt lists. These are the most solid numbers in AI search measurement.

They are also unstable in aggregate over months, in a way that is well documented and badly misread. Semrush's panel of more than 10 million keywords recorded presence at 6.49% in January 2025, peaking at 24.61% in July, and back to 15.69% by November — close to a quadrupling followed by a retracement of roughly a third inside eleven months. Presence is a Google product decision that moves in both directions.

Two errors follow from this series, and both are common. The first is quoting a presence rate without its month: the same panel produced 6.49%, 24.61% and 15.69% within a single year, and a screenshot of any one of them has been used to argue both that AI Overviews are everywhere and that AI Overviews are retreating. The second is comparing presence rates across vendors, which compares different keyword panels, countries, devices and parsers rather than different states of the world.

One further caution, because the research does not support what is often built on it. Semrush's published analysis is monthly aggregate and gives no day-to-day volatility figure for an individual keyword. Anyone asserting a specific daily flip rate for a single keyword's AI Overview presence is asserting something that study did not measure, and no source establishing one was found.

Inclusion on one keyword is close to a coin flip

Inclusion tracking is a citation-extraction problem: parse the links in the block, check for the tracked domain, record a result. Mechanically it is easy. Statistically it is the hardest thing on the page, because the block is regenerated rather than retrieved.

Ahrefs, working from 540,000 query pairs of September 2025 US data, reported on 15 December 2025 that 45% of AI Overview citations change between generations. Regenerate the same overview and close to half the source list is different. Nothing about the site changed, nothing about the index changed, and the answer cites different pages.

The practical consequence is direct. A domain can appear on Monday, vanish on Tuesday and reappear on Wednesday with no underlying cause, and a daily inclusion check on a single keyword is close to a coin-flip-grade instrument. Reporting "we lost our AI Overview citation" from one observation is reporting a draw from a distribution as though it were an event, and the same applies in the other direction to a citation that appears the week after a content change.

What carries signal is a rate over many keywords and many samples, with the sample count stated. What does not carry signal is any single reading, however carefully it was collected.

The organic top ten stopped being a proxy

For roughly two years the working assumption was that AI Overview citations were drawn largely from the classic top 10, which would have made ordinary rank tracking an adequate stand-in for citation tracking. That assumption was true, and it is no longer.

Ahrefs measured it twice. On 21 July 2025, 76.10% of cited pages ranked in the organic top 10, with 9.5% at positions 11 to 100, 14.4% outside the top 100, and a median organic position of 2 for the first citation. On 2 March 2026, across 863,000 keyword result pages and 4 million AI Overview URLs, that share had fallen to 37.9%, with the remainder splitting roughly evenly between positions 11 to 100 and beyond position 100 (Ahrefs, March 2026).

Two caveats belong with those figures and are usually dropped. Ahrefs states that the second study used improved parsing, so some of the movement may be measurement change rather than Google change, and the study does not quantify how much. The two runs are not a clean time series. Separately, this is one vendor measuring itself; the direction is well evidenced, the exact multiplier is approximate.

The direction is also what query fan-out would predict. If citations are selected against sub-queries the user never issued, there is no particular reason they would come from the visible query's top 10. Either way the operational conclusion holds: tracking organic position as a proxy for citation eligibility was defensible in mid-2025 and is not defensible now, because roughly a third of citations come from pages your rank tracker is not watching at all.

Why a tracker's count will never reconcile with Search Console

Teams routinely try to line up a third-party tracker's citation count against Search Console impressions and treat the gap as an error in one of them. It is not an error. They are measuring different events by construction.

Brodie Clark's crowd-sourced test, first published on 18 June 2024 and updated on 8 June 2026, found that AI Overview impressions only count when citations are scrolled or expanded into view rather than automatically on page load, and identified three tiers of citation visibility: visible, requiring a "show more" expansion, and requiring full expansion (Brodie Clark). A scraper expands the block. Most people do not. So a tracker records citations that Google may never count as impressions, and the gap will be widest exactly where citations are buried deepest.

Search Console will not close the gap from its side either. The generative AI performance report launched on 3 June 2026 carries impressions only, across four dimensions — pages, countries, dates and devices — with no position, no clicks, no queries, and no filter separating AI Overviews from AI Mode. Any number taken from it and described as AI Overview impressions is AI Overviews plus AI Mode.

That is also why an AI Overview click-through rate cannot be computed from Google's data. The clicks are counted — Google documents that clicking an external link in an AI Overview counts as a click — and then pooled into the Web search type with no way to isolate them. Organic click-through rate, which every search team has calculated since the Search Console query report existed, has no equivalent here. The division is undefined.

What can honestly be reported instead

None of this makes the practice pointless. It makes a short list of things reportable and a longer list not.

Reportable, with the method stated: presence rate across a defined keyword set, always with its month and its panel; inclusion expressed as a proportion of sampled generations rather than a status, with the number of samples on the same line; the set of URLs that have ever been cited for a keyword group, which is a more stable object than any single reading; and diagnostic work on which pages are being fetched at all, which is a different question from which are being selected.

Not reportable: a position within the block, an AI Overview click-through rate, a day-over-day change on one keyword, a presence comparison across vendors, and any reconciliation between a scraper's count and Search Console's impressions.

Two contextual facts should sit on any such report. Since 27 May 2026 a user's Preferred Sources selections influence AI Overviews and AI Mode, so no scraper's logged-out view is any real person's view. And country and device are dimensions in Google's own report precisely because results differ across them, which makes a tracker running from one data center in one country a measurement of one cell in a large table.

Frequently asked questions