What I measured

The machine-readable surface is advanced

MEASURED Fetched 2026-08-27, public endpoints, no authentication:

Per-model llms.txt at catalogue scale is ahead of most infrastructure companies. Any proposal that opens by telling this team to think about answer-engine optimisation was written by someone who didn’t look. You thought about it first, and I’d rather say so than pretend to find a gap where there isn’t one.

Usage proof already exists and isn’t crawlable

This is the cheapest high-value item I found anywhere.

MEASURED /openai/gpt-4o/apps, /activity, /providers and /uptime all return HTTP 200 with distinct content. Zero of them appear in sitemap.xml — 176 /openai model pages are listed, none of their sub-pages.

INFERRED So “which apps actually use this model” — real usage proof, the exact outcome signal a comparison table can’t carry — is already built and simply not exposed. That reframes it from a publishing programme into a sitemap change plus a template decision.

The content surface is product-generated

MEASURED First-path-segment tally of all 4,959 sitemap URLs: /compare 3,051 (61.5%), /docs 403, model pages ~600, /blog 119, /provider 103, /spawn 91, /works-with-openrouter 54, /apps 33, support/help 22.

MEASURED I opened the slugs rather than trusting the segment names, which changed the reading twice. /apps is a category taxonomy, not app pages. /works-with-openrouter is a real third-party integration directory — aider, browser-use, boltai, agent-zero, Maxim AI. /spawn is a product surface. /blog resolves to tutorials, announcements, insights — there is no case-study section.

INFERRED The gap is narrower and more specific than a raw tally suggests: 54 listings that a tool integrates, and no page describing what anyone achieved by integrating. No workload, no cost delta, no routing decision explained. The missing class isn’t third-party presence — it’s outcome.

UNKNOWN Whether templated comparison pages are more vulnerable to uncredited summarisation than experiential ones. I believe they are and I have no evidence for it. It is testable against a citation panel, and I’d test it before acting on it.

The community can be activated on demand

MEASURED #ori, 2026-08-21: 200 × $5 coupons cleared in about a minute. That single day produced 308 of the channel’s 452 lifetime messages (68%) from 73 of its 116 lifetime authors (63%). Volume fell to 6 messages/day within 24 hours, and 2/day by 08-25.

INFERRED Attention is acquirable on demand. Nothing currently converts a spike into retained usage — the readout that would tell you whether that spend was acquisition or leakage was available and never taken.

Two support surfaces failing in opposite directions

MEASURED On #community-help, an assistant replies in 0 minutes, 10 of 10 sampled threads — but 8 of those 10 never saw a human, and across aged threads roughly 6 in 10 got no human reply at all.

MEASURED On the five product-feedback forums, the inverse: zero bot messages across 83 threads, census, every thread opened. No-reply rate of 13% in each channel’s first 30 days against 77% after — using each channel as its own control.

MEASURED #providers corroborates outside that census: 52 threads, 45 of them in its launch month, then roughly twelve months of nothing, still in the sidebar.

INFERRED Because the two surfaces fail in opposite directions, no single response-rate metric can see either. A dashboard built on either one alone would report this community as healthy. That is the measurement gap, and it is the same gap that makes paid activation unaccountable.

Presence outperforms announcement

MEASURED In #announcements, normalised within year — reaction counts inflate over time, with yearly medians of 11 / 18 / 41 / 50 for 2023→2026, so the corpus-wide median of 25 is an invalid yardstick — advance event notices score 9–28 reactions. Posts announcing that something is happening right now score 92 and 101.

MEASURED #events holds six messages in its entire life, 13 reactions total, two events at zero, no threads, no recaps. MEASURED Meanwhile lu.ma/openrouter carries 12 events across SF, NYC, London and Zoom.

Positioning has drifted, and the job posting contains the drift

MEASURED Across the 853-post announcement corpus (2023-04-04 → 2026-08-26), "developers" falls from 8.4 to 1.3 mentions per 100 posts; "enterprise" first appears 2026-04-26.

QUOTED The posting describes OpenRouter as “the open AI routing and infrastructure layer that enterprises use” — then, two paragraphs later, locates growth in developers finding it, “nearly all of it starting in self-serve.”

PLATFORM FACT Discord messages sit behind authentication: not crawlable, not indexed. Whatever the community produces stays inside that boundary unless something carries it out.

Two directorships, one shared dependency, pointing one way

MEASURED Director, Channel Partnerships was published 2026-07-28 — four days after this req, at the same level, on the same remote-US terms, with the same start-it-from-nothing mandate.

QUOTED It names marketing exactly once, and names it as something it needs: “Partner with marketing to develop joint campaigns, webinars, case studies, events, and content that showcase partner solutions”.

MEASURED This posting never names that motion back. Searched across its full text, the channel vocabulary is absent outright — marketplace, reseller, distributor, integrator, GSI, hyperscaler, co-sell, alliance, indirect: zero occurrences each. Its four uses of partner mean the enterprise team, product and engineering, or the phrase thought partner; its four uses of channel mean paid media channels.

MEASURED And the marketing function it depends on is one open req of 23: Go To Market 11, Engineering 9, Operations 1, Product & Design 1, Marketing 1.