Decodo’s analysis, published August 11, 2026, found that machines sent 57.5 percent of all web requests in June 2026, the first month automated traffic outpaced human traffic on the open web. PPC Land assembled that finding alongside several other datasets on agentic infrastructure economics in an August 25, 2026 report. Decodo also put United States traffic at 53.5 percent of the global bot total, and found automated requests growing roughly eight times as fast as human traffic. For anyone measuring a website by its server logs, that threshold changes what the number in the log even represents.
Peter Walker compiled the figures, Moses Sternstein circulated them through a16z’s Charts of the Week newsletter, and the reporting landed on August 24, 2026. They explain why the shift matters beyond raw volume. OpenRouter’s own numbers show agents burning close to five times the tokens of a human working through the same model on the same task. More than 85 percent of that consumption comes from cached prompts: agents re-reading system instructions, tool definitions and prior steps on every turn instead of producing anything new. A person typing a request into a model sends one short instruction and gets one short answer. An agent doing the same job reloads its entire context each time it acts, which makes an agent’s “request” a fundamentally different unit than a human page view, even when both land on the same line in a log file.
The scale of that gap is not fixed. OpenRouter’s data shows a fourteenfold jump in agent usage since February 2026, concentrated among the heaviest users. The top decile of enterprises now generates eight times the token volume of a typical enterprise, and those leaders’ own output is more than seventeen times what it was in April 2025. In the information sector, the leading firms now produce 32.5 times what they generated in April 2025. Legal teams, not typically associated with rapid tooling adoption, increased Codex usage 108 times over the same window, according to the same dataset.
HUMAN Security’s April 9, 2026 report arrived at a comparable crossover finding using its own methodology, a separate measurement from Decodo’s. The two figures measure different things and should not be read as confirming one another. A marketer survey referenced in the same reporting, whose sponsoring organization is not named in the source, found that 75.6 percent of respondents believe they have lost ad budget to bots, and 51.1 percent pointed to automated bidding algorithms as the biggest risk agents pose to their stack. Only 5.3 percent said they operate a dedicated tool to catch invalid traffic.
These figures do not add up into one dataset. OpenRouter measured token consumption on its own routing layer. Decodo measured request composition across the open web. HUMAN Security ran a separate crossover analysis with its own methodology. The marketer figures come from an unnamed survey. What they share is a direction, not a shared number: the requests hitting a site, and the compute a model burns to serve an agent, have both stopped tracking human attention on a one-to-one basis.
For a search or content team, the operational consequence is specific. Raw session counts and server-log volume now blend three different populations: human visitors, crawler and agent fetches, and the re-read cache calls that inflate an agent’s own token count without adding a new page view anywhere. A team that still reports total sessions or total crawl volume as a stand-in for audience size should segment known agent user agents out of that number before its next budget or capacity review. More than half of what a server logs today is not a person at all.
PPC Land, reported by Luis Rijo on August 25, 2026.
