Needs Work

Live since 27 Oct 2026

These are totals read from the private control panel. They show how large the ranking and the weekly scan are, not what the owner has done with them: no site is named, and a count of zero is printed as zero.

sites in the ranking

154

measured 28 Oct 2026

in the red tier

6

measured 28 Oct 2026

in the amber tier

36

measured 28 Oct 2026

pinned at the top

6

a noindex homepage or a sitewide robots block

measured 28 Oct 2026

sites in the latest weekly scan

154

140 ok, 9 partial, 5 inconclusive

measured 28 Oct 2026

sites with half or more of their measured pages thin

36

at least five pages measured

measured 28 Oct 2026

Measured weekly. Counts taken on 28 Oct 2026. Scores from the run of 28 Oct 2026. A scan that could not be read is marked inconclusive and set aside; it is never counted as thin.

Busy-looking is not the same as worked-on

Counted from deploy history, 90 days to 28 Oct 2026

successful deploys (finished or ready) in the last 90 days across 150 sites

7,556

measured 28 Oct 2026

How it works, in five steps

Detect, rank, act. It recommends; the owner decides.

1. 1

   Work log

   Every day the system reads each site's Git commit history and sorts every commit into real work or routine (SEO, error-tracker, dependency, automated data refresh, fleet-wide sweep, bot and deploy plumbing).

2. 2

   Weekly site scan

   A polite crawler visits every website: homepage, robots.txt, sitemap, 20 evenly sampled pages, a fake URL to catch "soft 404s", internal links, and the homepages of competitors that outrank it.

3. 3

   Signals

   Search performance (Google Search Console daily data, traffic-cliff detection), AI-search citations, engagement, backlinks and open issues are joined to the scan. Indexing coverage is built but not yet active (see the signals below).

4. 4

   Scoring

   Each site gets a transparent score built from named terms. Every point carries its evidence, such as days since real work or the share of measured pages that are thin. Critical problems are pinned to the top.

5. 5

   Delivery

   A ranked page in the private control panel, and a daily email with the top 15, today's suggested plan and a copy-paste job prompt for each site, written to be handed straight to an AI coding agent. The email is switched on and is sent each morning.

What gets measured

The weekly site scan

• Page checks

  Per page: word count, noindex (meta tag and HTTP header, including bot-specific directives), headings, structured data, images, internal links, search box and navigation links. Also the sitemap, including sitemap indexes, and whether the site has an llms.txt.

• Thin content

  Pages under 300 words. Counting is language-aware: Thai, Chinese, Japanese, Lao, Khmer and Burmese pages are segmented with the Unicode word segmenter instead of being misread as thin.

• Template duplication

  8-word shingle similarity flags pages that are near-copies of each other.

• Outdated content

  Stale years in titles and headings, "last updated" dates more than 12 months old, and old copyright years. Historic context such as a founding year is deliberately ignored.

• Broken links

  Up to 60 internal links checked per site. Only a definitive 404 or 410 counts as broken; blocked or rate-limited checks count as unknown.

• Competitor homepages

  Up to 5 real competitors per site, taken from search results where they outrank it, with platforms such as video and encyclopedia sites filtered out, cached for 30 days.

• Polite by design

  Rate-limited, small fixed concurrency, an honest self-identifying user agent, 12-second timeouts and capped page sizes.

Search, AI and audience signals

• Traffic trend

  The last 28 days of clicks against the 28 days before, from Search Console. Sites with zero clicks and real content are flagged.

• Traffic-cliff dating

  The control panel's existing cliff detector runs over the daily series, so a collapse older than a month is still detected and dated, and matched to the nearest Google algorithm update where one is close.

• Indexing gap

  Pages in the sitemap against pages Google actually indexes. Built but not yet active: the panel does not yet hold a trustworthy indexed-page count, so this term does not fire.

• AI-citation gap

  Sites that get search impressions but are never cited in AI answers, using the Google AI Overviews and Perplexity checks the panel already runs.

• Engagement

  Content that ranks but makes visitors leave quickly, from GA4. Built but not yet calibrated on live data.

• Backlink loss

  Referring domains falling month over month.

• Open problems

  The count of serious open issues. Dependency-bot noise is deliberately excluded.

Coverage

Every website on the Hetzner Cloud fleet and on other hosting is in scope; infrastructure tools (monitoring, mail, admin panels), parked sites and sites still being set up are excluded by tag. Static Cloudflare Workers/Pages sites are also kept out of the triage ranking and uptime monitoring, by the owner's decision.

How a site is scored

Each site gets a 0 to 100 score built from named terms, each with its points and its evidence. The weights are not published here: they will be tuned as the system runs.

Pinned rows

A site that is noindex is always shown at the top. The owner keeps several unfinished sites deliberately noindex, and they stay pinned as a reminder that they need work. Thin pages do not pin a site: they are a large score term, and the email lists thin sites in a compact section of their own.

Priority tiers A, B and C

The owner sets each site's priority tier. The score is weighted x1.3 / x1.0 / x0.6 for tiers A, B and C, so a lower-priority site still surfaces when badly neglected but rarely crowds out a higher one. A pinned site is not weighted.

Colour tiers

Red (60+), amber (35 to 59) and green.

Effort tag

Every suggested job is labelled a quick win (add an about page, schema, a search box, llms.txt, fix a few broken links or old titles) or a big project (a content sprint, a relaunch, an indexing gap).

Site-type checklist

An AI model labels each site (directory, news, guide, register, tool, portfolio or other), and the owner can set the label by hand. For a directory, a register, a news site or a guide the score then checks a few things that type should have, such as a search box, enough listing pages, article schema or links between pages. The labels are still being filled in, so the checklist applies only to the sites that have one.

Hygiene terms

Three small terms, capped together, from checks the panel already runs: a poor mobile page-speed score on the last two captures, a fifth or more of the crawled pages missing a title or meta description or repeating a title, and console errors on the homepage in each of the last three runs. A capture that failed or is missing adds nothing and removes nothing.

Site down

A site with an open outage alert is flagged "DOWN, fix uptime first", is left out of today's plan, and has its scan-derived terms set aside when the scan was taken during the outage.

Competitor feature gaps

Features present on 60% or more of a site's real competitors but missing on the site, judged only once at least 2 competitors have been scanned.

The control panel page

A page called "Needs work" in the private control panel holds the full ranked list of every site: pinned first, then by score.

Per-site card

The full score breakdown (every term, its points and its evidence), a score history sparkline, the last 10 commits with their class badges and a Reclassify control, the search queries to target, before and after work-impact results, and a short log of when the site was proposed, through which channel, at what rank and with what outcome.

The morning email

A daily email, separate from the existing health digest, with the top 15 sites.

The email is switched on and is sent each morning. The first one went out on 28 Oct 2026.

Today's plan

One big project plus two quick wins, taken from the owner's Next-up queue first and skipping sites worked on in the last 3 days and sites that are down.

Next-up queue

Up to 5 owner-chosen sites shown first. They clear themselves once real work is detected.

Top 15 ranking

One line per site: rank, score, days since real work, the main problems, traffic at stake, the suggested job and a quick-win or project badge.

Fleet sweeps

Quick wins that the same job fits on 3 or more sites are grouped into one block with a single prompt to run across them.

Movers

The biggest risers and fallers since last week, with the score terms that changed.

Monday scorecard

Sites really worked on last week, and whether the neglect backlog is shrinking, plus a "what worked" look at traffic before and after recent work.

One-click actions

Next up, Dismiss and Worked today from the phone. Links are single-use, expire after 7 days and are stored only as hashes. Each opens a confirm screen, because email security scanners automatically click links.

Copy-paste job prompts

A ready-to-run brief per site (the repository, which project notes to read, the top reasons, sample URLs of thin, outdated or broken pages, the search queries to target, the suggested job, the site's own notes, its latest real commits and a "Done when" line), so a fresh AI coding session starts with full context.

Stale-ranking flag

If the overnight scoring did not complete, the email is still sent, but its subject starts with "[STALE]" and a banner says when the ranking was last scored.

The old "sites needing revisit" section of the health digest is retired.

Honesty and engineering

A ranking is only worth reading if it can be trusted, so the system is built around not fooling itself.

of routine commits mislabelled as real work (4 of 382), in a set of 1,186 hand-labelled commits (a set also used while tuning)

1.05%

measured 28 Oct 2026

of routine commits mislabelled as real work (3 of 212), in a second set of 615 hand-labelled commits from 16 repositories, measured after the rule changes (a set that also informed them)

1.42%

measured 28 Oct 2026

one 1.5 MB adversarial page stalled the first version of the HTML parser, found by the first verifier round

26 min

measured 26 Oct 2026

Also built, and what is switched off

Three more parts. The first two appear on each site's card in the control panel and in the email; neither has produced a result that this page claims. The third is switched off.

Queries to target

For each site, up to 5 real search queries it already gets impressions for but ranks poorly on (below position ten), each with its impressions, its position and the page Google actually ranks for it. "Improve these queries" replaces a vague "add content".

Work-impact report

30+ days after a burst of real work, Search Console clicks and impressions for the 28 days before are compared with the same length of time afterwards, beside the score change. The aim is to learn which kinds of work actually move traffic.

Weekly AI memo

A short note of up to 120 words per red or amber site on what it is missing. The code is built and the owner chose to launch without it, so it is switched off.