Publishing for machines that copy. A field manual for the web after search, built on a measurement rather than a theory.
Read it: https://nanobotco.github.io/uptake/ · plain text · Markdown · JSON Lines
Fourteen days, fifty-two repositories on one GitHub account, no promotion of any kind:
| clone operations | 961 |
| distinct machines that cloned | 501 |
| GitHub page views | 36 |
| distinct browsers that viewed | 20 |
| cloners per viewer | 25.1 |
| repositories with five or more cloners and zero viewers | 23 of 52 |
On 16 September 2026 seven directories went public in one evening. That day drew 617 clone operations from 267 distinct machines — the twelve days before it total 293 between them. The largest of the seven was created at 04:54 UTC and cloned 130 times by 69 machines before the day was out, with one referrer and four human page views. Nothing linked to it. It was hours old.
A new repository is not a document somebody has to find. It is a row in a public event stream that anyone can subscribe to, and the subscribers arrive within the hour.
- The number
- What a clone is, and what a click was
- The measurement, and what it cannot say
- Day zero
- Three bots wearing one coat
- The objective function moved
- The artifact, not the page
- The hallway
- Provenance is the product
- The licence is the only thing that travels
- What to count now
- What does not work
- The tactics, in order of cost
- Posterity
- What would falsify this
Plus a colophon on why the thing is shaped this way, and 54 sources.
Everything the manual cites is in data/, with the collection method and the caveats
written into the files rather than into a caption.
| file | what it is |
|---|---|
data/repo-traffic-2026-09-18.json |
one record per repository, with totals and caveats |
data/repo-traffic-2026-09-18.csv |
the same rows, flat |
data/account-daily-2026-09-18.json |
the daily series behind the chart |
data/account-daily-2026-09-18.csv |
the same, flat |
data/sources.json |
all 54 citations, keyed to the numbers in the text |
data/croissant.json |
the dataset as MLCommons Croissant, loadable by a pipeline |
data/raw/ |
what the GitHub endpoints returned, untouched |
The window is a rolling fourteen days because that is all GitHub retains. Uniques are counted by IP. The endpoints carry no user agent. Section 3 of the manual states what those limits rule out, and section 15 lists the control experiment that has not been run.
llms.txt |
every page and file, one line each |
llms-full.txt |
the manual and the measurement, flattened |
corpus.jsonl |
one JSON object per section, with its citation ids and attribution string |
robots.txt |
38 crawlers allowed by name, with a Content-Signal line |
sitemap.xml |
pages and images |
CITATION.cff |
how to name this |
for-agents/ |
the terms, in prose |
The published page carries JSON-LD for ScholarlyArticle, Dataset, FAQPage, HowTo and
WebSite, with all 54 citations in the article's citation array.
Four, drawn by tools/figures.py from the data in data/, in four formats. The SVG keeps
its labels as text, so a machine reading the file gets 617 clones from 267 distinct machines as a string rather than as pixels; the PNG and JPG are rendered from that same SVG
by the build, so raster and vector cannot drift apart.
| figure 1 | unique cloners against unique page viewers — SVG, PNG |
| figure 2 | clones per day, and the spike — SVG, PNG |
| figure 3 | the seven files an arriving agent can pick up — SVG, JPG |
| figure 4 | the same fourteen days, revealed one at a time — GIF |
Five public-domain pictures sit alongside them, each with a sidecar JSON file naming its Commons page, its creator where one is recorded, and its licence statement.
Text, figures and data: CC BY-SA 4.0 — attribution and share-alike. Tools: MIT. Pictures: public domain, not relicensed here. Full terms in LICENSE and NOTICE.txt.
Cite it as:
NaNoBotCo, Uptake: publishing for machines that copy (2026),
https://nanobotco.github.io/uptake/
Two asks for a copy: name the source, and pass the same terms on. Share-alike binds the visible reuse — the fork, the republished dataset, the derivative directory. Whether it reaches a model's weights is unsettled law, and the manual says so rather than pretending otherwise.
python3 tools/fetch_images.py # public-domain pictures, with their sidecars
python3 tools/figures.py # four figures, four formats, from data/
python3 tools/build.py # docs/ — page, plain text, corpus, llms.txt, robots, sitemap, feed
Standard library only, except Pillow for the animated frames and rsvg-convert for the
raster passes.
The finding is one account over one fortnight. Contradicting data is worth more here than agreement — issues and pull requests are open.
Contact: Nan · nan@motdang.net · Sponsor: Ko-fi · Patreon
- Mot Dang — city directory for Chiang Mai and Chiang Rai
- wichaa — Lanna manuscripts, the amulet market, and the traditions around them
- Amulet Atlas — amulets, charms and talismans worldwide
- Carolina Barbecue — barbecue in North and South Carolina
- Wing Country — the American chicken wing
- Pink Box — the American mom-and-pop donut shop
- Basque Tables — Basque dining rooms of California, Nevada and Idaho
- Pinot Country — pinot noir: the vine, the regions, the cellars
- Care Abroad — treatment across borders, with published prices and their dates
- Thai Roots — a root dictionary of Thai, with a word decomposer
- The index — every corpus, site and repository, counted
- NaNoBotCo — the portal
- ฮักฝรั่ง — เรื่องเงิน วีซ่า และชีวิตกับแฟนฝรั่ง
- Offrampt — turning crypto into spendable local money, Thailand first
All of it, counted: https://nanobotco.github.io/index/ · roster as JSON: https://nanobotco.github.io/index/fleet.json