seo: we rank 4–8 when we appear and appear almost nowhere — the GSC-measured program

[JSONbored]

From the owner's GSC exports, 2026-08-14 (coverage + 7 drilldowns + performance + breadcrumbs). The numbers change the strategy: this is not an authority problem.

Measured

• 3 months: 8 clicks, ~8,000 impressions, average position 5.7. When we appear, we rank well. We almost never appear: 75 distinct queries in 90 days.
• The queries we DO get are subnet lookups — literally pasted repo URLs (github.com/chronollm/sn38, actual-subnet-95, satori) and competitor subnet pages, ranking us 4–8. The demand is per-subnet; our supply of rankable per-subnet content decides everything.
• Impressions collapsed ~600/day (late June) → single digits (mid-July on). Cause visible in coverage: indexed count ballooned 2,542 → 4,099 with thin chain pages, then Google reassessed — Aug 5 it deindexed 488 pages, and 5,797 sit "Crawled – currently not indexed" (validation FAILED). The 1,000-row drilldown is 993 /extrinsics + /blocks + /accounts hash pages — thin by nature.
• 82 of the 83 404s are apex /metagraph/* paths — API-host artifact links crawled on metagraph.sh where they don't exist.
• 18 soft 404s: /blocks entity shells answering 200 for absent content.
• Structured data is absent: breadcrumbs report shows ONE valid item site-wide. No rich results anywhere.
• 124 canonical-alternates are the ?tab= variants working AS DESIGNED — canonicals are healthy, no action.

The program, in order of measured leverage

1. Let the thin-page purge finish (already in motion): the robots disallow on /blocks/, /extrinsics/, /accounts/ per-entity spaces shipped recently; the crawled-not-indexed backlog drains on its own. Do NOT try to validate-fix those 5,797 — they should leave the index.
2. Kill the 82 apex /metagraph/* 404s: find the emitter (links that should be absolute to api.metagraph.sh) or 301 the prefix apex→api. One change, 82 errors.
3. Soft-404s answer 404/410: the /blocks entity shell must send a real status for absent refs (entityNotFoundMeta exists; the status code doesn't follow it everywhere).
4. Structured data on the money pages: BreadcrumbList + Dataset (per-subnet data provenance) + SoftwareApplication/WebAPI on subnet pages, FAQPage on docs. One breadcrumb item site-wide is why zero rich results.
5. Title/description templates tuned to the demand we can see: subnet pages should carry the on-chain name, repo slug, and netuid aliases in titles ("SN38 · ChronoLLM — API, health, economics | metagraphed"), because pasted-repo-URL queries are what already finds us.
6. Per-page OG cards (feat(og): per-page OG cards — the bundle-budget argument that evicted the render no longer holds #11075) — CTR on the impressions we have (0.1% CTR at position 5.7 says our snippets/preview are the weakness).
7. New rankable surfaces from data we already serve: compare pages, "subnets requiring a GPU" (gpu_required facet), per-subnet economics summaries — each a stable URL answering a real query.

Owner actions, dashboard-side: sitemap resubmitted 2026-08-14 (1,607 discovered — plausible: subnets+providers+docs+news); start validation on the 404 and soft-404 buckets AFTER 2 and 3 land so validation passes.

[linear-code]

JSO-8265

[JSONbored]

Status, 2026-08-15 — items 2–5 shipped, and the diagnosis for item 7 changed

item	state
1 · let the thin-page purge finish	no action needed, as written
2 · apex /metagraph/* 404s	done — #11225
3 · soft-404s answer a real status	done — #11225
4 · structured data	done — #11230
5 · title/description templates	done — #11230
6 · per-page OG cards	still #11075, gated behind the containers epic
7 · new rankable surfaces	premise changed — see below

Item 7 was solving the wrong problem

This issue's read was that "our supply of rankable per-subnet content decides everything". The supply was already there; it was invisible to a crawler.

The registry hubs virtualize their tables (#8248), so the server-rendered HTML only ever carried anchors for the rows in view:

space	linked before	total	after
subnets	30	129	129 (#11231)
providers	25	138	138 (#11232)
api-reference groups	0 (all 25 paths 404'd)	25	25 (#11234)

So 99 subnet pages and 113 provider pages had no internal link anywhere on the site — reachable only from sitemap.xml, which is the profile that lands a URL in "Crawled – currently not indexed", the bucket 5,797 of ours sit in. Those are the same per-subnet pages the measured demand asks for.

Recommendation: measure before building item 7. Adding new surfaces was the bet that lost last time (indexed count 2,542 → 4,099 with thin pages, then Google reassessed and ~600/day became single digits). The existing 129 subnet pages are now linked, titled for the demand GSC records, and carry Dataset + breadcrumbs. Give that a crawl cycle and read the index report before minting more URLs.

Also fixed, not in the original list

• The Content Signal was inert. feat(discovery): declare the Content Signals, and serve auth.md where scanners look #11174 emitted Content-Signal: above the first User-agent: line on both hosts. robots.txt is group-scoped (RFC 9309), so it belonged to no group and was read by nothing — the declaration permitting AI use had never been in effect. Fixed both hosts in fix(discovery): the Content Signal was outside its group, and 113 provider pages had no link #11232, and both tests now pin the position rather than the line's existence.
• ~290 docs pages each linked a 404. Every generated operation page links its own group path, and none of the 25 had a page behind it. That also made the new BreadcrumbList name 404 URLs as crumb targets, which Google validates. Fixed in the generator (fix(docs): the 25 API-reference group paths every page links were 404s #11234), so a new API tag cannot reintroduce it.

Checked and deliberately left alone

• /validators links 50 of 1,023. Not fixed on purpose: a thousand links on one hub reads as link-farming and validator pages are not the measured demand. The input for that call, if the index report says they rank: 599 of 1,022 carry an operator-declared on-chain identity (coldkey_identity.name — Datura, TAO.com, K8STAKE.AI), and there is no top-level name fallback, so those are real. Link that subset, not all 1,022.
• Sitemap — 1,628 URLs, zero duplicates, zero wrong-host, zero blocked by robots.txt.
• AI discovery — every URL advertised across llms.txt, llms-full.txt, agent.md and agent-workflows.md resolves (147 links checked); 7 of 8 well-known resources serve 200 on both hosts, the exception being /ai.txt, a much less adopted proposal than the llms.txt we already serve.
• Docs reachability — all 321 sitemap docs pages are reachable by internal links (a BFS from /docs converges at depth 12 with 348 pages, 0 orphaned). The group indexes above shorten that considerably.
• "Alternate page with proper canonical tag" in GSC is the ?tab= variants working as designed — an info row, not an error.

GSC-side validation for the 404 / soft-404 buckets was started by the owner on 2026-08-15, after items 2 and 3 landed.