If clicks fall while AI Overviews appear for your topics, first locate the change in comparable queries and pages. A similar ranking, a lower CTR or an answer resembling your article does not by itself establish that AI caused the loss. Use Search Console and documented search observations together before deciding what to change.
What a click study can—and cannot—tell you
AI Overviews summarize information and provide supporting links. They do not appear for every query, and readers may either continue to a source or complete part of their task in the search result. Neither a traditional featured snippet nor an AI Overview guarantees a particular click outcome.
Ahrefs' February 2026 study used 300,000 keywords: 150,000 with AI Overviews in December 2025 and 150,000 informational keywords without them. It compared aggregated desktop GSC CTR data from December 2023 and December 2025, using the non-AI group's change to estimate an expected CTR for the AI group. For position-one pages, the observed CTR was approximately 58% lower than that estimate.
This is an observational comparison under stated sample choices and assumptions, not a randomized experiment on every website. It does not mean your site necessarily loses 58 of every 100 historical clicks. Device, period, query mix and position matter; use the study to form a question, not to supply your site's diagnosis.
Keep three evidence sets separate
| Evidence | Useful question | Important limit |
|---|---|---|
| Overall Web performance | Which pages and queries changed in clicks, impressions and CTR? | It does not isolate every AI-caused click change |
| Generative AI performance report | Which pages received AI Search link impressions? | It is not a dedicated AI-click report or complete answer archive |
| Saved search observations | Did a feature appear, and which URLs supported the answer? | A limited sample does not represent all users' searches |
The current Generative AI performance report includes link impressions for AI Overviews and AI Mode. Google says it was rolled out worldwide on August 31, 2026. Its data is already included in overall Web performance, so do not add the two totals.
Preserve reporting conditions: dates use Pacific Time, recent data may be preliminary, and property-level chart totals can differ from page-level table totals. Exported zeros originating from “~” or “-” mean unavailable or nonnumeric values, not confirmed zero impressions. A missing report is not proof that a page is unindexed or that there is no demand.
Check whether the CTR decline is a change in query mix
Hypothetical calculation: in period one, branded queries receive 1,000 impressions and 100 clicks, while non-branded queries receive 1,000 impressions and 20 clicks. Combined CTR is 120/2,000, or 6%.
In period two, branded queries receive 500 impressions and 50 clicks, while non-branded queries receive 1,500 impressions and 30 clicks. Combined CTR becomes 80/2,000, or 4%. Each group's CTR remains unchanged at 10% and 2%; the total falls by two percentage points because the mix changes.
Real data needs the same care. Use complete, non-overlapping groups with consistent filters and divide total clicks by total impressions. Do not average row percentages, and do not treat a truncated query export as the whole account. The high-impression, low-click diagnostic is a related starting point for query-level investigation.
Position also needs context. Google's counting rules assign links in one AI Overview the position of the overview itself, not their order within its source list. AI Mode follow-up questions count as new queries. A combined average cannot describe every result type.
Rule out other explanations before rewriting
- Clicks and impressions both fall: investigate demand, seasonality, indexing, access problems and search-position changes.
- Comparable impressions remain while clicks fall: inspect titles, competing results and result-page features under recorded conditions.
- Search clicks remain but analytics visits fall: check measurement, consent handling, redirects and page loading.
- AI citations fall in a question sample: check whether the question set, platform mode, conditions or failure rate changed.
- A whole page type declines: examine its shared template, content, technical state and market conditions.
These are investigation paths, not single-metric verdicts. Google's traffic-drop guidance helps organize the comparison.
Also inspect the effective Search generative AI control, including parent-property inheritance. It controls participation in the applicable generative Search features, not AI training or inclusion and rankings in other Search areas. Understand why an existing setting was chosen before changing it, and allow for processing and propagation.
Improve the task beyond the summary
Do not hide the basic answer to force a visit. Make the destination useful for the work the reader still needs to complete: selecting a model, checking conditions, reproducing a calculation, comparing options or obtaining a real document.
For example, a short definition of a material may answer a general query. A purchasing decision may additionally require the exact formulation, temperature and media compatibility, a drawing and a source revision. Add those only when you can substantiate them. A longer generic definition does not fill that gap.
For a procedure, provide actual inputs, steps, expected results and failure handling. For a comparison, keep price periods, included items and limitations aligned. For a case study, provide the method and permitted evidence rather than inventing a customer result. Keep consultation, download and related-information links relevant and working.
Google's generative Search guide emphasizes useful, distinctive content rather than tiny text chunks or endless question variants. The AI Overview eligibility and content guide covers the technical and editorial checks in more detail.
Evaluate the change without inventing causation
Record the exact change, live date, URL and verification result, then compare the same evidence sets. If prices, advertising, forms and content changed together, do not attribute every later movement to the article.
A factual error can be corrected immediately even while the traffic cause remains unresolved. AI link impressions, clicks, contact actions, successful submissions and qualified inquiries are separate outcomes. The practical goal is a better answer and a useful next step, followed by evidence of what actually happened.