Google AI Overviews now sit above the fold on roughly 27% of English-language searches and climbing. If you are still measuring only blue-link rank, you are missing the box that pushes every organic result below the scroll on more than a quarter of your keywords.
We have been tracking AI Overview appearances daily since the feature rolled out globally in mid-2024. The pattern is now clear enough to act on: appearance is common, citation is scarce, and the sources that get cited are not always the sources that rank. This post is how we monitor it, what we have learned about who gets picked, and the weekly cadence we recommend.
What an AI Overview actually is
An AI Overview is a Gemini-generated answer block that appears at the top of a Google SERP, above the organic results and often above the ads. It is composed of a short synthesised answer plus a set of citation cards – usually three, sometimes up to eight – linking to the sources Gemini pulled from.
The important thing to internalise: the cited sources are chosen separately from the ranking algorithm. A page ranking #7 organically can be cited card #1 in the Overview. A page ranking #1 can be absent from the Overview entirely. Treat them as two different competitions.
The citation-position bias
Across the roughly 42,000 AI Overviews we sampled across client accounts in the first half of 2026, citation share is heavily front-loaded. The first card gets the click, the second gets skimmed, the third gets ignored. Below the fold on the Overview itself is effectively dead space.
| Citation card | Share of Overview clicks | Typical CTR uplift |
|---|---|---|
| Card 1 | 58% | +11-14% vs. organic #1 |
| Card 2 | 22% | +3-5% |
| Card 3 | 11% | roughly flat |
| Card 4-8 | 9% combined | negligible |
The takeaway: getting cited is only half the win. Getting cited first is where the traffic actually is. When you audit a keyword, do not stop at "we appear in the Overview" – record the card position.
Who gets cited (and who doesn't)
The pages Gemini picks share a few consistent traits. None of them are secret, but the combination is uncommon:
- A direct answer in the first 40-80 words of the page – no throat-clearing intro.
- A specific number or claim the model can quote verbatim ("27%", "since 2024", "up to 8 sources").
- Clean article schema with an author, a publish date, and a stable canonical.
- A page that is already in Google's index and has some organic authority – Overviews do not reward pages Google does not already trust.
- Content type match: how-to queries cite how-to pages, comparison queries cite comparison pages. Category confusion kills citation rate.
The single strongest lever in our data is the first bullet. Rewriting the top of the page so the answer sits above the H2s lifts citation rate roughly 2-3x on keywords where we already rank in the top 10. It is the cheapest AI visibility work you can do this quarter – see the 2026 AI Visibility Playbook for the full rewrite pattern.
How to detect Overviews at scale
Manually eyeballing SERPs does not scale past about 30 keywords. For anything larger, you need a scraper that renders the page (Overviews are injected client-side after the initial paint) and a parser that pulls three fields: presence, citation URLs in order, and the answer text.
Our AI Visibility tracker runs this daily across all keywords in a project. If you are building your own, here is the minimum record we store per SERP snapshot:
{
"keyword": "how to track ai overviews",
"fetched_at": "2026-07-01T04:12:00Z",
"overview_present": true,
"overview_answer_hash": "a3f2...",
"citations": [
{ "position": 1, "url": "https://semoptimiser.com/blog/tracking-ai-overviews", "domain": "semoptimiser.com" },
{ "position": 2, "url": "https://searchengineland.com/...", "domain": "searchengineland.com" },
{ "position": 3, "url": "https://ahrefs.com/blog/...", "domain": "ahrefs.com" }
],
"organic_rank": 7
}Storing the answer hash – not the full answer – lets you cheaply detect when the Overview text has changed without ballooning your storage. When the hash flips, re-fetch the full answer and diff it.
The weekly monitoring cadence
Daily is overkill for reporting (the noise swamps the signal), but a weekly review is the right cadence for action. Here is the checklist we run every Monday for our own site and for retainer clients:
- Pull the week's Overview appearance rate across the tracked keyword set. Flag any keyword that gained or lost Overview presence.
- For every keyword where the Overview is present, record card 1, card 2 and card 3 domains. Compare to last week.
- Identify keywords where a competitor took your citation slot. These are your rewrite briefs for the sprint.
- Identify keywords where a new domain entered the citation set – that is Gemini expanding the source pool, and it is your window to enter it too.
- For keywords where you rank top 10 organically but are not cited, audit the page structure. Usually the answer is buried.
- Sanity-check any keyword where you rank #1 organically but a competitor is card 1 in the Overview. That is the highest-leverage rewrite you will do all quarter.
- Update the report. Share deltas, not absolute numbers – the board wants to know what moved.
What we still cannot measure
Two honest caveats. First, we cannot measure the clicks from inside the Overview with any precision – Google Search Console lumps them into the same "web" bucket as organic clicks. You can infer with UTM-less referrer analysis, but it is directional at best. Second, the Overview shown to a logged-in user with personalisation is not necessarily the Overview our headless scraper sees. Treat scraped data as the baseline, not the ground truth.
What to do next
Pick your top 50 commercial keywords this week. Run a single AI Overview snapshot across them and count three numbers: how many trigger an Overview, how many cite you at all, and how many cite you as card 1. If card-1 share is below 15%, you have a page-shape problem – start rewriting. If Overview appearance is above 40% across the set, you need weekly monitoring in place before your next quarter starts. Our AI Visibility tracker does both; if you would rather build it yourself, the schema above is the minimum viable version.