An AI Overview tracker answers a question that Search Console cannot: when the generated answer appears above your result, and what it cites. The useful version of that tracker measures four things across a frozen keyword set, on a schedule, and stores the raw result page so you can ask new questions later.
We built the first snapshot on September 12, 2026 and pointed it at 40 SEO keywords, United States, English, desktop. Four of them returned an AI Overview. The rest did not, including most of the queries the industry currently worries about.
That single number reframes the work. Before you build a reporting line for AI Overviews, it is worth knowing how rarely they appear on the queries you actually own.
What the tracker has to record
Presence alone is a weak signal, because a generated answer that cites you is a different event from one that cites three competitors. Four fields per query per day is enough to start.
- Presence. Did an AI Overview appear at all, yes or no.
- Position. Where it sat in the result list, in absolute terms. An overview at position 1 removes the top organic slot. An overview at position 4 does not.
- Citations. How many sources the overview referenced, and which domains they were.
- Organic overlap. Whether the domains cited by the overview also rank in the top three organic results, and whether you are one of them.
The fourth field is the one teams skip and then need. Without it you can see that AI Overviews increased and still not know whether that is a threat or free distribution.
The method, and why the raw payload matters
The snapshot came from one SERP API request per keyword, using the advanced organic endpoint with the location set to the United States, language English, desktop, and a depth of 20. We captured the full response for each query rather than a summary, because the fields that matter later are the ones you did not think to extract today.
The item type you are looking for is ai_overview. When it is present it arrives as its own element in the result list with its own rank_absolute value, alongside the organic items. That is the entire detection rule, and it is why a tracker built on a scraper that only reads organic results cannot answer AI Overview questions at all. If you already run a two-source tracker for positions, this is the same data pull with one extra element parsed, which we covered in the rank tracking API build.
The snapshot, September 12, 2026. 40 SEO queries, United States, English, desktop.
Query | AI Overview position | Citations | Top 3 organic domains |
|---|---|---|---|
how to check google rankings | 1 | 17 | seobility.net, developers.google.com, keyword-tools.org |
claude code seo | 4 | 13 | github.com, claude-seo.md, reddit.com |
google search algorithm update | 1 | 10 | status.search.google.com, developers.google.com, seroundtable.com |
ai search optimization | 1 | 13 | developers.google.com, npdigital.com, salesforce.com |
Four of 40 queries returned an AI Overview, which is 10%.
The split followed query shape, not difficulty. No query that named a tool or an API returned an overview.
What the other 36 queries had in common
The queries without an AI Overview were not random. They clustered, and the clustering is the useful finding for anyone deciding where to spend effort.
Splitting the set by query shape makes it concrete. Fifteen of the 40 queries named a tool, an API, or a data source. The remaining 25 were questions, definitions, and news queries. The AI Overviews fell entirely on one side of that line.
Query group | Queries | With an AI Overview |
|---|---|---|
Names a tool, API, or data source | 15 | 0 |
Question, definition, or news query | 25 | 4 |
Tool and API queries had none. rank tracking api, serp monitor, google search console api, google search console mcp, ai overview checker, ai visibility checker, and seo mcp server all returned a standard result page. These are commercial-investigation queries, and Google currently answers them with a list of products and documentation rather than a generated summary.
Narrow technical queries had none. serp volatility, core web vitals, schema markup checker, keyword difficulty, and featured snippet optimization were all clean. The interpretation we favour is that these queries have a well-defined answer that a ranked list serves better than a paragraph, and Google appears to agree on this sample.
Broad how-to and news queries did have one. The four that triggered are all wide questions: how to check rankings, what changed in the algorithm, what AI search optimization is, and how to do SEO with Claude Code. That last one is interesting precisely because it is new vocabulary. AI Overviews appeared on a query whose answer set is still forming.
Citations were diffuse, with two exceptions. Across the four overviews the tracker collected 53 citation slots across 24 domains, and no domain took more than a fifth of them. YouTube received 8 slots and appeared in three of the four overviews. developers.google.com received 6, and so did claude-seo.md, an agent-readable documentation file that showed up only on the Claude Code query. Fourteen of the 24 domains received a single slot. There is no shortlist of publications that reliably earns AI Overview citations in this niche, which means a citation strategy built on copying a competitor's placement is guessing.
Citation sets were diffuse. Fourteen of the 24 domains received a single slot.
What this does and does not prove
One day, one country, one device, 40 queries. That is a snapshot, not a trend, and the honest framing is that it establishes a baseline you can compare against rather than a law about Google.
Three specific limits are worth writing into the tracker itself.
Personalisation and location move the answer. The same query can return an overview in one city and not another, and a signed-in account sees different results from an anonymous request. Fix location and device in the request, record them in the output, and never compare across a change in either.
Presence is not impact. The tracker tells you an overview appeared. It does not tell you that it cost you clicks. Search Console tells you that part, which is why the useful report joins the two: overview present, position held, clicks down. If clicks are flat while an overview sits above you, the overview is not the problem you thought it was. This is the same discipline as separating a real drop from ordinary movement.
The query set defines the conclusion. We chose SEO keywords because that is our subject. A different set, for example commercial product queries in retail, would produce a much higher share of overviews. If you publish the number, publish the query list with it.
Building the rest of the tracker
The snapshot is the easy half. The value arrives when the same 40 queries are pulled daily and the differences are stored.
Freeze the list. Changing keywords between runs makes the trend meaningless, because a change in the share could be a change in your list. Add keywords deliberately, in a new group, and keep the original group intact.
Store raw, summarise second. Keep the full response per query per day for at least 90 days. A summary table of presence and position cannot answer the question you will have next quarter, which is always about composition: which source got cited, which element moved above the fold, whether video took the slot that used to be yours.
Log four numbers daily. Share of queries with an overview. Average citation count. Share of your tracked queries where you are cited. Share where the top organic result is also cited. Those four move for different reasons, and separating them stops you from reacting to the wrong one.
Set a recheck rule, not an alert. The natural instinct is to alert the moment presence changes, which produces noise for the same reason fixed ranking thresholds do: the underlying data moves daily. Recheck a shift after two consecutive runs, and treat a one-day move as unconfirmed. The band design we use for ranking signals, including the thresholds and the reasons they are per-query rather than global, is in the monitor guide.
Auspia view: AI Overview tracking is a measurement project, not a content panic. Ten percent of the SEO queries we checked had an overview, and the queries that did were broad informational questions. The teams that will get this right are the ones that can say which of their queries have an overview, whether they are cited, and what happened to clicks, in one table they trust.
FAQ
How do I check whether a query has an AI Overview? Request the result page from a SERP API and look for an element of type ai_overview in the response. If your only data source is Search Console, you cannot see it directly, though a click drop with a stable position is the pattern that usually points to one.
How often should an AI Overview tracker run? Daily if the keyword set is small, weekly if it is large. Presence changes less often than position does, so a weekly cadence captures most of the movement at a seventh of the cost. Our own daily movement data, measured separately, is a useful reminder that fine-grained daily readings on a large set mostly measure noise.
Why did so few queries have an AI Overview? In our set, the queries without one were tool, API, and narrow technical questions. The four with an overview were broad informational and news-shaped questions. Our sample covers SEO topics only, so do not generalise the 10% figure to other industries without running the same test on your own query list.
Should I try to get cited in AI Overviews? Being cited is worth having and hard to engineer directly, because the citation sets we observed were diffuse. The more reliable move is to be the page that answers the question in plain language, which is also the page that ranks, and to measure citations rather than assume them.
Can I track AI Overviews for competitors? Yes, and it is more informative than tracking only your own brand. The tracker records cited domains for every keyword, so you can see which competitors are being pulled into the overview while you are not, which is a clearer signal than your own citation count alone.
Author: Priya Nair, AI Search Measurement Lead Across 12,000 Tracked SERPs at Auspia. Priya writes about AI Overview measurement, citation patterns, and the difference between a change in the results and a change in your traffic.
