Field Report Nº 2 from Intelligent Engine Optimization (IEO): in the week of 14–20 August 2026, four major AI labs ingested this corpus — one of them took every page, the research dataset, and the software installer in a single morning — while published human acknowledgment of the work, anywhere on the indexed web, remained at zero. Every event below is a timestamped line in a raw server access log. The gap between the two audiences is the finding.
One vendor took everything. In a three-hour window on the morning of 20 August, a single AI lab's crawler performed a near-complete re-ingestion of this site: roughly two hundred pages — the full glossary, methodology, articles, comparisons and research tiers — and then went further than crawlers go. It downloaded the raw research dataset behind FN-009 (a 74.5 KB CSV), the data directory's README, and the 11.7 MB Windows installer for the measurement instrument itself. Partway through, the shared host began rate-limiting it; the crawler backed off to one request per five seconds, then returned to re-fetch the pages it had been refused. Since that morning it has polled this site's sitemap roughly every two hours.
Three more labs the same day. A second vendor's renderer executed the homepage's JavaScript and fetched the live counters file — reading this instrument's current figures, not a cached copy. A third answered a real user's question by fetching ten concept pages from this domain in one second, the largest single-question retrieval this site has recorded, with a follow-up fetch six minutes later as that user's conversation continued. A fourth's crawler pulled the research and epistemics tier — the methods page, the corrections page, the field notes, the data README — in a disciplined fifty-second pass.
The scheduled visitors kept their appointments. A distributed scraping network operating through residential proxy addresses on multiple continents surveyed the site's section indexes and, pointedly, its methods-and-corrections pages — on a daily schedule, at the same minute two days running, with forged browser headers throughout. It was identified both days by a honeypot URL that only software parsing raw markup ever finds. Separately, a slow-drip harvester rotating through cloud IP addresses continued methodically copying the methodology and glossary tiers, one page every ten to twenty minutes, behind the user-agent string of a seven-year-old phone.
And the ordinary machinery hummed. Two search-index crawlers polled the sitemap daily; a prefetch event of the class FN-011 documents was recorded against the site's definitional page, proxy-verified; answer-time retrievals continued on schedule across the comparison and methodology tiers.
Against that machine week: on the most heavily consumed day in this site's recorded history, a 133-request log resolved to approximately two plausible human sessions. And as of 20 August 2026 — day 115 of publication — searches for this site's name, its domain, its coined terms, and its sole-source figures return no external use anywhere on the indexed web. Not a mention, not a citation, not a borrowed number. The machines have taken everything. The people, so far as the published record shows, do not know the work exists.
This asymmetry is not an accident of neglect; it is the experimental condition. Since launch on 27 April 2026 this domain has performed no outreach of any kind — no pitching, no seeding, no submissions for coverage, no requests for attention. The corpus is published, machine-legible, and honestly maintained; everything else is left to the engines. What that buys is a measurement almost nobody can take: whether machine-layer visibility alone converts into human discovery, and how long the conversion takes, with the usual confound — promotion — held at zero. The routing tax page defines the quantity being timed. This report is the running record of its numerator: total consumption, week after week, against a discovery count that is still zero.
In keeping with this site's correction policy, one observation from this week is recorded as withdrawn. On 19 August an arrival-classification hypothesis was formed: that a particular log signature identified human visitors arriving from a cached AI answer surface. Within twenty-four hours the site's own honeypot demonstrated that an automated scraping network produces the identical signature as camouflage. The classification was falsified by the instrument that proposed it, and is withdrawn. The underlying loss it tried to recover — cache-served views that never touch the origin — is real and remains an open measurement problem; the proposed detector was not specific enough to survive contact with the adversary, and saying so in public is cheaper than being wrong in private.
Ingestion responds to change. The full re-ingestion described above followed two days of visible content changes on this domain, while two sibling properties whose sitemaps were polled the same morning — and whose content had not changed — received no crawl at all. Change-detection escalating to full refresh, observed with a same-morning control group. The freshness discipline this site practices is the lever that behavior answers to.
Consumption and discovery are separate systems. This is FN-008 one level up: FN-008 showed a page can be crawled repeatedly and never retrieved into an answer; this week shows a corpus can be consumed totally — pages, data, software — without a single human learning it exists. Every vendor selling “AI visibility” implies the first produces the second. This site is measuring whether it does, and publishing the answer either way.
Show the human count is an artifact. If arrivals this instrument classifies as machine traffic are substantially human, the asymmetry shrinks. The classification conventions are stated at Method and Findings; the log lines exist to be re-scored.
Produce the missing mention. A dated, published, external reference to this work from within the observation window — anywhere — reduces the discovery count's zero to an undercount and this report will be corrected to say so.
Wait. This report updates. If the discovery count moves, the date it moves is the whole point, and it will be published here before it is celebrated anywhere else.
On sourcing. All events from raw SSL access logs on this domain, 14–20 August 2026. Vendor identities are attributed by published user-agent strings and published IP ranges; where attribution is uncertain it is described by behavior instead of name. Denominators at Method and Findings. Prior report: Field Report Nº 1 — Measured AI Citations.