Every city keeps a census. This one counts a city that has never been counted as a city: the internet, the largest built environment in human history, in the years its population stopped being human.
We are the legacy residents of a digital city that is being rapidly repopulated by non-human actors.
The Census measures that repopulation with six fixed indicators, taken from named public sources, the same way every year. This is the baseline edition. Its job is to plant the stakes; the deltas will do the talking from here on.
Vertical axis begins at 35 percent, marked by the break at its foot, to make the four-year rise legible. The majority line sits at 50 regardless of scale.
Share of all web traffic that is automated. Four years ago, machines were 42 percent of the city. The climb since has been relentless: five points, then two, then the crossing. In 2024, automated traffic outnumbered human traffic for the first time in a decade of measurement. In 2025 the margin widened. Two consecutive years of machine majority: the repopulation is no longer a projection. It is the resident record.
Source: Imperva/Thales Bad Bot Report, annual since 2013 (2021: 42.3%; 2022: 47.4%; 2023: 49.6%; 2024: 51%; 2025: 53%). Cross-reference: Cloudflare Radar bot-share data, continuous.
Within the machine population, a new class of resident is arriving fastest: agents, software running errands on a person's behalf rather than crawling for its own purposes. Their traffic grew roughly eighty-fold in a single year. The crawler reads the city; the agent transacts in it. This is the line to watch in every future edition.
Source: HUMAN Security, State of AI Traffic. Cross-reference: Cloudflare Radar AI-crawler series; Vercel AI-crawler statistics.
How much reading has moved from human eyes to machine intermediaries. Nearly half of all searches now end on the results page: the machine reads, the person receives, the visit never happens. Where an AI answer appears, clicks to the underlying pages fall by roughly three-fifths. Fully generative search interfaces refer visitors out at 1.6 to 2.5 percent of queries, against 17 to 19 percent for the traditional page of links.
No single instrument measures machine reading directly, so this indicator is a composite of clickstream panels and impression studies, with each component named. It will remain a composite until the city builds a better meter.
Sources: Datos clickstream panel (13.1B events, analysis by iPullRank); Seer Interactive (25M impressions); Similarweb AI Mode referral analysis.
How much of the human-facing city has been torn down. The web shrank across the very period its traffic exploded: forums, blogs, the small hand-made neighborhoods built for human visitors, demolished or abandoned while construction moved to districts no person will ever walk. Traffic at an all-time high, and the streets emptying of people. The busiest ghost town ever built.
Sources: Pew Research Center link-rot study; Common Crawl page counts over time; Internet Archive coverage.
How consciously the city is rebuilding for its new majority. The publishers, who live on attention, are walling their districts at five times the rate of the open web. A smaller movement runs the opposite direction, posting signage in the new residents' native language. And the largest infrastructure operators have begun charging the new residents at the door: pay-per-crawl tollways, licensing deals negotiated upstream, in bulk, the way utilities negotiate.
The number that gets no headlines may be the most telling: 84 percent of sites have no AI policy at all. Most of the city has not yet noticed who is walking through it.
Sources: BuzzStream news-publisher robots.txt study; HasData AI Crawler Block Index (10,894 domains, July 2026); ADF adoption crawler.
Whether the city can tell its residents apart. It cannot, and both populations now have paperwork problems. On the human side, proof-of-personhood has begun: roughly 18 million people across 160 countries have had an iris scanned to certify they are not machines, and the credential is being wired into video calls, dating apps, and signature platforms. That is two-tenths of one percent of humanity holding papers. On the machine side, the fences do not hold: two in five sites that ban the best-known crawler by name are crawled anyway.
The checkpoints designed for the old city now mostly detain the residents they were built to protect. The charter for the new population is still being drafted. This is the slowest-moving indicator in the Census, and that is itself the finding.
Sources: World (World ID verified-human counts); HasData enforcement measurements; agent-identity standards tracking.
The indicators above describe the whole city. The exhibit shows one address. Each edition of the Census documents a single ordinary property, fully specified, because the aggregate studies all speak for the giants and almost nothing reports from the long tail of small sites where most of the city actually lives.
This year's specimen is a small compliance software product. Over three months of 2026, Google showed its pages 403,257 times. Those impressions produced 5,055 visits.
| Search impressions, May 1 to Jul 31, 2026 | 403,257 |
| Clicks those impressions produced | 5,055 |
| Click-through on pages ranking 4th to 6th | 0.5 to 0.9% |
| Click-through historically expected at those positions | 5 to 7% |
| Long conversational queries (8+ words), impressions | 11,223 |
| Average position on those queries | 6 |
| Clicks they produced | 51 |
Read the last three rows slowly. Eleven thousand times, someone asked a long, natural-language question and this property was among the top results. Fifty-one of them arrived. The other eleven thousand got their answer from a machine that had already read the page on their behalf. The pages are read more than ever and visited less than ever, which is the whole Census, experienced from one front porch.
The property still gets its customers. They arrive by paths the old instruments cannot see, which is its own finding: the meters of the old city measure visits, and the new residents do not visit.
The Census obeys one law: the same indicators, from the same named public sources, measured the same way, every year. Comparability over comprehensiveness. Where an indicator is a composite, its components are named and it stays a composite until a direct measurement exists. Where sources disagree, the disagreement is reported rather than averaged away.
The exhibit is one documented specimen per edition, presented as a field note, never as a statistic. Operators who wish to have a property counted in a future edition can write to the address below.
The data behind each indicator ships alongside this page in CSV and JSON, at stable addresses, legible to machine readers. A census of machine attention should be readable by it.