How AI Overviews Are Changing Google Search: What the Research Shows

AI Overviews are changing the relationship between being found on Google and getting a visit to your website. Several studies find lower click-through when an AI answer appears. However, the size of the effect varies, newer findings are not uniformly negative, and a fall in clicks is not automatically the same as a fall in enquiries.

For a small business, the useful question is how that change affects the searches your customers make, the pages that bring them to you and the work those visits generate.

Research reviewed: 12 October 2026. This article reviews published studies and official documentation; it is not an original experiment. Findings are dated because Google’s interface and reporting continue to change.

What are AI Overviews, and how are they different from AI Mode?

An AI Overview is an AI-generated summary within Google’s search results, with links to sources or further reading. Google says it appears when its systems consider it helpful; it does not appear for every search. Google also warns that these answers can contain mistakes. See Google’s explanation of AI Overviews.

AI Mode is a more conversational search experience, designed for follow-up questions and more complex exploration. It is a different interface, not simply another name for an AI Overview. It is also distinct from the standalone Gemini app. See Google’s AI Mode guidance.

This distinction matters. A study that forces people into AI Mode cannot be treated as a measurement of the effect of a summary above ordinary results.

What the research says about clicks

The studies below measure different populations, time periods and outcomes. They should be read together, not averaged into one “AI traffic loss” figure.

On smaller screens, swipe the table to see all columns.

The original studies and their methods are linked below. CTR means click-through rate: the proportion of measured impressions or search visits that produce a click, depending on the study’s definition.

Pew: people clicked traditional results less often

Pew Research Center’s July 2025 study examined 68,879 unique Google searches from 900 US adults during March 2025. Traditional results received a click on 8% of visits with an AI summary, compared with 15% without. Links within the summary received clicks on 1% of visits where a summary appeared.

That is a seven-percentage-point gap in traditional-result clicking. It is not evidence that every website lost seven per cent of its visitors.

There are meaningful limitations. The searches were observed in March, but their results were reconstructed in April. AI and non-AI searches could also differ in intent and complexity. The study offers valuable behavioural evidence, but does not isolate an AI Overview’s causal effect on an otherwise identical search.

Ahrefs: a substantial top-position CTR gap

Ahrefs’ February 2026 analysis compared desktop Search Console data for 300,000 keywords: half with AI Overviews in December 2025 and half informational keywords without them.

Between December 2023 and December 2025, first-position CTR in the AI Overview group fell from 7.3% to 1.6%. CTR also fell in its comparison group, from 7.6% to 3.9%. Ahrefs used that comparison-group decline to estimate an expected CTR of roughly 3.7% for the AI group. The observed 1.6% was about 58% lower.

The headline is therefore a modelled shortfall against a comparison, not a finding that Google took away 58% of all organic traffic. The adjustment is useful, but cannot control for every difference between keyword groups or every change over two years.

Seer: the decline was not a straight line

Seer Interactive’s April 2026 update covered 53 brands and 5.47 million tracked queries. Organic CTR for queries labelled as having AI Overviews rose from 1.3% in December 2025 to 2.4% in February 2026, after declining during 2025.

Its findings also associated AI citations with better CTR than being uncited on AI-affected results. That does not establish that securing a citation causes the improvement: ranking, brand recognition and query mix could help explain the difference.

The dataset required complete monthly paid, organic and search-results data. AI Overview status came from a recent snapshot applied retrospectively, with no seasonality adjustment. March 2026 figures in the report were projections; February was the last month of actual data.

This is a useful warning against projecting a decline indefinitely from a short window.

SISTRIX: affected-query losses and market losses differ

SISTRIX’s German analysis, published in February 2026 and updated in April, covered more than 100 million keywords. It estimated a 6.6% loss in organic clicks across the full keyword market, with wide differences between sectors.

That much smaller market-wide figure is not necessarily inconsistent with a large CTR reduction on specific affected searches. Not every query displays an AI Overview, and different searches contribute different amounts of traffic.

These are modelled German-market estimates, not directly observed losses for every business. Applying the percentage to a Sussex company’s revenue would be unjustified.

A field experiment adds stronger causal evidence

An August 2026 preprint by Stephanie T. Wang and colleagues reports a randomised Google Search experiment with 1,100 US participants. After a baseline period, users spent seven days with normal search, a version intended to hide search AI, or enforced AI Mode.

The compliance-adjusted estimate for removing search AI was an 8.8-percentage-point increase in publisher CTR, with a 95% confidence interval of 2.3 to 15.3 points. This “No AI” treatment covered AI Overviews and redirected AI Mode, so it is not a pure test of one feature.

A Google interface change weakened the intervention: only 51.1% of AI Overviews were successfully hidden overall. The adjusted estimate therefore depends on additional assumptions. The sample skewed younger and more educated, and the experiment was short.

Randomisation strengthens the evidence, but this remains a preprint rather than an established peer-reviewed conclusion. Its separate enforced-AI-Mode results should not be relabelled as AI Overview effects.

How does this fit with Google’s claim that clicks are stable?

In an August 2025 statement, Google said total organic clicks to websites were relatively stable year on year and that average click quality had improved. It described a quality click as one where someone does not quickly return to the results.

This is Google’s account, dated August 2025, rather than a new independently audited 2026 finding. The post did not provide a reproducible dataset for checking the aggregate claim.

Both stable total clicks and declining CTR can be true. As a purely illustrative example, 1,000 impressions at 10% CTR generate 100 clicks; 2,000 impressions at 5% also generate 100 clicks. That arithmetic is not a claim about Google’s actual search volumes.

Aggregate stability can also hide redistribution between websites. And a visitor staying longer is not automatically a qualified enquiry, purchase or profitable customer. Businesses need to measure those outcomes separately.

Why older research needs a date attached

Google has continued changing how AI answers display and link to sources. Its 2026 announcement about website-owner controls, updated in August, describes expanded inline links and other ways to surface websites. Google also reported more than 2.5 billion monthly AI Overview users.

That is a company-reported audience figure, not the proportion of searches containing an Overview. The interface changes also mean a study of March 2025 browsing cannot be assumed to describe every October 2026 search experience.

The sensible approach is to retain older evidence as a baseline while checking whether newer measurements confirm, weaken or complicate it.

What could this mean for a small business?

My interpretation is that businesses should separate three parts of the customer journey: learning about a problem, choosing a provider and taking action. A search feature can affect each differently.

Informational visibility may not produce the same visits

Consider the hypothetical search “what causes damp on an internal wall?” Someone might be satisfied by an explanation without visiting a contractor’s website. By contrast, “damp survey in Hove” suggests a different need: identifying a suitable provider and arranging work.

These examples illustrate intent, not checked live results or a claim that local searches are protected from AI. A useful review would look at the actual terms bringing customers to that business.

I would therefore question a content plan that counts every educational visit as equally valuable. Helpful guides can still demonstrate expertise, but their role should be explicit: attracting suitable enquiries, supporting a buying decision or answering questions existing customers ask.

A citation is visibility, not a guaranteed customer

Being named in an answer may expose someone to your business. It does not demonstrate that they read your page, trusted the recommendation or contacted you. Equally, a person could discover a name and search for it later.

That makes attribution harder. Treat citations as one diagnostic measure, alongside branded searches, relevant visits and recorded enquiries. Do not price their value as if every mention were a sale.

Publishers and service businesses have different economics

A publisher funded by page views may experience an immediate cost when readers obtain an answer without visiting. A service business can sometimes tolerate fewer visits if enough suitable customers still enquire. These are different commercial models, so the same traffic percentage can have very different consequences.

My concern is the incentive to invest in useful original material if its audience increasingly stays elsewhere. That is a business-model risk, not proof that every publisher will fail or that useful web content will disappear.

How to measure your own exposure

Start by asking whether there is a business problem, then investigate the search behaviour behind it. Avoid declaring an AI crisis from a single traffic chart.

Use the current Search Console reports

Google’s generative AI performance report documentation says its worldwide rollout completed on 31 August 2026. It reports impressions from AI Overviews and AI Mode, with breakdowns including pages, countries, dates and devices. A site may not see the report until it has enough impressions.

The documented report measures visibility, not a separate AI Overview conversion rate. Its data is also included in the main Web search performance data, so adding the two totals would double-count impressions. Dividing all organic clicks by AI impressions would not produce a valid AI Overview CTR.

Use the main search performance report to investigate clicks and query groups alongside this additional visibility information.

Compare like with like

• Separate branded and non-branded searches. People already looking for your business have a different relationship with you.
• Group pages by purpose. Compare advice articles, service pages and commercial enquiries separately.
• Keep country and device consistent. A change in the mix can distort an average.
• Check impressions, clicks, position and enquiries together. A lower CTR with more impressions does not necessarily mean fewer visits.
• Record other changes. New pages, redirects, indexing problems, advertising activity and seasonal demand are alternative explanations worth checking.

For a small business, this can start with Search Console exports and a simple enquiry log. Record the service requested, whether the enquiry was suitable and whether it became paid work. Ask customers how they found you, recognising that recollection is imperfect.

Aim for a consistent comparison over several weeks and, where available, the equivalent period last year. A single before-and-after snapshot is useful for forming a hypothesis; it is rarely enough to establish a cause.

What should businesses change?

Google’s current AI optimisation guidance keeps familiar SEO foundations central: useful original content, technical accessibility and a good user experience. It does not require a special AI text file or special schema to appear in AI features. Structured data can support relevant search features, but is not a guarantee of inclusion.

My practical response would be to improve the evidence available about the business and the experience of choosing it:

1. Make service pages genuinely useful. Explain who the service suits, where you work, what happens next and how customers can assess whether you are a good fit.
2. Show work you can substantiate. Use permissioned case studies, relevant experience, original examples and clear explanations. Distinguish measured results from opinions.
3. Make information consistent. Check business details, service descriptions and key facts across your website and the profiles you maintain.
4. Improve the enquiry journey. A relevant visitor should be able to understand the offer and contact you easily on a phone.
5. Review the marketing mix. Decide whether search, referrals, email, social activity or paid campaigns deserve the next portion of the budget based on your own results.

There is no need to launch every channel at once. I would agree a manageable monthly plan, work on the highest-priority problem and review what changed. Buying a package labelled GEO or AEO is not, by itself, evidence that the work will help.

Our free website SEO health check can help identify basic gaps. For a closer review of discovery and visibility, see AI visibility audits and SEO and local search support.

Can website owners opt out?

Google now documents a Search generative AI inclusion control. Excluding a site removes its content and links from the covered generative search features; Google says this setting is not used as a ranking or inclusion signal outside those features.

This is distinct from removing a website from ordinary search with noindex, and from controls concerning model training. Other sites’ material may still appear in AI answers, including information about the same topics.

My view is that opting out needs a commercial assessment. It gives up potential citation visibility and does not, by itself, restore clicks displaced by an answer based on other sources. Publishers and local service businesses may reasonably reach different decisions.

What we can conclude, and what remains uncertain

The evidence reviewed here supports taking reduced outbound clicking seriously. It does not support applying one headline percentage to every website, assuming local services are immune, or declaring that SEO has stopped mattering.

Important questions remain: how behaviour changes with newer interfaces, whether citations lead to later enquiries, which businesses gain from redistributed attention, and how durable the effects will be. The studies cited are not a representative forecast for small businesses across Sussex.

My advice is to keep the business objective clear. Understand where customers discover you, give them convincing reasons to choose you and measure useful enquiries rather than celebrating visibility in isolation.

If you would like to understand what this means for your website, tell me about your business and the changes you are seeing. We can agree a practical review around your priorities and budget.

Editorial note: this is a selected evidence review, not a systematic review or meta-analysis. It prioritises original research, published methodology and official product documentation. Commercial studies and Google’s own statements are attributed as such; practical recommendations are my interpretation. No universal traffic or revenue forecast is inferred from these sources.