How to Track AI Overviews for Your Keywords

Tracking AI Overviews means recording, for each keyword you care about, whether Google shows an AI Overview, whether your page is cited inside it, and what that does to your clicks. Search Console’s new Generative AI report counts how often links to your pages were shown inside AI Overviews and AI Mode. It does not say which query showed the Overview, whether it was an Overview or AI Mode, or who else was cited.

Those come from checking each keyword on a live Google results page with the delayed Overview loaded. Clicks come from the Performance report per query, kept as a baseline so you can see how many clicks a query still sends you after an Overview appears. The check is re-run every month.

• Read citation and position together. Google assigns every cited link the Overview’s slot, so a cited page reports a better position in Search Console. On windsor.ai, four in five queries where we rank in the top 10 organically are also cited, and one in five of the rest.
• If you check through a SERP API, a scraper or an LLM, make it wait for the delayed Overview, or the count is short. A check without the wait would have missed 6 of our 30 citations.
• Count clicks separately from citations, because a citation does not keep the click. One of our guides is cited on both of its head queries and lost more than half of its monthly clicks between June and August.
• Keep your own monthly log. I checked Ahrefs and DataForSEO for earlier results pages. Both store them only for keywords already in their databases, and for our list that meant a snapshot for 17 of the 62 queries.

The numbers in this article come from the 62 non-branded queries that send the most impressions to a subset of windsor.ai’s pages, each checked on a live Google results page through DataForSEO’s SERP API in September, next to 90 days of our Search Console data and the Generative AI report for 20 June to 19 September.

What Search Console now shows about AI Overviews

Since June 2026 Search Console has a Generative AI performance report under Performance, Search results. Google rolled it out to all sites on 31 August. It reports impressions from AI Overviews and AI Mode together, by page, country, device and date (down to the hour), with an export button. An impression here is a link to your page shown inside one of those features. It does not report clicks, queries or position, and it does not separate AI Overviews from AI Mode. The same impressions are still counted inside the main Performance report, under the Web search type, with no filter there either.

Three rules from Google’s documentation decide how those impressions look in the main report:

• An AI Overview occupies one position on the results page, and every link inside it is assigned that same position.
• A link in the Overview counts an impression under the standard rules, whether or not anyone expands the Overview.
• A click on a link inside the Overview counts as a click.

So a page cited in an Overview records a position of 1 or 2 and an impression on every search, and almost no clicks. The Generative AI report tells you how many of a page’s impressions came from AI features. It cannot tell you which query showed the Overview, whether it was an Overview or AI Mode, or who else was cited beside you. Those need the live check in step 3. Clicks per query come from the Performance report, in step 4.

Google also states that a page needs no special treatment to appear in an Overview beyond being indexed and eligible for a snippet. Step 5 shows what that looks like on our queries.

Step 1: read the Generative AI report next to the Performance report, per page

Do this. Open the Generative AI report for the last three months, export the Pages table, and put it next to total impressions and clicks per page from the Performance report for the same dates. The share of each page’s impressions that came from AI features is your first ranking of where to look. Pages with a high share and few clicks are answered on the results page. Pages with a high share and normal clicks are cited and still clicked, which is a different situation.

Through the Search Console connector for Claude or ChatGPT the Performance side is one prompt:

Pull Search Console impressions and clicks per page for 20 June to 19 September for these pages: [paste the pages from the Generative AI export]. Add a column for total site impressions and clicks in the same period.

The Generative AI report is export-only for now, so paste its Pages table into the same chat and ask for the share per page.

Why. For windsor.ai, 20 June to 19 September 2026: 1.43 million impressions from AI features out of 11.6 million Search Console impressions, 12% of the total. Per page:

Page	Impressions	From AI features	Share	Clicks
Definition guide A	131,306	109,021	83%	49
Connector page A (social network)	133,864	70,793	53%	7,238
How-to guide A (ad platform)	63,859	32,231	50%	3,267
Definition guide B	153,021	74,728	49%	261
How-to guide B (spreadsheet)	119,114	51,837	44%	2,300
How-to guide C (social network)	94,747	41,584	44%	2,479
How-to guide D (spreadsheet)	57,762	24,582	43%	243
How-to guide E (ad platform)	70,934	28,525	40%	969

Definition guide A gets 83% of its impressions from AI features and 49 clicks in three months. Connector page A gets 53% from AI features and 7,238 clicks. The share tells you the page is in the answers. Whether people still click depends on the query type: on definitions they do not, on “how to connect” questions some do.

How to tell it worked. Every page with more than a few thousand impressions has a dated row with its AI share and its clicks, and the pages with a high share and low clicks are at the top of your list for step 2.

Step 2: pick the keywords from Search Console, per page

Do this. Pull the query table for the last 90 days, drop branded queries, and keep queries with at least 100 impressions. Group them by the page they land on, so each page has its own short list, starting with the pages from step 1. Add the queries with a click-through rate under 0.5% at a position between 1 and 10, even if they fall below the impressions floor.

Pull Search Console queries for the last 90 days with clicks, impressions, CTR and average position, non-branded only, at least 100 impressions, grouped by landing page for pages containing “[folder or term]”. Sort each page’s queries by impressions.

Why. Our list was 62 queries across 36 pages. Two pages carried five queries each, most carried one or two. That is enough to see a pattern per page and cheap enough to re-run monthly. Below 100 impressions the query list fills with one-off phrasings that never show the same Overview twice, so the check produces noise.

How to tell it worked. The list is short enough to check by hand or through a SERP API in one sitting, and every page you care about has at least one query on it.

Step 3: check each keyword on a live results page, with the delayed Overview loaded

Do this. For each query, load a live Google results page in the country and device that matter for the page and record three things: whether an Overview is present, every domain it cites, and the organic position of your page. If you check by hand, the Overview is on the page when it loads. If you check through a SERP API, a scraper or an LLM with a browsing tool, Google often returns the page first and the Overview a second or two later, and a tool that captures the page as first served reports “no Overview” for queries that have one. Set the tool’s option to wait for the asynchronous Overview, or have it request the page again after a pause.

Why. On our 62 queries, an Overview showed on 57, and 8 of those 57 loaded asynchronously. An automated check without the wait would have recorded those 8 as no Overview, and 6 of them cite windsor.ai. That is a fifth of our 30 citations missing from the count, on the queries where the Overview loads slowest. The same thing happened in July on a smaller set: one page looked uncited until the check was re-run with the wait on.

Each Overview cited a median of five domains. YouTube appeared in 48 of the 57 Overviews, Reddit in 13. On “how to connect” queries the Overview is usually a numbered set of steps with one link per step, so the domain that holds a step is cited and the domains with the same steps in prose are not.

How to tell it worked. Each query has a row with the Overview yes or no, every cited domain and your position, dated. Overviews change between checks: Ahrefs measured the Overview text changing 70% of the time and the citations 46% of the time between checks. Some disagreement between runs is normal. If two automated runs a week apart disagree on most queries, the tool is not waiting long enough.

Step 4: record the clicks and click-through rate per keyword before you change anything

Do this. For every keyword in the list, keep the last 90 days of clicks, impressions, click-through rate and position from Search Console next to the Overview check. This is the baseline. Without it, a later drop in clicks cannot be attributed to the Overview, a rank change or a fall in demand. If you know the date an Overview appeared on a query, compare click-through rate for the same number of weeks before and after it, at the same position.

Why. I audited 49 educational and glossary pages on windsor.ai over three months to the end of July. Almost every page showed a click-through rate between 0.05% and 0.4% at positions between 5 and 15. Definition guide A had 54,599 impressions and 40 clicks at position 5.1. A live check of 10 head terms for that page set found an Overview on 9. windsor.ai was cited in none of them. With the baseline in place, the pattern was clear: the Overview was answering the question and the click never happened. The Generative AI report later showed the size of it: 83% of that page’s impressions came from AI features.

How to tell it worked. Each keyword has a dated row of clicks, impressions, CTR and position that predates any edit you make.

Step 5: read citation next to organic rank

Do this. For the queries with an Overview, take your organic position from the live check, meaning the results under the Overview and not the Overview’s own slot, band it (1 to 3, 4 to 10, 11 to 30, not in the top 30) and count how often you are cited in each band. Read the two together. A good position is not a requirement for citation, and Google states none beyond being indexed and eligible for a snippet. A citation does improve the position Search Console reports, because every link in an Overview is assigned the Overview’s slot. Work on citation for the pages that are absent from the Overview, whatever their organic position.

Why. On our 57 queries with an Overview:

Our organic position	Queries	Cited in the Overview	Share
1 to 3	17	14	82%
4 to 10	15	12	80%
11 to 30	7	2	29%
Not in the top 30	18	2	11%

Four in five queries where we rank in the top 10 organically also cite us, and one in five of the rest. A single snapshot does not say which came first. The same page is often both ranked and cited because both draw on the same content, and the two citations from outside the top 30 are on a page that ranks for a neighbouring phrasing of the same question.

The educational pages tell the other half. Across 27 head terms checked in July, 25 showed an Overview and windsor.ai was cited in 5: the queries about our own product mechanics (the Meta Ads API, BigQuery pipelines, attribution models). On the generic definitions (“what is bigquery”, “etl vs elt”, “modern data stack”) the citation went to whichever page held a one-sentence definition or a plain comparison table near the top. On one “what is” query a course provider holds the definition slot and our page is cited on the bare term for a features list further down, because our page opened with a narrative paragraph.

How to tell it worked. The share cited in each band rises month over month, and the Search Console position on the cited queries moves toward the Overview’s slot.

Step 6: split every low-CTR query into a title problem or an answer problem

Do this. For each query with a good position and a click-through rate under 0.5%, look at the check from step 3. If there is no Overview, or there is one and it does not answer the query, the title and description are the problem and a rewrite will recover clicks. If the Overview answers the query fully, that phrasing has stopped sending clicks, and the work moves to being cited in it and to the follow-up questions the Overview does not answer.

Why. Earlier in the year our LinkedIn Pages guide had the same symptom, high impressions and a poor click-through rate at position 3, and a title rewrite fixed it, because the title did not match how people phrased the query and no Overview was involved. The 49 educational pages looked identical in Search Console and were a different problem: nine of ten head terms showed an Overview that answered the definition on the results page. Rewriting those titles would have done nothing. The fix there is a citable opening sentence and coverage of the questions under the head term, which is what the guide to getting cited by ChatGPT and Claude covers.

How to tell it worked. Title rewrites on the no-Overview queries recover click-through rate within a few weeks. On the Overview queries, the measure is citation, not clicks.

Step 7: re-check monthly and log gains and losses per keyword

Do this. Re-run step 3 on the same list every month, on the same country and device, and keep every run. For each keyword record three transitions: Overview appeared or disappeared, citation gained or lost, and organic position moved. When a citation disappears, open the Overview and read who replaced you.

Why. I checked Ahrefs and DataForSEO for earlier results pages. Both store them only for keywords already in their databases, monthly or at their own update dates, and for our list that meant a snapshot for 17 of the 62 queries, spread across June to August. That is why I now keep my own runs. For those 17, between the snapshot and 17 September:

• Two pages lost their citations, and the cause was the same in both. In July the Overview on one of our connector queries cited our page. In September it cites the platform’s own announcement of its native integration, its marketplace listing and its developer material, and our organic position went from 6 to outside the top 30. For a second connector, the Overview on every query cited our guide in June, July and August. In September none of them do. The AI vendor’s own help page for its native connector holds the first organic position and the citation, and our page moved from 3 to 12 on its best query. The Generative AI report still shows that page at 44% AI share, which is the June to August citations plus AI Mode impressions the report does not separate, so the live check is the only way to see the loss.
• Three queries gained a citation: two acquired an Overview that cites us, and one that had an Overview without us now includes us.
• Four pages held their citations and their positions.

The pattern in the losses is that an official integration launched and Google rebuilt the results page around it. A page refresh will not win those head terms back. The realistic target is the questions the official page does not answer, stated in the first paragraph so the Overview has a reason to quote a second source.

How to tell it worked. Your monthly log shows the same keyword under the same conditions across runs, and every lost citation has a named replacement next to it.

Step 8: measure the clicks you lose even when you are cited

Do this. For the pages cited in an Overview, put monthly clicks next to monthly impressions and position. A cited page that holds its position and loses clicks is losing to the Overview itself: people read the answer on the results page and do not click through, even though your page is one of the sources. Report that as answer loss, separately from rank loss and from falling demand.

Why. How-to guide A ranks 1 and 3 on its two head queries and is cited in the Overview on both. Its monthly clicks went from 1,916 in June to 815 in August. The group of pages it belongs to lost 52% of its clicks between the peak week in June and the end of August, while the whole site’s clicks were flat. Part of that is competing pages taking the query set and part is the Overview answering the how-to question in place. The citation keeps the brand in the answer, which is worth having, and it does not keep the click. Ahrefs’ study across 150,000 keywords with an Overview found 34.5% fewer clicks at position 1 year over year, the same pattern on a larger sample.

How to tell it worked. Monthly, three numbers per page: position on its head queries, cited yes or no, and clicks. If position and citation stay the same and clicks fall, the Overview took the clicks.

How to run the measurement

1. Export the Generative AI report, Pages table, last three months, and pull the Performance report totals for the same pages and dates through the Search Console connector for Claude or ChatGPT:
   

   For 20 June to 19 September, pull impressions and clicks per page for these pages: [paste]. Add site totals for the period. Then compute each page’s share of impressions from AI features using this table: [paste the Generative AI export].

2. Pull the query list for the same pages, last 90 days, non-branded, 100+ impressions, grouped by landing page:
   

   Pull Search Console queries and landing pages for the last 90 days for pages containing “[folder or term]”: clicks, impressions, CTR, average position, non-branded only, at least 100 impressions. Normalise trailing slashes and sum rows that collapse to the same page.

3. Add the low-CTR queries that fall under the floor:
   

   From the same data, list queries with at least 30 impressions, average position between 1 and 10, and CTR below 0.5%.

4. Check the SERP features per keyword. Ahrefs records, per keyword and country, which SERP features show and where your page ranks, and per page how many Google AI Overview answers the page appeared in. Through the Ahrefs connector the prompt is:
   

   For [domain] in [country], list the organic keywords from this list with search volume, our best position and the SERP features count, and for each landing page the number of Google AI Overview responses it appeared in.

5. Run the live check for the keywords with an Overview, on the country and device that matter. If a tool does it, make it wait for the asynchronous Overview. Record Overview present, every cited domain and your position, dated.
6. Join the tables on keyword and add the rank band.
7. Keep the baseline: clicks, impressions, CTR and position per keyword for the 90 days before your first run.
8. Classify each low-CTR query as title problem or answer problem.
9. Repeat monthly on the same list and log the transitions: Overview gained or lost, citation gained or lost, position moved, clicks per page.
10. Report five numbers: share of site impressions from AI features, share of tracked keywords with an Overview, share of those citing you, the citation rate in the top-10 rank band, and clicks per cited page against the baseline.

The tools behind this

• Google Search Console: the Generative AI report in the UI (export only), and the Performance report read through the Search Console connector for Claude or ChatGPT, which is how the query tables, baselines and per-page totals in this article were pulled. In Claude, add it from the directory (Customize, Connectors, search “Windsor”, Add). In ChatGPT, install the Windsor app from the apps directory. One source runs on the free forever plan, no credit card.
• Ahrefs, through the Ahrefs connector, for SERP features per keyword and country and the AI Overview count per page. Search Console plus Ahrefs is two sources, which is a paid plan after the first 30 days, from $19 a month for three sources.
• A live results page, or a SERP API that can wait for the delayed Overview, for the citation check itself. I use DataForSEO for it. No Google report gives it.
• GA4 for landing-page sessions on the cited pages, through the same connector list, to see whether traffic to a cited page changes. GA4 cannot separate Overview clicks from other Google clicks.

Conclusion

Search Console now counts how often each page is shown inside AI features, and says nothing about which query, who else is cited, or how many clicks that query still sends. Take the pages with a high AI share, pull their queries, check them on a live results page with the Overview fully loaded, and record who is cited and where you rank. Citation and position move together: four in five of our top-10 queries were cited, one in five of the rest, and a cited page reports the Overview’s position in Search Console. Keep your own monthly runs, because Ahrefs and DataForSEO hold past results pages only for the keywords already in their databases, and log each lost citation with the page that replaced it. Then measure the clicks you lose even on the pages that are cited, so the report separates rank loss, demand loss and answer loss.

The Search Console connector for Claude or ChatGPT covers the query tables, baselines and per-page totals, and the guide to getting cited by ChatGPT and Claude covers what to change on the pages that are absent from the answer. Free forever plan, no credit card.