The Last Clicks: Who Measures the Internet When Nobody Clicks Anymore?

01 · The Number

276 out of 1,000

SparkToro analyzed US search traffic for January through April 2026. Result: 68.01 percent of all Google searches end without a single click. In 2024 it was 60.45 percent, in 2019 around 49, in 2016 around 45.

More interesting than the percentage is the conversion to a thousand searches, because it shows the ratio that matters to operators. Of 1,000 searches, 276 clicks arrive at the open web. In the 2024 survey it was 374.

A quarter of the clicks per thousand searches has disappeared in two years. Not because people search less, but because the answer now appears at the top of the page.

27.6%

Share of Google searches that still send a click to the open web in 2026. In 2024 it was 37.4 percent.

What does "without click" mean?

Not that someone was blocked. It means that the search was already answered before there was anything to click on. The answer was on the results page itself.

This is older than any AI. The weather has been displayed directly there for years. Exchange rate, math problem, opening hours, map section, song lyrics, sports result: Google answers everything on its own page. Since 2024, the AI summary has been added at the very top, and it can answer far more questions than a weather tile.

For you as a searcher, this is an improvement. You get the answer faster and don't have to load a page with banners.

For the website whose text this answer comes from, it's a visit that never happens. Both statements are true at the same time, and that's why this matter is so hard to discuss.

In the image below, both are therefore the same unit: one symbol is a search query. A filled dot sent a click, an empty circle was answered at the top. What shifts is only the ratio.

Fig. 00 · Each time slice from 2016 to 2026 gets exactly 20 searches, so both stacks together are always the same height and the view compares shares instead of quantities. The four white marks on the center line are the actually measured values (45, 49, 60, 68 percent), in between is linearly interpolated. Source of measurement points: SparkToro, panels from Jumpshot, Datos and Similarweb, see section 03.

68.01%

Searches without click

USA, January through April 2026. Ten years earlier it was around 45 percent.

−26%

Clicks per 1,000 searches

From 374 to 276 in two years. This is the loss that operators actually feel.

20%+

Searches with AI answer

Where an AI summary appears, the click rate drops by nearly 60 percent.

0.34%

Share AI mode

The pure chat mode is tiny so far. The loss comes from the normal search page, not from the chat.

The last tile is the one that gets corrected the most. In the debate it sounds as if everyone is migrating to a chatbot. AI mode so far accounts for one-third of a percent of searches. The click loss occurs almost entirely on the completely normal results page, through the summary at the top and through everything else Google places there to keep you on the page.

02 · How to Measure This

To see the not-clicking, you have to watch the clicking

Google does not publish this number. There is no statistic about how many searches end without a click that comes from the search engine itself. Every number in section 01 comes from outside.

And from outside there is exactly one method: you observe people searching. Not the search engine, not the destination pages, but the devices in between. Exactly the panel method from Data 01, just applied to a different question.

This results in a chain with four links, and every single one of them is an observed browser. No link can do without.

The consequence of this runs through the rest of this report: the statement "the click is dying" is only as reliable as the instrument watching it die. And this instrument is currently having problems itself.

Link 01

Extension

A tool in the browser records which address is called. Usually a product with a different purpose.

→

Link 02

Session

From the address sequence it is reconstructed where a search began and whether something else came after. On mobile, a session is considered ended after 10 seconds of inactivity.

→

Link 03

Classification

Each subsequent call is sorted: click to the open web, click back to Google, or no click.

→

Result

A rate

68.01 percent. Weighted with the assumption two-thirds mobile, one-third desktop.

Two assumptions in links 02 and 03 carry the entire result. The session boundary of ten seconds decides whether a later call still belongs to the search. The split of two-thirds to one-third decides how strongly the mobile world impacts, where the rate is significantly higher. Both are disclosed and both are reasonable. But both are settings, not measurements.

03 · The Curve Everyone Cites

Four points, three panels, three different companies

The time series from 2016 to 2026 is shown everywhere as one line. It is not. It is composed of three different data sources that have never seen the same people and never the same devices. SparkToro says this itself, and that is why we say it here too.

Fig. 01 Where each point of the curve comes from

Share of Google searches without click, USA. The bars are comparable in statement, not in survey.

2016 · Jumpshot

~45 %

2019 · Jumpshot

~49 %

2024 · Datos

60.45 %

2026 · Similarweb

68.01 %

Point	Panel	Who owns it today	Value
2016	Jumpshot	No longer exists. Subsidiary of antivirus manufacturer Avast, shut down in January 2020.	~45 %
2019	Jumpshot	ditto	~49 %
2024	Datos	Semrush held the majority since December 2023. Semrush has belonged to Adobe since April 2026.	60.45 %
2026	Similarweb	Publicly traded, NYSE. The panel from Data 01.	68.01 %

SparkToro explicitly points out that the earlier datasets come from other panels, thus "not the same users or devices". The curve is a statement about a development, not a continuous measurement series.

And then there's the matter with Jumpshot. The two oldest points of the most important curve about the state of the web come from a company that no longer exists because it was shut down for exactly that reason. Jumpshot was the data arm of antivirus manufacturer Avast, had access to around 100 million devices and sold browsing histories to Google, Microsoft, Yelp and McKinsey, among others. After a joint investigation by Vice and PCMag in January 2020, Avast shut down the subsidiary within days. In 2024, the US Federal Trade Commission banned Avast from selling such data and imposed 16.5 million dollars.

The history of the click is written with instruments that in some cases are no longer allowed to exist.

04 · The Contradiction

The only one who knows exactly says the opposite and shows nothing

In August 2025, Liz Reid, head of Google Search, stated that total organic click volume to websites was "relatively stable" year-over-year and that click quality had even increased slightly. Regarding external studies, she said they were based on "flawed methods, individual cases, or traffic changes that occurred before the introduction of AI features."

This is a serious statement, because Google is the only entity that actually knows the truth. However, it came without a dataset, without a time series, and without a chart.

And it stands against three independent measurements using very different methods, all pointing in the same direction.

The third carries the most weight because it is the only one that is a true experiment rather than an observation: Two researchers built their own Chrome extension and randomly divided 1,065 participants into groups that were shown or had the AI summary removed. This allows cause to be separated from correlation, which a panel observation fundamentally cannot achieve.

This experiment also ran, you guessed it, via a browser extension.

Fig. 02 Three measurements, three methods, one direction

All three measure something different and therefore cannot be converted into one another. That is precisely what makes the agreement in direction significant.

Study	Method	Finding
Pew Research Center, July 2025	Observation. 900 adults with installed history recorder, March 2025, 68,879 searches of which 12,593 with AI summary.	With summary, 8 percent clicked on a result, without 15 percent. On the reference in the summary itself, 1 percent clicked.
SparkToro, May 2026	Observation. Similarweb panel, USA, January to April 2026.	68.01 percent without click. Per 1,000 searches, 276 clicks to the open web.
Agarwal (ISB) and Sen (Carnegie Mellon), 2026	Randomized field experiment. Own Chrome extension, 1,065 US participants, random assignment, January and February 2026.	Where the summary appeared, 38 to 40 percent fewer clicks outward. Reported satisfaction remained nearly the same.
Google, August 2025	Statement without published data.	Click volume "relatively stable", click quality slightly increased.

The last entry is deliberately in the same table. It is the position of the best-informed party and belongs alongside, even though it is not verifiable.

Both sides can be right. If the number of searches overall increases sharply, the absolute number of clicks can remain stable while the share per search falls. That is exactly what Google's wording aims at. For an individual operator, this is still no consolation, because they see the rate, not the global volume. As long as Google does not show a time series, this cannot be resolved.

05 · The instrument goes blind

The panel sits in the browser. Less and less is happening there.

Here comes the second loss, and it does not hit the websites, but the measurement itself.

A panel of browser extensions sees what runs through a browser. On mobile phones, according to common measurements, that is around eight percent of usage time, the rest happens in apps. And search is moving there: into the Google app, into ChatGPT, into assistants that no longer open a page at all.

SparkToro writes this limitation itself in the footnote, and the sentence is the most honest of the entire study: the data does not include the mobile Google app, but only mobile searches in the browser.

The definitive study on whether people still click therefore cannot see the place where most people search. This does not make it worthless, it makes it a statement about the browser portion of the world. This limitation is the first to fall away when cited further.

Fig. 03 What a browser panel still sees of the web

Three different quantities, deliberately side by side: time on mobile, share of machine traffic, and what an AI system sends back.

Mobile time in browser

~8 %

Mobile time in apps

~92 %

HTML traffic from machines

57.5 %

HTML traffic from humans

42.5 %

Mobile time: common market measurements, magnitude 90 to 92 percent in apps. Machine share: Cloudflare, as of June 3, 2026, based on HTML requests in their own network.

And what comes back?

Cloudflare has built a metric for this that is more precise than any survey: how many pages does a system fetch before it sends back a single visitor? It is derived from their own network traffic, without panel and without survey.

Fig. 04 Pages fetched per visitor sent back

Logarithmic view necessary: the values span four orders of magnitude. Therefore shown here as a table with bars on a square root scale, the number next to it is the truth.

DuckDuckGo

1.5 : 1

Google

5.2 : 1

OpenAI

1,276 : 1

Anthropic, mid-2026

11,122 : 1

Anthropic, early 2026

23,951 : 1

Bar length on square root scale, so that 1.5 and 23,951 fit in one image. The number column is the true value. Cloudflare Radar, values fluctuate greatly with the measurement period: for Anthropic, 70,900 : 1 was also measured at the end of June 2025.

The comparison is unfair and yet meaningful. Unfair because Google operates a search engine and Anthropic does not: those who have no product that sends people onward cannot send visitors. Cloudflare itself mentions another limitation, namely that requests from an app carry no referrer header and the ratio can therefore be too high. It is meaningful because it shows the direction of value creation: content flows in one direction, attention does not come back in the same direction.

06 · The New Guild

When there are no more clicks to count, you count mentions

An entire industry is currently building the measurement tool for the world after. The question is no longer "what rank am I," but "does the system mention me when someone asks."

The method behind this is remarkable because it flips the panel. Instead of watching people search, you ask the models themselves, thousands of times, with invented but typical questions, and count which brands appear in the answers. Semrush renews 158 million questions monthly for this purpose, according to its own statements.

This is not a panel, this is a sensor. And it has a property that shifts everything: it doesn't measure what people have actually experienced, but what the system would answer if you asked it.

This eliminates the calibration standard from Data 01. With traffic, there were websites that provided their real numbers, and the model could be calibrated against them. With an AI answer, there is no truth to calibrate against. There are only more measurements.

Fig. 05 Who is currently building the measurement tool

Over 300 million dollars in venture capital flowed into this category between summer 2025 and spring 2026.

Provider	Approach	Status 2026
Profound	Complete platform: measure, analyze, generate content, suggest actions.	Around 155 million dollars raised, valuation 1 billion dollars early 2026.
Peec AI	Pure measurement accuracy: replicate real user interactions, compare regionally, across over 115 languages.	Over 30 million dollars, 4 million dollars recurring revenue in ten months.
Otterly	Self-service, transparent pricing, four systems in basic plan.	The affordable entry point of the category.
Semrush (Adobe)	Clickstream panel from 200 million users plus 158 million monthly renewed questions to ChatGPT, Gemini, Perplexity and Google's AI mode.	Part of Adobe since April 2026.
Similarweb	Own panel plus a product that reports traffic from chatbots as a separate source.	Publicly traded, see next section.

Semrush describes its panel as built from "browser extensions, antivirus software and VPN apps." The same construction as in Data 01, just larger.

With clicks there was a truth to calibrate against. With an answer there are only measurements.

07 · What the Stock Market Thinks

Two publicly traded measurement firms, two opposite paths

Until recently there were exactly two companies whose business is measuring other people's web traffic and whose numbers are public: Semrush and Similarweb. Both experienced the opposite of each other in the same eighteen months.

Semrush was acquired. Adobe announced the acquisition on November 19, 2025, 1.9 billion dollars, 12.00 dollars per share in cash. The stock jumped 74 percent on the day of the announcement. On April 28, 2026, the purchase was completed, and the stock has been delisted since then.

Similarweb was written down and then revalued. The stock fell from nearly 21 dollars after the IPO to 2.59 dollars in February 2026, a decline of 88 percent. Precisely during these weeks, the narrative of the end of search was at its peak.

Since then, the stock has nearly tripled to 7.51 dollars. The trigger was not better traffic numbers, but the contracts for training data and the products around AI traffic. The market has recategorized the company: from measurement service provider for a dying metric to data supplier for the systems replacing it.

Fig. 06 Similarweb (NYSE: SMWB), month-end in US dollars

September 2021 to August 8, 2026. IPO in May 2021 with 1.6 billion dollar valuation, today around 658 million.

Milestone	Price	What happened
Sept. 2021	20.93 $	Highest month-end close after IPO in May 2021
Jan. 2025	16.21 $	Last interim high
Feb. 2026	2.59 $	Low. Down 88 percent compared to 2021
May 2026	4.15 $	Succession search for founder announced, AI data contract in quarterly report
Aug. 8, 2026	7.51 $	Up 190 percent from low, still 64 percent below high

Month-end prices. Market capitalization on August 7, 2026: 658 million dollars at 87.6 million shares. Revenue last twelve months 289 million dollars, earnings minus 30 million.

The stock price tells the thesis of this report more precisely than any survey. Until early 2026, what was valued was a company that measures how people click in browsers. What is valued since then is a company that sells behavioral data to the systems replacing clicking. Same company, same panel, different market.

08 · The actual problem

No truth, no calibration

In Data 01, the calculation path for every traffic estimate had four steps: collect, adjust, calibrate, extrapolate. The third step carries the whole thing. Without a set of cases where you know the truth, an extrapolation is just a guess with decimal places.

For website traffic, this truth exists: tens of thousands of operators provide their real analytics data. For the questions in this report, it doesn't exist.

Nobody can verify whether 68.01 percent is correct, because only Google has the comparison number and doesn't show it. Nobody can verify how often a brand appears in AI answers, because the answer varies for each user and is logged nowhere.

This means the fourth step is no longer secured in both cases. What remains is step one and two: collect and adjust, without control afterwards. That's not worthless, but it's a different kind of number, and it should be read differently.

Step 01

Collect

continues

Panels and sensors deliver more raw data than ever before.

Step 02

Adjust

becomes harder

The panel sits in the browser, usage migrates to apps. The imbalance grows faster than it can be calculated out.

Step 03

Calibrate

drops out

For click-through rates and AI mentions, there is no public truth to calibrate against. This is where the chain breaks.

Step 04

Extrapolate

uncovered

Continues to be done and continues to be cited, but now without the safeguard that justified the previous step.

That's why section 03 shows the source of every single data point. When calibration is missing, the origin of the data is the only thing you can still use to assess reliability. A number without a truth anchor must name its sample all the louder.

09 · Context

Four sentences to get the picture right

First: the direction is undisputed. Three studies with three different methods show the same thing. Anyone who doubts that search is delivering fewer clicks is arguing against the data.

Second: the magnitude is not. 68.01 percent is a number for the USA, for searches in browsers, with a set device distribution and a set session limit. Transferring it to Germany, to apps, or to a single industry is not supported.

Third: the measurers are stakeholders. SparkToro, Semrush and Similarweb sell tools that are needed precisely when the situation is dramatic. That doesn't make their numbers wrong, their methodology is disclosed and more verifiable than Google's counterargument. It still needs to be mentioned.

Fourth: for you, only your own trajectory counts anyway. All numbers here are world averages across millions of pages. Your own analytics software measures your case directly and without a panel in between. It's the better source for this question, because here you happen to possess the truth that everyone else lacks.

The click was never the goal. It was just what was easiest to count.

10 · What this means practically

Measure your own case

Compare the impressions from Google Search Console against your clicks and look at the rate over time. That's your version of 27.6 percent, and it's the only number on this topic that truly belongs to you.

Separate share from volume

Falling click-through rate with rising impressions is different from falling impressions. The first case is the industry trend, the second is a problem you caused yourself.

Never cite without the limitation

Anyone who mentions 68 percent should also mention USA, browser, and the time period. Without these three words, the number is a claim about the entire world, and it doesn't support that.

Monitor what's being said about you

If the answer replaces the click, the new question is whether you appear in the answer. This doesn't require a tool costing 2,000 dollars a month: ask the four major systems the same twenty questions your customers ask once a month, and write down who gets mentioned. That's your own little panel, and it costs half an hour.

Sources

Everything this report stands on

Ten sources, each with what it contributes and the full address. They are intentionally set large and expanded: those who don't need them scroll past, those who want to verify shouldn't have to search.

SparkToro: In 2026, Less than One Third of Google Searches Still Send a Click

The main source. 68.01 percent without click, 276 clicks per 1,000 searches, the comparison values 2016, 2019 and 2024, and the explicit note that the earlier data points come from other panels and do not include the mobile Google app.
sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click/
Pew Research Center: Google users are less likely to click on links when an AI summary appears

The independent cross-check with its own panel: 900 adults, March 2025, 68,879 searches. 8 percent click rate with summary versus 15 percent without, and only 1 percent on the reference in the summary itself.
pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Agarwal (Indian School of Business) and Sen (Carnegie Mellon): Field experiment on AI Overviews

The only real experiment on the topic. Own Chrome extension, 1,065 US participants randomly divided into groups, January and February 2026. Result: 38 to 40 percent fewer outbound clicks, with nearly unchanged satisfaction.
searchenginejournal.com/ai-overviews-cut-organic-clicks-38-field-study-finds/573145/
Cloudflare: The crawl before the fall of referrals

The metric of pages fetched per visitor sent back, from their own network traffic rather than from a panel. Contains the calculation formula and the caveat that app accesses carry no referrer header.
blog.cloudflare.com/ai-search-crawl-refer-ratio-on-radar/
Vice Motherboard and PCMag: Leaked Documents Expose the Secretive Market for Your Web Browsing Data

The investigation from January 2020 that ended Jumpshot. Around 100 million devices, customers including Google, Microsoft, Yelp and McKinsey. Without this panel, the two oldest data points of the zero-click curve would not exist.
vice.com/en/article/avast-antivirus-sells-user-browsing-data-investigation/
TechCrunch: FTC bans antivirus giant Avast from selling its users' browsing data

The aftermath, February 2024. Sales to over a hundred companies, 16.5 million dollar fine, ban on resale. Proves that the Jumpshot case was officially closed and not just a press scandal.
techcrunch.com/2024/02/22/ftc-bans-avast-selling-customers-sensitive-browsing-data/
Adobe: Adobe Completes Semrush Acquisition

The original announcement of completion on April 28, 2026. Plus the announcement from November 19, 2025: 1.9 billion dollars, 12.00 dollars per share in cash, share price jump of 74 percent on announcement day.
news.adobe.com/news/2026/04/adobe-completes-semrush-acquisition
Semrush Knowledge Base: Semrush Data & Metrics

Their own description of the panel: over 200 million users in more than 190 countries, built from browser extensions, antivirus software and VPN apps, plus 158 million monthly refreshed queries to AI systems.
semrush.com/kb/997-semrush-data
Search Engine Land and Search Engine Roundtable on Google's counter-position

Liz Reid's statement from August 2025 that organic click volume is "relatively stable," along with criticism that Google presented neither Search Console data nor a time series.
seroundtable.com/google-clicks-stable-with-ai-overviews-39900.html
Similarweb Investor Relations and price data

Quarterly figures, market capitalization and the month-end price history for Fig. 06. As of August 7, 2026: 7.51 dollars, 658 million dollars market capitalization, 87.6 million shares, 289 million dollars revenue for the last twelve months.
ir.similarweb.com/financials/quarterly-results

Own work McGrinsey: Conversion of zero-click rate to clicks per 1,000 searches, square-root scale for Fig. 04, price series for Fig. 06 from month-end prices September 2021 to August 2026, and the animation in section 01, which runs 1,000 particles with the measured ratio 276 to 724. All translations from English are ours. Where sources contradict each other, the contradiction is stated in the text.

Connection

Data 01 asked where a number comes from. Data 02 asks what happens when it no longer exists.

Data 01: The Super Data Collectors dissected the calculation method of every traffic estimate and showed that everything depends on the third step, the calibration against real numbers. This report shows what happens when this step is no longer available for a new question.

Data 03 addresses the third level: When models are trained with data from panels, and their answers are measured again by panels, where does an independent observation come from? And what has actually become the price of a data point?