Descriptions finished (36, not 17), and the R-queue restocked from live GSC

[pftg]

Two queued items, both finished. Each turned up a measurement error worth more than the task itself.

1. The last truncated descriptions - 36, not 17

The 17 I'd left were the ones a script should not guess: announcement posts and link roundups with no usable opening sentence. Written by hand after reading each post. The award posts now claim only what the post supports - "Techreviewer included JetThoughts in its 2024 list", not the body's "thrilled to announce its inclusion in Techreviewer's prestigious list".

Then 19 more turned up that every prior pass missed, including my own "234 → 17" figure. Their ... sits inside single quotes:

description: 'TL;DR: ... use defined?(@_result)...'

/\.\.\.$/ never matches that - the trailing quote is the last character, not the dot. They were invisible to every source-level count I ran.

I only found them by grepping the rendered meta tags instead of the source, which is the rule this repo already states: a source-level validator cannot see rendered output. Verified after the fix, site-wide and rendered:

Surface	Ending in ...
<meta name="description">	0
<meta property="og:description">	0
<meta name="twitter:description">	0

Source is 0 too, under a pattern that now catches every quoting style. All 36 carry seo_override, so the 10-minute sync cron cannot revert them.

2. R-queue restocked from live GSC (§13)

The R-queue is exhausted. §13 replaces it, and every row carries the live 90-day number it rests on.

The audit changed the conclusions. Two traps, both documented in §13a because either produces a confident wrong answer:

Page impressions are mostly queries GSC won't name. rails-virtual-attributes reports 7,902 impressions / position 7.8 / 0.09% CTR at page level. Its per-query breakdown totals 265 impressions. On the queries GSC does name, it sits at position 3.7-5.9 with 4.26% CTR on the head term. The alarming page-level CTR is an aggregation artifact - a snippet rewrite there would chase a number, not a reader.

One page contaminates every site-wide average. elital's upwork login draws 46,218 impressions at 0.01% CTR - roughly half the property - for a navigational query the site has no business ranking for.

Corrections to two claims, one of them mine. §4's F1 row reads pos 12.8 / 8,832 impr; live is 6,373 impressions at position 10.6, so the row's premise holds. My earlier "1 impression at position 16" was the wrong slug - that's a different propshaft post (288 impr / pos 16.9). Two exist.

The restocked rows:

#	Candidate	Verdict
N1	Falcon post snippet rewrite	Highest confidence, and an upgrade not a new post. 804 named impressions on-topic, ruby falcon pos 5.4, falcon ruby pos 4.8, 5 clicks total. Position ~5 should return ~8%; this returns 0.62%. The ranking is won; only the snippet fails.
N2	Narrow single-problem posts (a format, not a topic)	rails install dependencies 15.5% CTR, tailwind triangle 7.3% vs the long guides at 0.08-0.16% at comparable positions. 20-50x.
N3	LangChain Ruby	Upgrade, don't add - R2 just shipped into that space.
N4	falcon vs puma	DO NOT WRITE. 1 impression. Logged so nobody re-proposes it from intuition.
N5	fractional cto services at position 83.3	Flagged - commercial page, not a blog slot.
N6	elital upwork login	Flagged for a noindex decision.

§13c's recommendation: fix what already ranks before publishing more. Named-query clicks across the whole property are in the low hundreds per quarter, while pages sit unclicked at positions 3-8.

3. OKF sync

analytics-access gains the page-vs-named-query trap, recorded with its inverse so the rule is usable rather than just cautious - the Falcon post's 804 on-topic impressions at position 5 for 5 clicks IS a real snippet failure. Same shape of number, opposite verdict; only the comparison separates them.

Gates

Content-only and docs: bin/hugo-build 8/8 validators, marketing_copy_test 3 runs / 0 failures, okf validate --strict conformant. The blog diff is frontmatter-only - 0 lines changed outside description/seo_override - so the visual suites correctly do not apply.

Not done: the agent 4-eyes pre-commit review, because this session runs under a standing instruction not to spawn agents. Say the word and I'll run a reviewer over the diff before merge.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: fd461bc8-880f-4802-9ba5-dd5bfc8e08cc

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[pftg]

Review verdicts (4-eyes gate)

Two independent reviewers, distinct lenses, briefed with the diff and the standards but not my conclusions. Both found real defects. Fixes in 628163eec and 99f972217.

Data verification — 6 claims failed

The serious one: N1, the row I labelled "highest confidence", was wrong in every part and its inference was backwards.

claimed	804 named impressions, 5 clicks, 0.62% CTR, "the named queries ARE the volume"
actual	920 named across 27 queries; the page takes 37 clicks at 1.03% and is the site's best blog earner

804 came from a row_limit=20 call — get_search_by_page_query truncates at 20 rows and its totals field sums only the rows returned. The "5 clicks / 0.62%" were named-query totals quoted as page performance: a denominator splice, in the section written to warn about denominator splices.

Applying §13a's own test kills the row — named 920 vs page 3,590 = 26%, so it is the same artifact class as rails-virtual-attributes, not its inverse. N1 downgraded to a hypothesis; §13c now rests on N2.

Also corrected: "roughly half the property" → 13.4% (this disagreed 4x with a figure already in the OKF bundle); N6 mixed a query row's impressions with a page row's clicks; "low hundreds" of named clicks → 97; N2's "20-50x" compared query-CTR against page-CTR and was really 24–194x, replaced with a like-for-like page-vs-page comparison (~52x at near-identical position).

Confirmed unchanged: trap #1, the F1 propshaft correction, N2's raw figures, N3, N4, N5.

Cold-eyes copy — 12 descriptions fixed

Blocking: from-chaos-flow opened "A founder running three remote teams was drowning in context switching" — that is Sarah, the persona the claims canon records Paul ordering removed on 2026-08-20 ("remove she is not real"). A SERP snippet asserts to a stranger that we had this client. Rewritten to the mechanism.

Separate issue, not fixed here: that post still carries 7 "Sarah" references in its body. Same fabricated persona; needs queueing.

Five more claims the posts don't support — invented Techreviewer provenance; a stack post whose thesis was inverted; a promised monolith-splitting threshold that isn't in the post; a "free-tier" promise that stopped being true when Heroku ended free dynos in Nov 2022; and "what we changed about our articles" in a post that never says. Plus five P2 overreaches and one negative-parallelism construction.

Coverage limits, stated plainly

The copy reviewer fully read 17 of 36 posts, keyword-verified 11, and did not check 8 (placeholder-color, truncate-text, memoize, minitest-names, docker-rails-7, automation-test-plan, sales-onboarding, single-source-of-truth). Those were written from their openings earlier, so they aren't unread — but they have not had a second pair of eyes. It also read only §0–§2 of the voice guide, not §3's full banned-pattern tables.

Flagged and deliberately not acted on: ~19 of 36 share one sentence template. Several are varied; a full pass would be a rewrite, not a fix.

Gates: bin/hugo-build 8/8 · marketing_copy_test 3 runs / 0 failures · okf validate --strict conformant · 0 truncated descriptions in rendered output.