What AI Overviews Actually Mean for Your Traffic in 2026

AI Overviews reduce clicks to websites. That much is settled, and every credible study measuring it has found a decline. What is not settled is the size of the effect, and the published figures vary so widely that quoting any single one of them tells you more about which study someone read than about what is happening to your website. The honest position in July 2026 is that the direction is certain, the magnitude is contested, and almost none of the underlying research is South African.

That nuance matters commercially, because the strategic response to a 15 percent decline in informational traffic is entirely different from the response to a 60 percent collapse, and both figures are currently in circulation with reputable names attached. This article works through what the data actually says, why it disagrees with itself, and what a South African business should change as a result.

The numbers, and who produced them

Taking the most methodologically transparent studies in turn.

SparkToro, published 8 June 2026. Using Similarweb clickstream panel data covering United States Google searches from January to April 2026, 68.01 percent of searches ended without a click, up from 60.45 percent in 2024. Put another way, only 276 of every 1,000 United States Google searches now reach the open web, down from 374 two years earlier. Two caveats belong with that figure: the panel was weighted two-thirds mobile and one-third desktop, and it excludes searches made inside Google’s mobile app, where Rand Fishkin expects zero-click behaviour is higher still. The often-quoted companion figures, that AI Overviews appear on more than 20 percent of searches and cut click-through by roughly 60 percent when present, are Ahrefs data cited within SparkToro’s write-up rather than SparkToro’s own measurement.

The same study found that only 0.34 percent of searches transitioned into AI Mode during the period, which sits awkwardly beside Google’s announcement at I/O on 19 May 2026 that AI Mode had passed one billion monthly active users. In fairness to both, Google also stated that AI Mode queries were more than doubling every quarter, so a very small share today is entirely compatible with a large share within a year.

Ahrefs, published 4 February 2026. Click-through rate for the top-ranking organic result was 58 percent lower on queries where an AI Overview is present. The study examined 300,000 keywords, split evenly between 150,000 with an AI Overview and 150,000 informational keywords without one, using aggregated Google Search Console desktop data from December 2025 against a December 2023 baseline. This updates their April 2025 measurement of 34.5 percent, so the measured effect close to doubled between the two studies. The impact thins further down the page: position two was down 50.8 percent, position three down 46.4 percent, and pages ranking tenth still lost close to 20 percent.

Pew Research, published July 2025. Based on the browsing behaviour of 900 United States adults during March 2025, covering 68,879 unique Google searches, users clicked a traditional search result on 8 percent of visits where an AI summary appeared, compared with 15 percent of visits where one did not. Only 1 percent clicked a link inside the AI summary itself. Users were also more likely to end their browsing session entirely after a page with an AI summary, at 26 percent against 16 percent. This remains the most rigorous user-behaviour study in the space because it observes real people rather than inferring from aggregate traffic. Worth noting for the section below: in that dataset, 18 percent of all searches produced an AI summary.

Each of these is defensible. They also produce different pictures, and the reason is methodological rather than anyone being wrong.

Why the studies disagree

Four sources of variance account for most of the spread, and knowing them lets you read any new study critically rather than reacting to the headline.

Keyword set selection. This is the largest single source of variance, and the published prevalence figures are genuinely far apart. Semrush tracked more than 10 million keywords through 2025, recording a peak near 25 percent in July before settling at 15.69 percent by November. Conductor measured 25.11 percent across 21.9 million queries in the first quarter of 2026. BrightEdge, whose tracker follows nine commercial verticals, recorded 48 percent of tracked queries by early 2026. Pew, sampling real consumer browsing rather than a marketing keyword list, measured 18 percent in March 2025.

None of those organisations is wrong. They are measuring different keyword universes. A tracker weighted towards informational and commercial marketing keywords will find high prevalence, because those are precisely the queries that trigger Overviews. A sample drawn from real browsing behaviour, full of navigational and branded searches, will find far less. When a supplier quotes a single prevalence figure without naming the keyword set behind it, the number is close to meaningless, and the 48 percent is the one most often repeated without that context.

Panel composition. Clickstream panels skew towards particular demographics, devices and countries. Mobile behaviour differs sharply from desktop, and studies weighting them differently reach different conclusions about the same underlying reality. The long-run zero-click trend line is a good example of the problem: it stitches together Jumpshot data from 2019, Datos data from 2024 and Similarweb data from 2026. SparkToro flags this themselves. The direction is consistent, but the year-on-year comparisons are not strictly like for like.

Measurement period. Google has been changing this feature continuously. A study running January to April 2026 and one running February 2026 are measuring different products, and the I/O changes in May merged AI Overviews and AI Mode into a single flow, which makes pre-May and post-May data awkward to compare at all.

Geography. This is the one that matters most for readers of this article, and it is barely acknowledged in the industry conversation. Almost all of the research is United States data. Feature rollout timing, competitive density, language mix and device profiles all differ in South Africa. When someone quotes a 58 percent decline to a business in Bryanston, they are quoting a number measured in a different market.

The asymmetry that determines your actual exposure

The most useful finding across the entire body of research is that the impact is wildly uneven by query type, and this is what should drive your assessment rather than any headline percentage.

Informational queries are hit hardest. Definition questions, how-to content, comparison research and general explainers are precisely what generative answers handle well. Encyclopaedic and reference content has seen the steepest declines anywhere in the data.

Transactional and local queries are affected far less. Someone searching for a plumber in Randburg needs a phone number, an address and a booking. Someone searching for a WooCommerce developer needs to evaluate a supplier and start a conversation. A generative summary does not complete either of those jobs, and the data consistently shows commercial and local-intent queries retaining click behaviour that informational queries have lost. Fishkin made the same point directly when publishing the zero-click figures, naming branded searches, local business queries and high-intent transactional searches as the categories where search still returns meaningful click volume.

So the honest exposure assessment for a South African business is not “how much traffic will I lose”, it is “what proportion of my traffic came from queries that a generated answer can now satisfy completely”. For a professional services firm in Johannesburg North whose traffic is dominated by service and location queries, the answer is often surprisingly little. For a business whose traffic came primarily from broad informational blog content, the answer is a great deal.

This is worth doing properly rather than by intuition. Export your Search Console query data for the past 12 months, segment it into informational, commercial and navigational buckets, and calculate what share of clicks sits in the informational bucket. That percentage is your real exposure, and it will be specific to you rather than borrowed from an American study.

The pattern in the data most businesses miss

There is a finding buried in most of this research that reframes the entire problem, and it is more useful than the decline figures.

The clicks that survive are qualitatively different. A user who reads a generated summary and then chooses to click through has already absorbed the basic answer and is clicking because they want something the summary could not give them: a supplier, a price, a deeper analysis, a decision. The generated answer has performed a filtering function. It absorbs the casual traffic and passes through the committed traffic.

Pattern worth checking: We have seen this pattern in client data over the past year with reasonable consistency. Sessions down, session duration up, and enquiry volume roughly stable or improved. That is not a universal law and anyone stating it as one is overreaching, but it is a pattern worth checking in your own numbers before concluding that a traffic decline is a business decline.

The businesses that panicked at the traffic chart and cut their content investment generally did the wrong thing. The businesses that looked at enquiry volume and conversion rate first made better decisions, because those metrics were closer to what actually pays the bills.

What to measure instead

Impressions against clicks

In Search Console, rising impressions with flat or falling clicks is the signature of appearing in results that resolve without a visit. This is the closest thing to an AI Overview exposure metric currently available, because Google does not break out AI Overview impressions as a separate search appearance. It is a proxy, and it should be described as one internally so nobody mistakes it for a direct measurement.

Branded search volume

If people encounter your business inside a generated answer and later search for you by name, branded search volume rises even when the article that earned the mention shows no click. This is the most underrated metric of the current period and one of the few that captures value created without a session.

Enquiries per thousand sessions

As casual traffic is absorbed, absolute session counts become a worse proxy for commercial health. Enquiries per thousand sessions normalises for that and tells you whether the traffic you are still receiving is worth more or less than it was. If sessions halve and this figure doubles, your business is fine and your dashboard is lying to you. Setting that measurement up so it reports something commercially useful is part of our website design and SEO services in Johannesburg.

Assisted conversions and direct traffic

A person who reads about you in an AI answer, does not click, and later types your website address directly appears in your analytics as direct traffic with no attributable source. Rising unattributed direct traffic alongside falling organic traffic is a strong signal that discovery is happening somewhere you cannot see.

What this means for content strategy

The tempting conclusion is to stop publishing informational content because it no longer earns clicks. That is the wrong lesson, and it misreads the mechanism described in our companion piece on AI Overview citation.

AI Overviews are generated from passages retrieved from Google’s live index. Content that does not exist cannot be retrieved, and a business with no published expertise has nothing to be cited for. Withdrawing from content publication does not protect you from the shift, it removes you from the candidate pool entirely, while your competitors remain in it.

The real adjustment is in what you publish and what you expect from it. Thin content that restates general knowledge has lost its remaining commercial rationale, because it earns no clicks and no citations. Content with genuine specificity, actual South African figures, real regulatory detail, and operational experience that cannot be synthesised from other sources retains both citation value and click value, because it gives the reader a reason to come to the source.

The uncomfortable implication is that the content marketing volume play is finished. Publishing forty mediocre articles a year to chase long-tail informational traffic was a viable strategy in 2019. In 2026 it produces content that gets summarised without attribution and generates nothing. Twelve genuinely authoritative pieces will outperform it decisively. We set out the case for that approach in Topical Authority: The Foundation of Modern SEO.

The strategic read

The businesses navigating this well are the ones that were already measuring enquiries, pipeline and revenue rather than sessions. They can see clearly that fewer visitors of higher quality is not a problem. The businesses struggling are the ones whose entire marketing narrative was built on a traffic graph, and who now have to explain a declining chart that may not correspond to anything happening in the business at all. That shift in what actually gets measured is the substance of Search Experience Optimisation: The Next Evolution of SEO.

Before you respond to what AI Overviews are doing to your traffic, work out whether they are doing anything to your revenue. Those are different questions, and only one of them matters. If you would like that assessed against your own numbers rather than a headline statistic, request a quotation.

Frequently Asked Questions

How much traffic will AI Overviews cost my website?

It depends almost entirely on your query mix. Informational queries have seen the steepest declines, while transactional and local queries are far less affected. Segment your Search Console data by intent and calculate what share of your clicks come from informational queries. That percentage is your actual exposure.

Why do different studies report such different numbers?

Mainly because of keyword set selection, panel composition, measurement period and geography. Some of the spread is also misreading: the widely quoted 48 percent figure measures the share of industries where AI Overviews appear, not the share of searches. Almost all of the underlying research is United States data.

Should I stop publishing blog content?

No. AI Overviews are generated from passages retrieved from Google’s live index, so content that does not exist cannot be cited. Withdrawing removes you from the candidate pool while competitors remain in it. The adjustment is towards fewer, genuinely specific pieces rather than high-volume thin content.

Can I see AI Overview impressions in Google Search Console?

Not as a separate search appearance. Google does not currently break out AI Overview impressions. The usable proxy is the ratio of impressions to clicks, where rising impressions against flat clicks suggests visibility in results that resolve without a visit.

Is South African search affected the same way as the United States?

There is very little South African data, so this is genuinely unknown. Rollout timing, language mix, competitive density and device profiles all differ. Treat United States figures as directional rather than as a forecast for your own market.