Google AI Overviews and your traffic: how to measure the impact

What Google says about clicks from AI Overviews and AI Mode

Google has made the same argument since launch: AI answers lead to more searches and to better clicks. Its main statements, in order:

• May 14, 2024. Announcing that AI Overviews were rolling out to everyone in the US, Google said the links included in an AI Overview get more clicks than the same page would get as a traditional web listing for that query, and that people who use AI Overviews search more.
• October 28, 2024. When AI Overviews expanded to more than 100 countries and territories, reaching more than a billion users a month, Google said its new link placements (a right-hand link display on desktop and in-line links in the text) had increased traffic to supporting websites in its testing, compared with earlier designs.
• May 21, 2025. The Search Central blog told site owners that clicks from results pages with AI Overviews are higher quality, meaning visitors are more likely to spend more time on the site, and suggested judging Search by conversions such as sales and sign-ups rather than by clicks alone.
• August 6, 2025. Liz Reid, head of Google Search, wrote that total organic click volume from Google Search to websites had been "relatively stable" year over year, and that Google was sending slightly more quality clicks, which it defines as clicks where the user does not quickly come back. She called third-party reports of steep declines inaccurate, often based on flawed methods or on changes that predate AI features. She also acknowledged that quick-answer searches may end without a click, as they already did with Knowledge Graph panels and sports scores, and that traffic is shifting from some sites to others.

Read these as statements about averages across the web. None gives a baseline, the periods compared or a figure for any one site, and a stable total is compatible with a large loss for a site whose traffic came from quick informational answers. Google's documentation adds the detail that matters most for measurement: AI Overviews appear only when its systems judge them additive to classic results, and often do not trigger. Which of your queries trigger them is the thing to find out.

How AI Overviews and AI Mode appear in Search Console

There is no separate click count for AI features. Google's documentation says sites that appear in AI Overviews and AI Mode are included in the overall search traffic in Search Console, reported in the Performance report under the "Web" search type. The counting rules follow the ones for other result types, with one detail that shapes every analysis below:

Two dates change what older data means. In August 2024 Google clarified that AI Overviews were already counted this way; it was a documentation change, not a reporting change. AI Mode data started counting towards the Performance report's totals in June 2025, according to Google's documentation updates dated June 16. Experiments in Search Labs are not included in Search Console at all.

The Generative AI performance report

On June 3, 2026, Google introduced a Generative AI performance report in Search Console, first for a subset of sites and, as of August 31, 2026, for all websites worldwide. It is the first Google report that isolates AI features. According to its help page, it shows impressions in AI Overviews and AI Mode by page, country, device and date (hourly to monthly), with a filter that separates text searches from multimodal searches made with an image, and the Performance report's usual limits, such as a 1,000-row table. Discover has a report of its own. An empty report means the rollout has not reached the property, the site has too few impressions in AI features, or it has been excluded from them with the control described at the end of this guide.

Each tool sees a different part of the picture:

What independent research has measured

Three kinds of study answer three different questions: how people behave on a results page, what happens to a publisher's traffic, and how click-through rate moves for pages that rank. The first two come from independent researchers. The third comes mostly from SEO vendors, whose data is useful but should be read as vendor data.

How searchers behave

Pew Research Center, a nonpartisan research organization, analyzed the browsing of 900 US adults on its KnowledgePanel Digital panel, who had installed an app that logs the pages they visit. It took every Google search they made in March 2025 (68,879 unique searches), re-ran each one with a scraping service in April to record whether an AI summary appeared, and classified what the person did next. The results, published on July 22, 2025:

• People clicked a traditional result link on 8% of visits to a results page with an AI summary, against 15% without one, and a link inside the summary on 1%.
• They ended their browsing session on 26% of pages with a summary, against 16% without.
• 18% of searches produced a summary, most often for questions: 60% of searches that began with who, what, when or why, against 8% of one- or two-word searches.

A correlation in a panel could reflect the kind of query rather than the summary. In a follow-up paper presented at IC2S2 2026, the Pew researchers fitted a mixed-effects model that controls for the panelist and for the query attributes that make a summary more likely, and the associations held.

Causal evidence comes from a preregistered field experiment by Stephanie Wang and colleagues, posted in August 2026. In March 2026, US Chrome users installed a browser extension that, after a three-day baseline, randomly assigned them for seven days to Google with its AI features hidden, Google unchanged, or every search sent to AI Mode; 1,100 of them searched during the experiment and form the sample. Hiding AI Overviews raised click-through to external sites by 8.8 percentage points; sending every search to AI Mode cut it by 18.8 points. The limits, which the authors state: a younger and more educated sample than the US population, only about half of AI Overviews actually hidden after Google changed their markup mid-study, and a forced AI Mode condition for people who had used it on 0.6% of searches. The paper is a preprint.

What happens to a publisher's traffic

Mehrzad Khosravi and Hema Yoganarasimhan used the staggered geographic rollout as a natural experiment. Their difference-in-differences design compares monthly search referrals to English Wikipedia articles (about 40% of their traffic comes from the US, where AI Overviews became the default in May 2024) with referrals to the same articles in German and French, using Wikimedia's public clickstream data for about half a million article pairs from December 2023 to December 2024. The estimate in their September 2026 revision: default AI Overview availability reduced English Wikipedia's search traffic by 5.45% relative to German and 4.82% relative to French. That is an effect on a whole publisher, averaged over every query that sends it traffic, which helps explain why it is far smaller than estimates for single queries.

Click-through studies from SEO vendors

Ahrefs, which sells SEO software, compared 150,000 keywords that showed an AI Overview with 150,000 informational keywords that did not, using average desktop click-through rates from aggregated Search Console data. Between March 2024 and March 2025, it estimated that an AI Overview lowered the top-ranking page's CTR by 34.5%. Its re-run, published in February 2026 and comparing December 2023 with December 2025, put the reduction at 58% for position one, shrinking to 19.4% at position ten.

Seer Interactive, an agency that also sells AI answer tracking, followed 3,119 informational queries across 42 client organizations from June 2024 to September 2025, using Search Console and Google Ads data. Organic CTR on queries that showed an AI Overview fell from 1.76% to 0.61%. Two of its findings matter more than the headline: CTR also fell by 41% on queries without an AI Overview, and the queries that showed one in September 2025 already had a lower CTR in June 2024. Seer's own reading is that AI Overviews appear on queries that already drew fewer clicks, so a plain before-and-after comparison overstates their effect.

The studies agree on direction and disagree on size because they measure different things: a whole publisher's traffic, clicks per search across all queries, and the CTR of a top-ranking page on the most exposed informational queries. Your site sits somewhere in that range, depending on how much of its traffic comes from quick informational answers. Two lessons carry over to your own analysis. Compare against a control group in the same period, as the experiment, the Wikipedia study and Ahrefs all do, because queries that attract AI answers were different to begin with. And treat absolute CTR from Search Console between May 13, 2025 and April 27, 2026 with care, for the reason in the next section.

Zero-click searches, and why impressions rise while clicks fall

A zero-click search is one that ends without a click on any result: the searcher found the answer on the page, refined the query or gave up. It predates AI Overviews. In a 2024 study based on the multi-million-device clickstream panel of Datos (a Semrush company), SparkToro found that 58.5% of US and 59.7% of EU Google searches in the first five months of 2024 ended without a click, before AI Overviews reached most countries. Pew saw the same pattern in 2025: about two-thirds of searches ended with the person moving elsewhere on Google or leaving without clicking a result, with or without an AI summary.

AI features push more searches over that line, and they also change the numbers Search Console reports, which is how a site can see impressions rise while clicks fall. The first mechanism is position. The Performance report records the topmost position any link to your site held, and every link inside an AI Overview gets the AI Overview's position, which is often the top of the page. A page that ranks ninth but is cited in the AI Overview is reported near position one, with an impression whether or not anyone clicks. Average position improves, impressions grow, and CTR at "position one" drops.

Two more effects add to it. Google says AI Overviews put more links on the page than before and that people search more and ask longer questions. Both features may use query fan-out, issuing several related searches across subtopics to build one answer, which lets them link to a wider set of pages than a classic results page. In AI Mode, each follow-up question is a new query with its own impressions. More links and more queries mean more impressions for more sites, each with a small chance of a click.

Before blaming AI features, match the pattern in your Performance report to its likely causes:

Build a baseline before you compare anything

Search Console keeps 16 months of Performance data; Google's traffic drop guide points to the Search Analytics API or bulk exports if you want to keep more. As of September 2026 that window starts in late May 2025: after AI Overviews had reached more than 100 countries, and inside the logging error. Unless you exported data earlier, you cannot build a clean before-and-after from the report alone. That is not fatal. The more careful studies above compare groups within the same period, and you can do the same with your own data.

1. Start the bulk data export to BigQuery if the site has enough traffic to justify it. It has no history: the first export arrives up to 48 hours after setup and contains that day's data. For older dates, use the API or the reports.
2. Fix the scope. Use the Web search type, look at one country at a time where AI features launched on different dates, and keep mobile and desktop apart.
3. Classify pages as well as queries. The query rows leave out anonymized queries, rare searches that Google withholds to protect privacy, and the report's table shows at most 1,000 rows. Grouping by page (guides and articles against service and product pages) keeps all the clicks in the analysis.
4. Export the Generative AI performance report for the same dates, so you know which pages appear in AI features and which do not.

Split queries into informational and transactional

AI features concentrate on informational queries, so a site-wide average hides their effect. Pew found summaries on 60% of question-word searches. A 2026 measurement study that issued 55,393 trending queries between March 13 and April 21, 2026 saw AI Overviews on 13.7% of them overall and on 64.7% of question-form queries, and Ahrefs reports that 99.2% of the keywords in its database that trigger an AI Overview are informational. Queries with commercial or transactional intent, such as a price, a quote or a service in a city, make the natural comparison group, and branded queries a third.

Use the branded queries filter for the first class rather than a hand-written list; its classifier also catches misspellings and product names. For the other two, the Performance report's query filter accepts a custom regex in RE2 syntax, which is not case-sensitive by default. A starting point for informational searches:

^(who|what|when|where|why|how|which|is|are|can|does|do|should)\b|\b(meaning|definition|example|guide|tutorial)\b

And for commercial and transactional ones:

\b(price|pricing|cost|quote|hire|buy|near me|company|agency|services?)\b

Adapt the word lists to your market and language, and read a sample of the matches before trusting the split. A query such as "how much does a website cost" matches both patterns; decide which class it belongs to and keep that rule for every period.

Compare CTR by position over matching date ranges

CTR depends heavily on position, so compare like with like: the same position band, query class, country and device, over date ranges of the same length and season. In the Performance report, the Compare tab of the date filter sets up the two ranges (Google's traffic drop guide uses "Compare last 3 months year over year" as its example), and the weekly or monthly view added in December 2025 removes the weekday mismatches that make daily comparisons hard to read.

1. Pick two equal periods in the same season. Year over year is best. Where your history is too short, compare two recent periods and rely on the class comparison in step 4.
2. Filter to one class at a time and note clicks, impressions, CTR and average position for each period.
3. Group the results into position bands: 1, 2 to 3, 4 to 10, and 11 or lower. Export the query table and band it in a spreadsheet, or use the API or the bulk export to get past 1,000 rows.
4. Compare the change for informational queries with the change for transactional queries in the same band and period. The difference between the two changes, not either change alone, is your best estimate of what AI features cost you.
5. Repeat at page level: pages that appear in the Generative AI report against comparable pages that do not.

With the bulk export, one query does steps 2 and 3 for every month. The fields come from Google's table reference: sum_top_position is zero-based, so average position is sum_top_position / impressions + 1.

WITH q AS (
  SELECT
    DATE_TRUNC(data_date, MONTH) AS month,
    country,
    device,
    query,
    clicks,
    impressions,
    sum_top_position / impressions + 1 AS position
  FROM `your_project.searchconsole.searchdata_site_impression`
  WHERE search_type = 'web'
    AND NOT is_anonymized_query
    AND impressions > 0
)
SELECT
  month,
  country,
  device,
  CASE
    WHEN REGEXP_CONTAINS(query, r'(?i)\b(price|pricing|cost|quote|hire|buy|near me)\b') THEN 'transactional'
    WHEN REGEXP_CONTAINS(query, r'(?i)^(who|what|when|where|why|how|which|is|are|can|does|do|should)\b') THEN 'informational'
    ELSE 'other'
  END AS query_class,
  CASE
    WHEN position < 1.5 THEN '1'
    WHEN position < 3.5 THEN '2-3'
    WHEN position < 10.5 THEN '4-10'
    ELSE '11+'
  END AS position_band,
  SUM(clicks) AS clicks,
  SUM(impressions) AS impressions,
  SAFE_DIVIDE(SUM(clicks), SUM(impressions)) AS ctr
FROM q
GROUP BY month, country, device, query_class, position_band
ORDER BY month, country, device, query_class, position_band;

Two cautions when you read the output. Position one is the band AI Overviews distort most, because a citation inside an AI Overview reports the AI Overview's position; read it next to the Generative AI report for the same pages. And if your periods straddle May 13, 2025 or April 27, 2026, compare classes with each other rather than a class with its own past. An error that inflates impressions for all queries moves every class's CTR in the same direction, so a comparison between classes is far less exposed to it than a before-and-after within one class.

Annotate the rollout dates

A change in your chart means little until you know what happened that week. Search Console's custom annotations, added in November 2025, let you right-click the Performance chart and add a dated note of up to 120 characters. Everyone with access to the property sees them, so keep personal details out. Add the dates that apply to your markets, then your own releases, migrations and content changes, and the core updates on Google's list of ranking updates.

AI features reached other countries and languages on other dates, announced on Google's Keyword blog. If your market sits outside the table, look up its date and use that instead.

Read branded search trends

Branded searches show whether demand for you is holding while the research phase moves into AI answers. Search Console's branded queries filter, announced in November 2025 and available to all eligible sites since March 11, 2026, splits the Performance report into branded and non-branded queries with an AI-assisted classifier. The Insights report adds a card comparing branded and non-branded clicks. The filter works only on top-level properties (not URL-path or subdomain properties) with enough queries and impressions, and Google warns that some queries may be misclassified.

Outside your own data, Google Trends shows interest in your brand name over time, along with the rising and top searches made with it; Google suggests checking them over the past 30 or 90 days. Read the two sources together:

• Non-branded informational clicks down, branded clicks and brand interest steady or rising. People are doing their research in AI answers and still come to you when they decide. Some may have met your name in an AI answer first, which shows up as branded search rather than as a referral.
• Branded clicks and brand interest falling too. The cause is broader than AI features: demand, reputation or competition. Treat it as a traffic drop and work through Google's guide.

Track referrals from AI assistants in analytics

Chat assistants are a separate channel from Google's AI features, and Google Analytics now treats them that way. Since May 13, 2026, GA4 assigns the medium ai-assistant and the campaign (ai-assistant) when a visit's referrer matches its list of AI assistants, and reports those sessions in an AI Assistant channel; Google names ChatGPT, Gemini, Claude, Copilot, DeepSeek and Grok as examples. Visits from AI Overviews and AI Mode are not in it: GA4 counts them as Organic Search, mixed with every other visit from Google Search.

For dates before May 2026, or for assistants the default list misses, build a custom channel group: add a channel called AI assistants with a "matches regex" condition on the source, and move it above Referral so it is evaluated first. Custom channel groups apply to reports retroactively. Google's help includes an example pattern; a stricter one that names the assistants is:

chatgpt|openai|perplexity|gemini\.google|copilot\.microsoft|claude\.ai|deepseek|grok

Check the source values your property actually records before relying on any pattern, and expect some undercounting: a visit lands in this channel only when the assistant passes a referrer GA4 can recognize. Then compare the channel's engagement rate and key events with Organic Search, not only its sessions. Google's guide to Search Console and Analytics data treats sessions as the volume of traffic and engagement rate as its quality, and recommends reading them together.

Two more sources fill gaps that analytics cannot see:

• Server logs. OpenAI documents that ChatGPT may visit a page with the ChatGPT-User agent when a user asks it a question, separately from its OAI-SearchBot crawler. Requests from ChatGPT-User show your pages being read on a user's behalf, whether or not that user then clicks through.
• Bing Webmaster Tools. Microsoft's AI Performance report, in public preview since February 10, 2026, shows how often your pages are cited in Microsoft Copilot, Bing's AI-generated summaries and some partner integrations, with the grounding queries used to retrieve them. It counts citations, not clicks.

What to do with the result

Once the numbers are in, most sites land in one of a few situations:

Then change what you report. A single site-wide CTR will keep falling for reasons that are not all bad news. A more honest monthly view has four parts: clicks and key events by query class and page type, AI feature impressions by page, AI Assistant sessions with their engagement, and branded demand. The search demand that remains is still worth winning, and finding untapped opportunities in your Search Console data covers where it usually hides.

Search Console also offers a way out. The Search generative AI control, available to all sites since August 31, 2026 under Settings, excludes your site's links and content from AI Overviews, AI Mode and generative AI features in Discover. Google says the change generally takes effect within a few days, is not a ranking or inclusion signal for the rest of Search, and does not affect AI training, which Google-Extended controls. The cost is plain: an excluded site gets no impressions or traffic from those features, while other sites' content may still answer the question. Estimate what you would give up in the Generative AI performance report first, as Google suggests, and treat it as a business decision rather than an SEO fix.

If you want this measurement set up and run for you, our SEO service covers measurement and AI visibility: Search Console and GA4 set up properly, citations in AI answers tracked, and reporting in Looker Studio or BigQuery, read against a baseline with dates and sources.