stats

GitHub Search API, /search/commits, one query per tool per week using author-date ranges. The count is GitHub's total_count, which fluctuates between identical queries; runs since September 2026 sample each week more than once and keep the median. Every row of every CSV carries the GitHub search URL that produced it.

• Public repositories only. Private and enterprise usage is invisible.
• Detection is uneven: Claude Code adds a trailer by default, Copilot inline completions add nothing, Cursor and Codex changed their trailer behaviour during the window.
• GitHub total_count is approximate and drifts for weeks after the fact, so the latest run overrides earlier runs for the same week; the per-run values are kept in the data file.

Runs: 2026-03-27 (312 rows), 2026-09-27 (312 rows), 2026-11-01 (374 rows).

GitHub Search API, /search/commits, with only an author-date range as the query. This is the denominator for the AI share chart.

• GitHub counts here spike unpredictably (early March 2026 was roughly double the norm), so the share series inherits that noise.
• From w/c 10 Aug 2026 to w/c 5 Oct 2026 the Search API returned two to ten times the surrounding weeks (596M for w/c 21 Sep 2026 against about 55M either side). The level may be accurate. GitHub's Octoverse 2026 report (28 Oct 2026) counts 34.3 billion commits from Oct 2025 to Sep 2026 against 12.9 billion the year before, says the rise accelerated sharply after June, and puts September 2026 at about six times September 2025; that report covers public and private repositories, while this series is public only. The sampling is the weaker part: repeated queries of the same week on 1 Nov 2026 differed by more than 20% (w/c 5 Oct 2026: 451M, 476M, 577M), so a single query is not a stable reading. From Nov 2026 each week is queried three times and the median kept, and earlier weeks are re-sampled on later runs.

Runs: 2026-03-28 (25 rows), 2026-11-01 (31 rows).

GitHub Search API, /search/issues with is:pr and a created date range per calendar month.

• Author-account queries only see bots that open PRs under their own account; Claude Code PRs are opened by the human, so the body marker is used instead.
• The chatgpt-codex-connector bot account has authored almost no pull requests since March 2026 (three in August, one in October), so the Codex line here no longer reflects Codex use; PRs made with Codex now appear under the user's own account with no marker this series can match.
• The sweep-ai bot account no longer exists on GitHub (the Search API returns a validation error for it), so its rows stop at the last successful run and it is not re-measured.

Runs: 2026-03-28 (95 rows), 2026-11-01 (211 rows).

GitHub Search API, /search/issues with is:pr and an updated date range per calendar month. A PR is counted in every month it was updated, so monthly figures are not strictly additive.

• The post's six-month totals were single queries over the whole window; the monthly rows here come from separate queries and do not sum to exactly those totals.
• The chart shows one line per bot, from its commenter query; the reviewed-by and text-marker variants of the same bots stay in the table.
• The codiumai-pr-agent-pro and sweep-ai bot accounts no longer exist on GitHub (the Search API returns a validation error for them), so their rows stop at the last successful run and they are not re-measured.

Runs: 2026-03-28 (223 rows), 2026-11-01 (385 rows).

GitHub Search API, /search/issues with is:issue and an updated date range per calendar month.

• GitHub Actions, Dependabot and Renovate are kept as context rows and left out of the chart.

Runs: 2026-03-28 (217 rows), 2026-11-01 (248 rows).

Chrome Platform Status feature popularity timeline for the ServiceWorkerControlledPage use counter (https://chromestatus.com/metrics/feature/timeline/popularity/990), averaged per week. This measures page loads in Chrome, not sites; the Web Almanac figure in the post measures sites in the HTTP Archive crawl.

• Chrome-only, page-load weighted, so a few heavily visited sites dominate.
• Different denominator from the Web Almanac figure the post cites; the two are shown side by side, not compared.

Runs: 2026-11-01 (56 rows).

Monthly origin counts from the HTTP Archive tech report (https://httparchive.org/reports/techreport/tech?tech=React), fetched by tools/stats/collect-httparchive.py from cdn.httparchive.org/v1/adoption, the endpoint the report itself reads. Each crawl is dated the first of its month; the full history is re-fetched on every run so later corrections to the crawl show up as revisions.

• The crawl covers roughly 12-16 million origins with enough traffic to be in the Chrome UX Report; parked and unused domains are not in it, which is the main reason it is far below BuiltWith.
• The top-1m figure in the essay was read off the report's chart in October 2025; the API gives 113k mobile and 106k desktop for that month, so the 140k as written was closer to the desktop and mobile series combined than to either one.

Runs: 2026-11-01 (132 rows).

Hand-recorded readings of https://trends.builtwith.com/javascript/React (Site Totals and Traffic). The page sits behind an image puzzle, so each month open it in a browser, copy the page text and run `python3 tools/stats/manual-add.py react-builtwith --paste`, which extracts the live-sites and top-1m figures and rebuilds; other figures can be added by name.

• BuiltWith counts hundreds of millions of root domains, including parked and inactive sites, so its totals are not comparable with the HTTP Archive or W3Techs series, only with earlier BuiltWith readings.
• Between October 2025 and November 2026 BuiltWith moved about 42 million React sites from "live" to "expired" (expired went from 33.7M to 76.2M while live and expired together grew from 87.4M to 107.9M), so the live count fell 42% without React losing sites. The top-1m and top-100k figures are the ones to compare over time.
• The October 2025 rows come from the Wayback Machine's copy of the page; later readings are pasted from the live page because it sits behind a bot check.

Runs: 2026-11-01 (8 rows).

tools/stats/collect-w3techs.py reads the monthly JavaScript-library table at https://w3techs.com/technologies/history_overview/javascript_library/all: the share of websites in the W3Techs survey (the top ten million by traffic) that use each library, as read on the 1st of each month and stored against the month that had just ended. It is the automated stand-in for the BuiltWith figures, which can only be read in a browser.

• A share of a traffic-ranked sample, not a count of all domains, so it is not comparable with BuiltWith or the HTTP Archive origin counts, only with itself.
• W3Techs shows one trailing year per page; each run re-fetches the whole year.
• Only libraries that appear in the W3Techs table can be tracked; a tool missing from the page is reported on stderr and skipped.

Runs: 2026-11-01 (39 rows).

tools/stats/collect-openrouter.py reads the weekly chart behind https://openrouter.ai/rankings (model-rankings-chart, view=week): the top ten models plus an Others bucket for every Monday-to-Sunday week in the trailing year, summed to a weekly total. The programming estimate multiplies the complete weeks inside the trailing 30 days by the code share of tokens from the task breakdown on the same page.

• OpenRouter removed the per-category model rankings (rankings/programming) during 2026. The programming figures in the essays were read from that ranking and cannot be re-measured the same way; the estimate here uses OpenRouter's task classifier instead, which counts code tasks from any app rather than tokens from programming apps.
• The chart's first week (w/c 27 October 2025) only has a day or two of data, so the series starts the week after.
• The chart only reaches back one year, so each run re-fetches the full window and earlier weeks are revised if OpenRouter restates them.
• OpenRouter traffic is one router's view of the market, not total LLM usage.

Runs: 2026-11-01 (55 rows).

tools/stats/collect-openrouter.py reads the task breakdown behind https://openrouter.ai/rankings (task-spend): OpenRouter classifies requests into task types and reports each macro-category's share of tokens and of spend over the trailing 30 days. One row per monthly run.

• Only the trailing 30 days are published, so the history here starts with the first run (November 2026).
• Classification is OpenRouter's; a coding agent's planning calls land under Agent, not Code.

Runs: 2026-11-01 (8 rows).

Hand-recorded readings of the OpenRouter programming ranking (rankings/programming, later rankings?category=programming) at the time each essay was written: tokens, model cost from the pricing of each model, and the input share of tokens. The page no longer exists, so no new rows can be added the same way; keep it for the as-written figures.

• OpenRouter is one gateway; the number is a lower bound on programming traffic, not a market total.
• Model cost uses list prices at the time of the snapshot.
• OpenRouter removed the per-category model ranking these figures came from during 2026; the weekly series above (Tokens routed through OpenRouter) is the automated replacement.

Runs: 2026-03-27 (8 rows).