Methods & sources

How we count humans and machines, where every number comes from, and what we can't know.

What counts as human, what counts as machine

• Human: traffic a person causes in real time: page views, app use, video and music streams, calls, their own uploads.
• Machine: traffic with no person in the loop at that moment: crawlers, scrapers, good and bad bots, AI training and AI-agent fetches, API automation, IoT/M2M telemetry, server-to-server replication.
• Borderline, counted as human: software updates, backups and CDN cache fill. Vendors can't split them out.
• AI agents acting for a person count as machine. This is the choice that matters most for the "1 in 1,000" question.

Requests are not bytes

Bot detection works per HTTP request. A bot page fetch is a few kilobytes; a person watching video moves gigabytes. So by requests machines are already the majority (≈51–53%), while by bytes people still dominate (machines ≈9%, modelled). We always show both and never mix them.

Same basis for both: bytes, plus requests

The main chart compares human bytes and machine bytes in exabytes per month. A linked view shows request counts and bytes per request. Request counts start in 2012, the first year with a measured bot share.

• Global requests = Cloudflare's 81 million HTTP requests/s (2025) ÷ an assumed Cloudflare share of all HTTP requests of 15–35% (mid 22%). Cloudflare does not publish its share, so this is the biggest assumption. That gives 609 trillion–1.42 quadrillion requests per month (mid 968 trillion).
• Earlier years are back-cast from total bytes, assuming bytes per request grew 5–11% a year (mid 8%). HTTP Archive's median desktop page grew from 1,208 KB (2014) to 2,652 KB (2024), at about 70 requests per page.
• Bytes per request = modelled bytes ÷ modelled requests, for each group. In 2025 a human request averages about 1.50 MB and a machine request about 85 KB.

How we model machine bytes (never measured)

No one publishes machine bytes, so we build the figure two ways and use the widest range.

• Bottom-up: 0.9%–10.0% of bytes (mid 2.9%).
• Top-down: Cisco's M2M share of IP traffic (~3% in 2017 → ~6% forecast for 2022), trended to 2025 and with crawler bytes added: 5–14% (mid 9%).
• Used: 0.9%–14.0%, mid 5.9% (the average of the two mids). The forecast samples across this whole range.
• Not added, to avoid double counting: API traffic (bot detection already labels unattended API calls as bots) and backups, updates and CDN fill (counted as human by our definition). CDN-observed automated bytes aren't published. They are a phase 2 cross-check.

The Moore's-law scenario

The prediction says the machine-to-human ratio compounds, almost like Moore's law. We test that directly: ratio(2031) = ratio(today) × 2^60/N, for a doubling time of N months. By requests, today's ratio is about 1 : 1, so 1 in 1,000 human (999 : 1) needs ~10 doublings in five years: a doubling every ~6 months. By bytes, the ratio is about 0.06 : 1, so it needs ~14 doublings: one every ~4 months. A classic Moore doubling (18–24 months) gives a ratio of about 6–10 : 1 by requests, with humans at 9–15%. The measured request ratio doubled about every 7 years from 2019 to 2025. The prediction therefore needs a step change. Agents talking to agents is the most plausible driver.

Agents talking to agents (A2A scenario)

History alone can't see what comes next: AI agents calling tools and other agents, which call more agents. So the forecast has a separate, fully parameterised agent-to-agent model. Visible A2A calls per month = active agents × tasks per agent per day × (fan-out + fan-out² + … to the chosen depth) × a capped network-effect term (agents/agents₂₀₂₅)^α × 30.4 × the share of calls that cross the public internet. Bytes = calls × bytes per call, which grows as agents fetch media and data. Only growth above 2025 is added, because 2025's agent traffic is already in the measured bot share. Every input has a low/mid/high range (site.json → a2a.params), and you can change them all with the sliders on the front page.

What we can't see: a large share of agent-to-agent traffic runs inside clouds, as east-west traffic between services and private APIs. It never crosses the public internet, so Cloudflare Radar, Imperva and similar measurements never see it. We show it as "modelled hidden" agent traffic. It loads data centres, but it is not "the internet" that people and websites share.

Infrastructure impact

Costs tied to requests include per-request CDN and API fees, WAF and bot management, edge and serverless compute, origin CPU, database queries, TLS and logging. Machines drive about half of these, and their requests grow ~20%/yr. Costs tied to bytes include egress, transit, cache and video delivery. People still drive ~94% of these, growing ~15%/yr. The "machine bytes overtake human bytes" forecast runs the same 20,000 simulated futures to 2060. After 2031 it is an extrapolation, not a forecast.

"Live" and "real time"

No public source measures all internet traffic in real time. The live counters are modelled: a cited annual total, spread evenly over the year and grown at the forecast rate. The human/bot split comes from the latest daily Cloudflare Radar reading (last 24 hours). Every widget shows when its data was last updated.

The forecast ("weather forecast")

Five models, 4,000 simulations each, equal weights. The shaded bands are the middle 50% and 90% of outcomes. Once daily Radar data has two weeks of history, the starting point moves with it.

Caveats

• Imperva and Cloudflare see their own customers' sites, not the whole internet. Neither sees private networks, streaming CDNs or IoT backhaul.
• The jump between 2016 and 2017 is a change of method (Incapsula sample to Imperva Bad Bot Report), not a real collapse.
• Before 1990 volumes, and before 2012 bot shares, are order-of-magnitude models. No measurements exist. They are drawn dashed.
• 1990–2000 volumes are US backbone traffic (Odlyzko), used as a proxy for the world. Cisco's figures include managed IP video.
• Machine share of bytes is modelled (M2M ≈3–6% of IP traffic per Cisco, plus crawler bytes).
• Licence: Cloudflare Radar data is CC BY-NC 4.0, attributed to Cloudflare Radar (radar.cloudflare.com). This site is non-commercial and must stay so, unless Cloudflare gives permission.
• Figures marked ⚠ below were taken from the publisher's headline and could not be re-read on 2 Nov 2026. We will re-check them.

Sources

Every data point names one of these. Full table: history.csv (columns kind, confidence, source_id, source_url).