A sharp traffic decline is easy to notice and hard to explain. Artificial intelligence (AI) summaries now appear directly in Google Search through the AI Overviews feature.[5] A chart can show visits fell after launch. It cannot, by itself, prove the feature caused the fall.
That distinction matters for every publisher, documentation site, and small business that depends on search. If the cause is a ranking change, the response may be to improve a page. If demand for a topic fell, rewriting the page may not help. If users see an AI answer and stop before clicking, the business problem is different again.
Nicolas Bouliane, who runs the Berlin guide All About Berlin, reported that the site had lost 75% of its traffic to Google AI Overviews. The post includes a long-term traffic chart. It does not share the analytics export or a method that isolates AI Overviews from other causes. It also gives no detailed measurement window. The 75% figure is the author's account, not an independently audited estimate of the feature's effect on the site.[1]
There is also experimental research on the broader question. A 2026 working paper by Saharsh Agarwal and Ananya Sen was posted to the Social Science Research Network (SSRN) on April 3 and last revised July 8. It reports that AI Overviews reduced outbound organic clicks in its randomized browser experiment. That result is evidence about the study participants and searches. It is not a forecast that every website will lose the same share of its traffic.[2]
This article lays out a practical way to investigate a decline. It separates what the data can show from what still needs testing.
First, separate the numbers people mix together
"Traffic fell" can describe several different changes. Before looking for a cause, name the metric and the scope.
- Search impressions count how often a site appears in Google Search.
- Search clicks count visits that start with a click from Google Search.
- Click-through rate (CTR) is clicks divided by impressions. A lower CTR can mean the same visibility produced fewer visits.
- Analytics sessions measure visits recorded by the site's analytics system. They are not the same unit as Search Console clicks.
- Revenue and sign-ups measure outcomes after a visitor arrives. They may move differently from visits.
Google explains that Search Console covers activity before a visitor reaches a site, while Google Analytics covers activity on the site. Clicks and sessions will not match exactly because they use different systems and definitions.[3] Use one tool to understand search visibility and another to understand what visitors do after arrival. Do not treat their totals as interchangeable.
A fall in total sessions is not the same as a fall in Google clicks. A fall in clicks is not necessarily a fall in search visibility. And a fall in visits does not automatically mean revenue fell by the same percentage.
What the 75% report does and does not establish
Bouliane's post is useful as a first-person report from a site owner. It includes a long-term traffic chart and describes Google results that place ads and an AI-generated answer before the site's link for some searches.[1] Those statements describe his experience and his explanation of the chart.
But the public post does not show enough detail to estimate the feature's independent effect on this site. It gives no query-by-query comparison or control group (a set used for comparison). It also lacks matching figures for clicks and impressions. It gives no breakdown of rankings or visits for the same dates. A screenshot can show timing and scale. It cannot rule out a ranking update or a drop in demand. Seasonality, a technical problem, or another site change could also explain the decline.
That is not a reason to dismiss the report. It is a reason to label it correctly: an owner-reported decline with a proposed explanation. Independent validation would require more data than the post makes public.
The distinction becomes clearer when compared with the Agarwal and Sen working paper. The researchers randomly assigned Google users to a normal-search group or a group where a browser add-on hid AI Overviews. Conditional on an AI Overview appearing, they report 39.8% fewer clicks from unpaid results to outside websites. Searches that ended without an outbound click increased by 34.5%.[2] The authors also report no change in ad clicks or overall search frequency. They found no measurable improvement in perceived search quality or ease of finding information.
Random assignment gives this result stronger cause-and-effect support than a before-and-after chart. Still, its scope is limited. The study measured search behavior during its experiment. It did not measure traffic or revenue for every publisher. The 39.8% figure describes clicks from unpaid results under the study's conditions. It does not predict that a given site's monthly visits will fall by 39.8%.
Keep the two findings separate. One is a specific site's owner-reported traffic decline. The other is an experimental estimate of user click behavior. They point to a plausible mechanism, but one does not verify the exact cause or size of the other's result.
A search-traffic investigation you can repeat
Google's traffic-drop guide suggests a long date range and comparisons between periods. It also recommends filtering web or image searches and reviewing changes by page or query. Google Trends can help check for wider demand changes.[4] Use that sequence to structure the investigation.
1. Confirm the decline in more than one system
Choose one stable measure first. For Google search, begin with Search Console clicks. For visits and business outcomes, check analytics separately. Mark the first week when each measure changed and note whether both moved at the same time.
Check for measurement changes too. A consent-banner or analytics-tag change can alter recorded sessions. A domain move or reporting filter can do the same, even if Google clicks stay steady. Compare Search Console clicks to analytics sessions from unpaid Google traffic, but look for a shared pattern rather than identical totals.[3]
2. Compare with the same period last year
Search demand can rise and fall at the same time each year. A month-over-month comparison may mistake that pattern for a platform effect. Compare the affected period with the same weeks in the previous year. Then widen the view to cover at least a full annual cycle when the data is available.
For seasonal topics like travel or education, compare the same weeks and relevant holidays. Google recommends looking at a longer history and using Google Trends to see whether interest in the affected queries changed across the wider search market.[4]
3. Break the change down by page and query
In Search Console, compare clicks and impressions by page and query. Start with pages that lost many clicks. For each one, check:
- Did impressions fall along with clicks?
- Did impressions stay similar while CTR fell?
- Did average position move sharply?
- Did the change affect one query, a group of related queries, or many unrelated topics?
- Did the same pattern appear on desktop and mobile, or in one country only?
These patterns help narrow the next question. Falling impressions with a position decline may point toward visibility or ranking changes. A stable impression count with a lower CTR means the page still appeared but drew fewer clicks. Several causes can fit that pattern. The results layout or title may have changed. New competitors or an AI answer may also affect clicks. The pattern alone does not identify which cause matters.
The average position metric combines ranking data across many searches. It can move because the mix of queries changed, even if rankings for each query did not. Inspect the important queries and pages instead of relying on one site-wide average.[4]
4. Check demand and site changes before assigning cause
For affected queries, compare demand in Google Trends. If interest dropped across the market, the site may have lost some visits because fewer people searched. Review technical changes. Check site releases and redirects. Inspect your robots.txt file, which sets rules for web crawlers. Check canonical URLs too. A canonical URL is the address you want Google to treat as the main version of a page. Also review page speed and indexing reports. Then check whether a Google ranking update or competitor change overlaps the same dates.
Keep a dated event log. Record when the decline began. Add site changes, Google updates, and major content changes as they happen. This makes it easier to compare possible causes instead of choosing the obvious explanation after the fact.
5. Track AI visibility and its limits
As of August 31, 2026, Google says its generative AI performance reports in Search Console are available to all websites worldwide. The reports show impressions for pages appearing in AI Overviews and AI Mode. They can be grouped by country, device, and date. Google says these impressions are also included in the overall performance report.[5]
Google's August 31 note also says AI Mode data counts in the overall Search Console Performance totals. The dedicated AI report provides impression-based visibility, but its listed fields do not include a separate click count for each AI feature. A rise in AI-feature impressions alongside a fall in search clicks is an important signal, but it is not a complete accounting of clicks lost to the answer.
Pair the AI visibility report with the standard Search Console report and your analytics. Ask whether the same pages and queries are losing clicks, whether ordinary search impressions or positions changed, and whether the timing lines up. Keep a note of the exact date range and filters so another person can reproduce the comparison.
If you build your own reporting, Google's November 16 Search Analytics application programming interface (API) update adds filters and groupings for multimodal web search, searches that use images. Google says this includes activity from Lens, Circle to Search, and image uploads. A site with image-driven discovery can use that data to check whether its search mix changed, rather than treating every change as a text-query effect.[6]
Stronger evidence needs a comparison
A publisher usually cannot randomize which users see an AI Overview. But it can avoid drawing a conclusion from a single line chart.
Compare affected and less-affected pages. Group similar pages by topic, search intent (what a searcher wants to do), and past traffic. If one group appears more often in AI features, compare its change with a similar group that appears less often. Match the groups as closely as possible. A page that answers a simple fact is not a fair control for a page that helps someone complete a complex process.
Use a before-and-after comparison with a control group. Look at both groups before and after the change. If the more-exposed group falls more than the comparison group, that supports the hypothesis. It still does not prove it if another change also affected that group.
Check whether the result survives alternate date windows. Repeat the analysis with different start and end dates. A result that appears only after choosing one convenient week is less convincing than one that holds across reasonable windows.
Look at outcomes beyond page views. Track potential customers who match your business needs. Track subscriptions, downloads, and other goals too. Search changes can alter the number and type of people who arrive. A smaller audience may still complete more important goals, or it may not. Measure rather than assume.
For many independent publishers, this method will not produce a controlled experiment. It can still give a likely cause and a clear confidence level. That is more useful than claiming AI caused every decline or assuming it had no effect.
Turn the finding into a decision
If impressions and rankings fell across many query groups, investigate search visibility and technical health. If impressions stayed stable but CTR fell on queries where AI features appear, test titles and page presentation, then compare those results with the AI visibility report. If searches themselves declined across the market, update demand expectations. If search visits fell but referral traffic grew, include it in the review. Check direct visits and email traffic too.
Do not rewrite every page simply to make it more attractive to an answer engine. First identify which pages lost clicks and what role they play. A reference page that answers a short factual query may face a different risk from an interactive service. That service could be a calculator or a community. It could be a maintained database or a step-by-step tool that people need to visit.
Help readers return directly. Email subscriptions can create that path. Useful tools and saved resources can too. Clear author or organization identity can help build a relationship beyond a search result. These steps do not restore a lost Google click. They reduce dependence on any one route to an audience.
Measure the change you can see
AI answers can change how people use search, and randomized research now offers evidence that they can reduce outbound clicks in specific conditions. But a site's traffic chart still needs careful reading. It cannot isolate a cause without comparisons, and an industry study cannot supply an exact forecast for one publisher.
For a practical diagnosis, combine three views. Use Search Console for impressions and clicks. Use analytics for visits and outcomes. Use the AI performance report to check visibility in Google's generative features. Then check seasonality and query demand. Review rankings and site changes. Compare the result with a reasonable control group.
A credible conclusion may be narrower than a headline. It will also give a publisher a better basis for deciding what to fix, what to measure next, and how much to invest in channels it controls.
Sources
[1] AI is killing All About Berlin. Nicolas Bouliane, May 26, 2026.
[2] The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment. Saharsh Agarwal and Ananya Sen, SSRN working paper.
[3] Using Search Console and Google Analytics Data for SEO. Google Search Central.
[4] Debugging drops in Google Search traffic. Google Search Central.
[5] Google's Search Generative AI performance reports. Google Search Central, updated August 31, 2026.
[6] New filtering and reporting capabilities in the Search Analytics API. Google Search Central, November 16, 2026.
Top comments (0)