The Web Gained 1.2 Billion Users in Five Years. Visits Per User Fell 29%.
There are roughly 1.2 billion more people online than there were at the end of 2021. They are visiting the web's biggest sites less, not more.
New research published on 10 November by SparkToro co-founder Rand Fishkin, built on Similarweb panel data, puts total visits to the world's top million websites 12.2% below their early 2022 peak. Over the same period, global internet users grew by about 24%. Divide one by the other and visits per web user are down 29.2% (named-party figure, SparkToro/Similarweb).
That is the headline for publishers. The more useful question, for anyone selling or buying attention on the open web, is what replaced those visits.
What did SparkToro and Similarweb measure?
Fishkin asked Similarweb for the total visits flowing to the top one million websites worldwide, across mobile and desktop browsers, from 2021 to 2026. The data was split into rank buckets (top 10, 11 to 100, 101 to 5,000, 5,001 to 100,000 and 100,001 to 1,000,000) and set against DataReportal's quarterly estimates of global internet users.
Three findings stand out:
- Total visits peaked in Q1 2022 at 2.243 trillion. They hit a low of 1.914 trillion in Q1 2026 (down 14.7%) and stood at 1.970 trillion in Q3 2026 (down 12.2%).
- Users kept growing. DataReportal's estimates rise from 4.95 billion to 6.15 billion over the period.
- Distribution barely moved. The top 100 sites gained only about 3 percentage points of share. Fishkin's starting theory, that Big Tech was hoarding a growing slice of browser traffic, mostly did not hold.
Definition: in this study, the "open web" means the top million websites as visited in a browser. It excludes in-app activity, and it counts visits by people in Similarweb's panel, not requests made by software.
Why did the headline change from 24% to 12%?
The study first circulated on 10 November under a "24%" headline, and that number travelled through newsletters. The post's title now reads 12%; its URL still contains 24. Fishkin said on X that he had replaced an incorrect graph. The 24% figure is the growth in users, not the fall in traffic. Anyone citing the work should use 12.2% for the decline in total visits and 29.2% for the decline in visits per user.
Is AI the cause?
The study does not say, and it is careful not to. Fishkin lists several candidates: people moving to apps, search engines, AI tools and social networks answering needs in place, and the fading of the 2021 to 2022 lockdown peak. He also notes that the shift predates AI answers. SparkToro's earlier work found that 68% of US Google searches from January to April 2026 ended without any click (named-party figure, SparkToro/Similarweb).
Two caveats pull in opposite directions. The data is global, so declines at US, UK and EU sites can be masked by growth in Asia and Africa; the picture for Western publishers is probably worse than the aggregate. And the panel only sees browsers, so app consumption is invisible.
A third limit matters most for our beat: a human visitor panel does not count machine reads at all.
Where did the reading go?
A visit records a person arriving on a page. It does not record whether the page was read. Increasingly, pages are read by software acting for a person who never arrives.
When someone asks ChatGPT, Perplexity, Claude or Google's AI Mode a question, a Live Search Agent (a bot dispatched in real time to fetch sources for that answer) can pull a publisher's page, extract what it needs and hand the result back inside the chat. These agents generally do not execute JavaScript. Analytics tags do not fire, ad tags do not load and no auction runs. The read happens; the visit is never logged.
Other data shows the scale of that shift. IAB Tech Lab estimates that 51% of internet traffic is now automated and projects 90% by 2030 (named-party figure). OPA Europe and Rossel found that 1 in 50 publisher visitors is an AI bot, and that publishers see 7x more AI bot traffic than other industries (named-party figure, January 2026). Cloudflare reports that personal AI agent traffic grew 15x by 2025 (vendor-stated).
None of those figures prove that agents explain SparkToro's missing visits. They do show that demand for publisher content did not disappear along with the clicks.
What does it mean for publishers?
- The decline is per head, so it is structural. A quarter more people online should have lifted visits. It did not. Waiting for audience growth to refill the funnel is not a plan.
- Your dashboard undercounts your reach. If agents are reading your pages, your audience is larger than your analytics shows. Server-side and edge logs are the only place that audience appears.
- Blocking will not bring the visits back. Content still reaches AI answers through syndicated copies, open platforms such as YouTube and Reddit, and bots that ignore the rules; OPA Europe and Rossel found 13% of AI bots ignore robots.txt (named-party figure). Blocking mostly hands authority to whoever stays open.
- The unit to price is the read. A referral model assumes a visit follows the read. On these numbers, fewer of them do every year.
What does it mean for advertisers?
Fewer visits per person means fewer impressions per person on open web inventory, at the same time as decisions move into AI answers. Display, paid search and social units do not appear inside an AI answer. A brand that is absent from the answer is absent at the moment of choice, however well its open web buy performs against a shrinking visit base.
That makes brand visibility in LLM answers (how often, and how, a brand appears when an assistant responds to a relevant question) a reach metric in its own right, and one that needs to be measured and bought, not just hoped for through organic optimisation.
What should we watch next?
- Whether the follow-up validation Fishkin has promised, using additional data sources, confirms the per-user decline.
- Regional cuts: publisher-only figures for the US, UK and EU would show how much the global average is hiding.
- Whether any measurement vendor starts reporting agent reads alongside human visits. Until one does, the industry is pricing a market it can only half see.
the blankspace read: A 12% fall in visits is not a 12% fall in reading. The web's audience grew by a quarter while its visit count shrank, and over the same years AI systems learned to read pages on a person's behalf without ever registering as a visit. The open web's problem is not that people stopped wanting its content. It is that more and more of that content is now consumed in a form that does not count, and does not pay. Chasing the referral back is a losing game. Monetise the read, not the referral.
Key figures
| Figure | Value | Provenance |
|---|---|---|
| Peak total visits, top million sites (Q1 2022) | 2.243 trillion | Named-party figure (SparkToro analysis of Similarweb data) |
| Low point (Q1 2026) | 1.914 trillion, down 14.7% from peak | Named-party figure (SparkToro/Similarweb) |
| Latest (Q3 2026) | 1.970 trillion, down 12.2% from peak | Named-party figure (SparkToro/Similarweb) |
| Global web users | 4.95B (Q4 2021) to 6.15B (Q3 2026) | Reported (DataReportal, as cited by SparkToro) |
| Change in visits per web user | Down 29.2% | Named-party figure (SparkToro calculation) |
| Share shift to top 100 sites | About 3 points | Named-party figure (SparkToro/Similarweb) |
| US Google searches ending with no click (Jan to Apr 2026) | 68% | Named-party figure (SparkToro/Similarweb) |
| Internet traffic that is automated | 51% now; 90% projected by 2030 | Named-party figure (IAB Tech Lab) |
| Publisher visitors that are AI bots | 1 in 50 | Named-party figure (OPA Europe/Rossel, Jan 2026) |
| AI bot traffic to publishers vs other industries | 7x | Named-party figure (OPA Europe/Rossel, Jan 2026) |
| AI bots ignoring robots.txt | 13% | Named-party figure (OPA Europe/Rossel, Jan 2026) |
| Growth in personal AI agent traffic | 15x by 2025 | Vendor-stated (Cloudflare) |
Primary source, opened and verified: SparkToro, "New Research: Traffic to the Open Web Has Fallen 12% in 5 years; Visits/web User Are Down 30%" (Rand Fishkin, 10 November 2026, data from Similarweb)