I went looking for gaps and kept finding that the gap was in my own method. Each of these was a working hypothesis I would have shipped if I hadn’t checked it. They’re listed here because the claims that survived are only worth believing if you can see which ones didn’t.
1. “Their answer-engine surface is unowned”
Wrong. MEASURED llms-full.txt is 3.6 MB, /docs/llms.txt exists, and
per-model llms.txt is served at catalogue scale. This was the headline of an earlier draft.
It was the first thing I tested and the first thing that failed.
2. “The docs miss the failure modes users actually hit”
Wrong. MEASURED In llms-full.txt: BYOK 391, credits 198, 403 153,
429 113, deprecat* 85, rate limit 69, troubleshoot 16.
I then ran an absence guard on the one error string I still expected to be missing — and found it absent from the community corpus too. There was no gap in either direction.
UNKNOWN A caveat against my own test: the community corpus I could search was
#general and #announcements, not #community-help and the product forums. That comparison is
underpowered as well as negative, and I wouldn’t rely on it without redoing it properly.
3. “The model pages have no call to action”
Wrong, and wrong for an instructive reason. MEASURED The anonymous model page
carries Try this model, Playground and Compare. I had searched for Sign in, Sign up,
Create API Key, Get started and Buy credits — five variants, which is the right discipline,
and all five my vocabulary rather than yours.
INFERRED The surviving observation is narrower and is a question rather than a finding: the anonymous model page offers exploration, with account creation deferred behind it. A playground-first funnel is a legitimate strategy that may well out-convert an early wall. It’s the first thing I’d A/B test on the highest-traffic template, not something I’d call a defect.
4. “There are 33 builder pages”
Wrong. MEASURED /apps is a category taxonomy — apps/category/coding,
apps/category/creative/video-gen. I’d read a segment name instead of opening a slug. Correcting
it is what surfaced the finding that per-model app data exists and isn’t in the sitemap, which
turned out to be the most useful thing in this document.
5. “Five scattered one-off events, no cadence” and “no public events surface”
Both wrong. MEASURED The OpenRouter Show ran at least four episodes
(2025-11-07 → 2026-02-06) with named guests — I missed it because the posts say “join us later
today” and never “webinar.” MEASURED And lu.ma/openrouter carries 12 events
across SF, NYC, London and Zoom.
6. “The community corrected the record about the acquisition itself”
Wrong, and this one had already reached a draft of this proposal. I watched a member push back on outlets reporting the deal as finalised, and wrote that the community did the correction work itself.
MEASURED A second pass over the same corpus killed it. subject to, customary,
regulatory, antitrust, approval and definitive each occur zero times; 3 of ~206
acquisition messages carry an external link; correction-or-status-hedge is 4 messages from 2
authors.
INFERRED One person pushing back on a headline is not a community correcting itself. The replacement is the better finding and it’s the one now in Context for the record: the correction was available, load-bearing, and never made. I had generalised from a single thing I watched happen, which is a different failure from the absence claims above and a harder one to catch — nothing was missing, so nothing prompted me to look again.
The pattern, stated plainly
Four of these six were absence claims. Two of those four failed the same way: I searched for a thing using my own vocabulary, didn’t find it, and read the silence as neglect. The other two failed more plainly — I was simply wrong about what existed. The sixth failed in the opposite direction: I saw one thing happen and wrote it as though the channel had done it.
The finding that survived — per-model sub-pages rendering but missing from the sitemap — came from the opposite move. Not “does this exist?” but “this exists; is it reachable?” Reachability is binary and checkable. Absence almost never is.
That’s the discipline I’d bring to a measurement function, and it’s the reason the numbers elsewhere in this proposal are worth the time it takes to check them. I’d rather hand you six corrections than one confident story.