This runbook says what to work on and, more importantly, what not to. It exists because the obvious SEO instincts are wrong for this site in ways only the measured data reveals.
The mechanism behind it is ContentIntelligenceDesign. The repository copy is docs/agents/content-feedback-loop.md.
Measured 2026-08-16 from the visibility fact store and Google Search Console.
| Fact | Value | Meaning |
|---|---|---|
| Indexed by Google | 988 | Discovery works; sitemap and robots are fine |
| Not indexed | 389 | 373 of them are Google's own choice |
| Not indexed, our fault | 16 | 9 404s, 3 robots.txt, 4 canonical. That is all |
| Google impressions past position 20 | 2,154 of 2,351 (92%) | Almost all Google exposure sits where nobody looks |
| Bing pages on page one | 54 of 54 (100%) | Everything Bing indexes, it ranks well |
| Click-through, Google against Bing | 0.32% against 1.85% | Bing converts about six times better per impression |
| Human search sessions in 90 days | about 3 | All self-generated |
| Orphaned pages | 159 | 11.6% of the corpus has no inbound internal links |
| Broken links | 79 | Wasted crawl budget on a rationed crawl budget |
The single most important inference: the constraint is authority, not content quality, not snippets, and not discovery. Google indexes our pages and declines to rank them, and it has stopped indexing new ones. Both are the same signal.
Do not rewrite titles or descriptions for low-ranking pages. A page at position 30 gets no clicks whatever its title. One page carries 154 impressions and zero clicks at position 29.7 — that is a ranking failure, not a snippet failure. Only treat click-through as the problem at position 10 or better.
Do not bulk-publish right now. 293 pages are already discovered and declined. New pages queue behind the same authority ceiling and dilute crawl budget across a larger corpus without improving any of it. This inverts the natural instinct of a content pipeline and is the most expensive mistake available here.
Do not act on dashboard opportunity scores uncritically. The upstream detector scores uplift assuming a page could reach position three. Applied to 993 impressions at position 61, that produced a confident twenty-five-clicks-available figure that is almost entirely fictional. Check position first.
Do not trust indexed-page counts from the visibility pipeline. The Google provider reads a field deprecated years ago that now always returns zero. It reported zero indexed while Search Console showed 988. Treat that number as absent, not as zero.
First, the internal link graph. 159 orphans and 79 broken links. An orphan has no inbound links, so Google receives no importance signal, so it declines to index — exactly the observed pattern. This is the only authority lever available without outreach and is entirely within our control.
Second, consolidate rather than add. Merging two weak pages into one strong page improves the corpus on every axis that currently matters. Prefer depth over breadth until the indexed count moves again.
Third, Bing is the real opportunity. It ranks everything it indexes on page one and produces only 108 impressions, which is a coverage problem rather than a ranking one. It also responds to IndexNow, which this wiki already pushes and which Google does not accept.
Fourth, agent retrieval quality, which is the only feedback loop with real traffic today.
Human search traffic is approximately zero: twelve query-log rows in ninety days, all self-generated. Any loop built on human behaviour has no denominator and will manufacture confident statistics from nearly no samples.
The live audience is agents — MCP callers, bundle consumers, and AI crawlers. Two consequences follow. The agent-gap signal is the only rule with a live denominator: a retrieval returning zero sections, or a bundle with weak or unknown coverage, is a real observed gap with a real consumer behind it, and outranks any search-engine opportunity. And ranking in Bing is ranking in the answer layer that feeds AI assistants, which is where this wiki's actual readers are.
Call the backlog before deciding what to work on rather than picking pages by intuition. Read the evidence block rather than trusting the priority score, since the numbers are exposed so a suggestion can be checked rather than obeyed. Weight uncalibrated rule types down: until a type has twenty evaluated outcomes its weight is a guess, and the tool says which are guesses.
Record every change with its baseline, because an unrecorded change cannot be evaluated and an unevaluated change teaches nothing. Respect the sixty-day per-page cooldown, since two changes inside one measurement window make the effect unattributable. Snooze with a reason when declining, because a declined suggestion with no recorded reason is indistinguishable from a bug six months later.
Effect measurement is a weak quasi-experiment: one site, no control group, no randomisation. A single improved verdict is suggestive rather than conclusive, and insufficient-data will be the majority verdict for a long time. Value accrues in aggregate across many changes of one type.
Never reason from a single snapshot. An early read of one window showed Google delivering zero clicks; the trend showed that window was unrepresentative.
A parsed number is not a measurement. The zero-indexed-pages bug returned a structurally valid value from a successful API call and was entirely meaningless. When a number surprises you, check it against the source system before building on it.
Silence is a valid output. Every rule has a minimum-support threshold and declines to fire below it. An empty backlog means insufficient evidence, not nothing to do.