OpenGSC Drops Research

Goal

Turn a raw list of (possibly dropped) domains into a ranked, vetted shortlist: alive names retired for free, free names corroborated by two sources, link history attached, spam and burn-vetoes applied — ending in a buy/pass table a human can act on.

Required inputs

• A domain list (pasted text, CSV, or a donor's refdomains). If none given, ask — the funnel is useless without candidates.

OpenGSC MCP tools

• drops_ingest: import the list as a run. Rows are normalised (URLs/www. hosts/IPs handled, everything reduced to its registrable apex); the report says what was rejected and why.
• drops_prefilter: one bounded slice of DNS resolution. Retires every delegated (alive) domain for free — usually ~90% of the list. Run this before anything else; loop while remaining > 0 and done: false.
• drops_check: one bounded slice of the registry stage (RDAP → WHOIS, throttled per zone, ~35s per call). Loop while remaining > 0. Rules baked into the verdicts: an RDAP 404 is only "free" when WHOIS agrees (corroborated), a refusal is never "taken", and zones with no public registry (.gr today) come back uncheckable — they need a registrar API and cannot be swept.
• drops_enrich_wayback: Wayback snapshots for ≤12 domains per call — months archived, first/last seen, days since death. Free.
• drops_enrich_dr: Ahrefs DR for ≤60 domains per call via the free public endpoint. Free (needs the free DR key configured). Attribution "Domain Rating by Ahrefs" is mandatory in anything you publish.
• drops_enrich_refdomains: PAID — Ahrefs refdomains/backlink counts (~100 units per domain). Needs confirm: true. Only for shortlisted rows, never for a whole list.
• drops_history_ai: PAID — the AI history pass for ≤5 hand-picked domains: three Wayback snapshots, the configured LLM decides clean / topic_shift / spam_period / unknown. Needs confirm: true.
• drops_list: the catalogue view — filters, sort, the score, and the funnel counts.

Workflow

1. drops_ingest the list. Report the rejection reasons — a list that ingests 100% clean usually means the parser was tricked.
2. drops_prefilter until done. State the survival rate: from 50 000 rows expect a few thousand.
3. drops_check until done. Treat only available rows as candidates; corroborated: false means "probably free, needs a registrar look before buying". uncheckable rows are a registrar-API matter — do not retry them.
4. Enrich survivors: drops_enrich_wayback + drops_enrich_dr (free, go wide), then — with the user's explicit OK — drops_enrich_refdomains for the shortlist only.
5. Vet the finalists with drops_history_ai, five at a time. spam_period vetoes the row outright; topic_shift discounts the links but does not disqualify a donor-glue play.
6. drops_list sorted by score for the final table: domain, DR, refdomains, snapshots, verdict, score. Flag every veto explicitly.

Judgment rules (from the report consensus, baked into the score)

• Score ≤ 0 rows with a named veto (spam_history, idle_over_2y) are pass — do not talk the user into them.
• DR is a contributor, not a target: a DR40 with a spam interlude loses to a DR15 with a clean, single-topic life.
• The score needs enrichment to mean anything — before step 4 it is mostly a domain-count proxy, and saying so honestly beats quoting it.

Cost discipline

• DNS, registry checks, Wayback and free DR are free — run them freely.
• Refdomains bill Ahrefs units; the AI pass bills LLM credits. Both are gated behind confirm: true and both are meant for shortlists. If the user has not said "go", present the shortlist and ask.