The publisher traffic collapse now has its own genre of headline, and nearly every number in that genre was measured on properties far larger than yours. Then a site owner doing 1,400 sessions a month reads a 58 percent figure, opens their analytics, sees a bad July, and concludes the same thing happened to them. It might have. It might also be one page slipping two positions, a 28 day month, or a consent banner change. This piece tells those apart.
Where the headline numbers actually come from
Start with the strongest non-vendor evidence. The Pew Research Center recruited 900 U.S. adults from KnowledgePanel Digital and recorded their browsing for the whole of March 2025, then collected the corresponding search results between April 7 and April 17, 2025. Across 68,879 unique Google searches, 12,593 produced an AI summary. Users clicked a link on 8 percent of visits where a summary appeared, against 15 percent where none did, and clicked a link inside the summary itself on 1 percent of visits. The numbers are at Pew Research Center.
Alongside that sits the zero-click line. SparkToro, using a Similarweb clickstream panel, put 68.01 percent of Google searches ending without a click across January to April 2026. Ahrefs, using aggregated Google Search Console data across 300,000 keywords, first reported a 34.5 percent lower click-through rate for position one on keywords that trigger AI Overviews, then raised that to 58 percent on May 18, 2026.
Three instruments, three units. Pew counts human visits. SparkToro counts panel searches. Ahrefs counts clicks reported by Google itself. None is measuring your site.
The 76 to 38 percent drop is partly an instrument change
The most repeated statistic of 2026 is that the share of AI Overview citations coming from pages ranking in Google’s top 10 fell from about 76 percent to about 38 percent. The Ahrefs study behind it, published March 2, 2026 by Louise Linehan and Xibeijia Guan, analyzed 863,000 keyword result pages and 4 million AI Overview URLs. The earlier figure came from roughly 1.9 million citations in July 2025.
Buried in the post is the sentence that should have led the coverage: "Since that last study we’ve also improved our parsing methodology in Ahrefs, so that we can see even more of the citations that appear in AI Overviews." The citation volume tracked more than doubled. If the newer instrument detects citations the older one missed, and those skew toward sources outside the top 10, then part of that swing is the detector improving rather than Google changing its mind. Ahrefs also notes the analysis ran after AI Overviews moved to Gemini 3 in January 2026, so the system changed too. Both can be true, and neither is separable from the other with the published data.
The same discipline exposes a smaller problem with the CTR numbers. The 34.5 percent study compared March 2024 with March 2025. The 58 percent update compared December 2023 with December 2025. The baseline moved and the window doubled, yet the two were reported as a before and after.
What publisher traffic looks like below a million sessions
Large publishers have a statistical luxury small sites do not: their traffic spreads across thousands of URLs, so no single ranking movement can shift the total much. Their aggregate is a genuine average.
Small sites are the opposite. Ahrefs’ analysis of roughly 14 billion pages found that "96.55% of all pages in our index get zero traffic from Google, and 1.94% get between one and ten monthly visits." That distribution does not stop at the site boundary. Inside a small site the same shape repeats: a handful of URLs carry most of the search traffic, and the rest carry almost none.
So when publisher traffic at a small site falls 20 percent in a month, the arithmetic is usually local. Take a site doing 1,200 clicks a month where five URLs supply 800 of them. One of those five slips from position two to position five and loses 200 clicks. The site total drops 17 percent. Nothing about AI Overviews or the future of the open web is required to explain it, and no industry study will diagnose it.
Your effective sample size is not your click count
Here is the statistical core. If clicks arrived independently, a site with 1,000 monthly clicks would see counting noise of roughly the square root of 1,000, about 32 clicks, or 3 percent. At that level a 20 percent drop would be unmistakable.
But clicks do not arrive independently. They arrive in clusters, one per ranking URL, and everything inside a cluster moves together when that URL’s position changes. In survey terms this is a design effect: your effective sample size is closer to the number of independently ranking pages than to the number of sessions. If 20 URLs carry your search traffic, expected variation is nearer one over the square root of 20, around 22 percent, before seasonality or month length.
So a 20 percent monthly swing is roughly one standard deviation of noise for a site resting on 20 pages, and roughly six for a national publisher resting on 20,000. The same percentage is a shrug in one case and an emergency in the other.
Google’s own documentation adds a second layer. The Search Console performance report "omits some queries that are searched a very small number of times," and its table "can display a maximum of 1,000 rows." Small sites have proportionally more of those rare queries, so the query table you are staring at is a thinner and more volatile slice of reality than it looks. The caveats are listed at Google Search Console Help.
Google’s own account, and why it settles nothing
On August 6, 2025, Liz Reid, VP and Head of Google Search, published a rebuttal to the decline coverage. Google’s position is that "total organic click volume from Google Search to websites has been relatively stable year-over-year" and that "average click quality has increased and we’re actually sending slightly more quality clicks to websites than a year ago," where a quality click is one where users do not quickly click back. Google characterized the contrary studies as "often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features in Search." The full post is on Google’s blog.
Treat this as you would any vendor benchmark: self-reported, unaudited, measured with an instrument nobody outside the company can inspect. Google published no aggregate figures, no methodology and no definition of the click-back threshold. That does not make it false. It makes it unfalsifiable, which is a different problem. It is also compatible with the studies, because aggregate stability across the whole web says nothing about the distribution underneath it, and the sites that lose are the ones least able to detect the loss.
Four checks that separate signal from noise in publisher traffic
Run these in order before concluding anything. Each one kills a common false positive.
2. Compare impressions against clicks. Flat impressions with falling clicks points at what is happening on the results page, which is the AI Overview story. Falling impressions points at ranking, indexing or demand, which is your story.
3. Split by query intent. Ahrefs found that 99.2 percent of keywords triggering AI Overviews are informational in intent. If your traffic is mostly navigational or transactional, your exposure is structurally lower than the headlines imply.
4. Check your non-search channels as a control. If direct, email and referral fell by a similar amount over the same weeks, you are looking at a tracking change, a site problem or a seasonal dip.
Check four catches more errors than the other three combined. Consent mode updates, a new tag manager container and a switch to server-side tagging all produce clean double-digit drops that look exactly like an algorithm penalty. If you run WordPress, a plugin update touching analytics or caching belongs on the same list.
What to measure when your numbers are too small to test
If your effective sample size cannot support a monthly comparison, stop making monthly comparisons. Move to a 13 week rolling median compared year over year. It absorbs month length, day-of-week composition and single-page volatility.
Measure at the log level rather than the tag level. Raw access logs from your CDN or origin see requests that client-side analytics never records, including crawler activity absent from any session report. Access varies by host, as our WP Engine and Pantheon comparison covers.
Track what AI systems take against what they return. Cloudflare’s crawl-to-refer ratio, published July 1, 2025 by David Belson and Sam Rhea, divides HTML requests from a platform’s crawler by HTML requests carrying that platform’s referrer. For the week of June 19 to June 26, 2025, Anthropic’s ratio sat at roughly 70,900 crawls per referral. Cloudflare states plainly that traffic referred by Claude’s native app carries no referrer header, so "these calculations may overstate the respective ratios, but it is unclear by how much." That candor is what a usable measurement looks like, and the Cloudflare analysis is worth reading.
Finally, build a denominator you own. A subscriber list measured through something like Mailchimp has no ranking volatility, no anonymized query threshold and no design effect. It is a smaller number than your search traffic and a more honest one.
Frequently Asked Questions
Is publisher traffic really collapsing, or is it a measurement artifact?
Both, in proportions that vary by number. Pew’s click-rate gap of 8 percent versus 15 percent came from recorded human browsing and is hard to explain away. The widely quoted citation and CTR swings are more fragile, because the instruments that produced them changed between measurements.
Why did the top-10 citation figure fall from 76 percent to 38 percent?
Partly because Google changed, partly because Ahrefs did. Ahrefs states it improved its parsing methodology between the two studies and tracked more than double the citation volume the second time. Better detection of previously invisible citations produces some of that drop on its own.
How many monthly clicks do I need before a percentage decline means anything?
Click count is the wrong denominator. What matters is how many separate URLs carry your search traffic. If it is fewer than about 30, expect double-digit monthly swings from normal ranking movement and compare quarters instead of months.
Should I trust Google Analytics or Google Search Console?
Neither alone. Search Console reports impressions and clicks as Google counts them, filters rare queries and caps tables at 1,000 rows. Analytics reports sessions as your tags count them, which consent settings and script changes silently alter. Disagreement between the two is normal and its direction is diagnostic.
Does Google agree that traffic to publishers is falling?
No. In August 2025 Google said total organic click volume to websites had been “relatively stable year-over-year” and that click quality had improved. It published no supporting figures or methodology, so the claim cannot be checked independently.
Do AI assistants send any traffic back to small sites?
Very little relative to what they crawl, on available evidence. Cloudflare measured Anthropic at roughly 70,900 crawls per referral in late June 2025, while noting the figure may be overstated because app-based referrals carry no referrer header. Read it as an order of magnitude.
What is the single fastest check for signal versus noise?
Compare your search decline against direct, email and referral over the same weeks. If everything fell together, the cause is almost certainly your tracking, your site or the season, not the search results page.
Are vendor studies on this topic reliable?
They are the best data available and they are still self-reported. Ahrefs, Similarweb and Cloudflare each measure through their own panels or networks and each has a commercial interest in the story. Use them for direction and magnitude, and read the methodology note before quoting any figure.
