claude-directory-rank-decoded

Claude directory rank — decoded (the browsable listing over time)

Sibling to claude-registry-search-decoded. Different surface — keep them separate: that page decodes the backend search_mcp_registry (what the model retrieves at prompt time, a deterministic queryable engine, where metadata edits reliably move rank). THIS page decodes the public browsable directory listing (claude.ai/directory/connectors) — an algorithm we do NOT control and can only observe once per daily scrape. Both are “Rank” in discoverability-framework; they behave oppositely on the one lever clients ask about (editing). Feeds the signal inventory in signal-set-and-weights.

Method (one line): read-only reconstruction of a daily per-connector rank time series from the platform’s own SCD2 catalog history (a new version is minted on every rank change; valid_from is the scrape clock), + four natural-experiment studies + a backtested forecast model. 63-day window (2026-05-04 → 07-06), 550 Claude connectors, ~32.5k (connector, day, rank) cells — observational, single directory, algorithm unobserved. Full data, SQL, and model live in the product repo (discoverability-platform → experiments/directory-rank-predictor/, findings 00-SYNTHESIS.md), not here. Competitively sensitive.

Current truth

FINDINGS — evidenced 2026-07-06 (first-party corpus analysis, single 63-day window, observational)

• The history is reconstructable; the “clean” table was empty. The intended durable rank-history table was never populated, but a daily series is fully rebuildable from SCD2 version rows joined to the scrape observation grid. Rank-only changes mint a version (63% of all transitions), so intermediate ranks are preserved. This is the enabling capability behind everything below.
• Freshness dominates — the launch boost is the one large effect. New connectors debut at the very top (median entry rank ~8–22, top ~4% of the directory), then decay. Net of popularity, a <7-day connector ranks ~275 positions above an equally-popular mature one (Spearman(age, popularity- adjusted residual) = +0.53, n≈25.5k) — recency, not popularity, is the dominant non-popularity driver of directory rank.
• Decay is two-phase and gated by a fixed 14-day “New” badge. Launch advantage half-life ≈ 2–3 days to a “new-connector shelf” (~top 25–30%); the “New” badge is a fixed ~14-day timer from the listing date (115/125 connectors drop it at day 12–14, independent of rank), and decay runs ~4× faster once it drops (≈1.7 → 7.2 rank positions/day). The badge materially holds rank up.
• Crowding is mechanical. Each new entrant shoves those ranked below it down ~1 position; corr(daily new-entrant intake, mean incumbent rank shift) = 0.90, on top of a constant ~3.7 position/day churn that runs regardless of intake. A connector decays partly just because others are born above it.
• Predictability is inertia, not insight (backtested, strict temporal holdout). A structural forecast (age/decay/badge/popularity/crowding) beats a naive “yesterday’s rank” persistence baseline by only 11% at 1-day / 13% at 7-day (rank MAE 7.97 / 31.5), and that entire edge is the model tracking the directory growing ~4.4 connectors/day — re-expressed in percentile space the edge vanishes. It ships as a forecast (“where will you drift”), needs no probe channel (the daily scrape is the data), but is not a prescriptive advisor. The one regime that is a real lever — brand-new connectors — is the worst-predicted (per-connector high variance, MAE up to 65 at 7-day).

REFUTED — 2026-07-06 (tested, came back negative)

• “A metadata/description edit buys back directory rank” — refuted for the body of the directory. Diff-in-diff around real text edits vs matched non-editing controls: core effect −9.9 positions, 95% CI [−22.8, +3.1] — crosses zero (n=34), transient (gone within ~3 weeks), and confounded with a co-moving popularity surge. The eye-catching outliers (rank ~500 → top-100) are connector re-registrations (new identity → a fresh launch boost), not renames. Contrast with backend retrieval, where edits DO reliably move rank → claude-registry-search-decoded. Do not promise directory-rank gains from copy edits the way we can for retrievability.
• “Popularity is a leading indicator of directory rank” — refuted as a lever. Popularity and rank co-move in levels (per-day Spearman −0.66) but their day-to-day changes are near-decoupled and neither leads the other (cross-correlation peaks contemporaneously, k=0, symmetric). “Drive popularity and wait for rank” is not supported.
• “The trending flag drives rank” — refuted / unusable. The trending boolean is dead in the data (0 true); high trending_score weakly precedes better forward rank but is confounded with age and popularity. Not a clean signal.

CAPABILITY (what we can do with this — not a fact about the engine)

• We can reconstruct and forecast any Claude connector’s directory-rank trajectory from the daily scrape alone, and characterize its lifecycle (launch boost, badge window, decay, crowding). Useful for diagnosis and expectation-setting (“you’ll decay off the shelf in ~2 weeks, here’s the curve”), not for a copy-edit advisor. The honest client story: the directory rewards freshness/launch timing, not metadata optimization.

Open questions

• Window is only 63 days — every decay/edit magnitude is n-bounded; re-fit as history accrues.
• Category segmentation — is the boost/decay/badge behavior category-dependent? (unsegmented so far)
• Mechanism of the ~275-slot boost — algorithmic recency vs editorial/“featured” curation? Look for a placement signal that isn’t pure age.
• Re-registration as a deliberate lever — it demonstrably re-triggers the launch boost (n=7), but costs the connector its identity/history; needs more events + a risk write-up before we’d advise it.
• ChatGPT directory port — same method, different (unobserved) algorithm; untested. Everything here is Claude.

Timeline

• [2026-07-06] (research) Page created (entry below). Provenance note — the eyeball observation that preceded the study (2026-07-01, Vincent + Elliot afternoon call): scanning rank-across-updates charts, Vincent spotted update→jump patterns for some connectors and none for others; his explanation then [idea] was local MCPs get a fresh “added” date on update (≈ new connector → launch boost) while remote connectors (e.g. Direct Booker) don’t — consistent with, and refined by, this page’s findings (the big jumps are re-registrations/fresh-identity launch boosts; ordinary metadata edits refuted as a rank lever). Same call also observed [idea] that many connectors’ tool updates coincide with regressions in our picked-rate data (probably tech fixes, not optimization).
• [2026-07-06] (research) Page created. Work done 2026-07-06 in the discoverability-platform repo (experiments/directory-rank-predictor/, studies T1–T6: time-series reconstruction, birth/decay/ crowding, edit-rebound, popularity-coupling, backtested forecast, synthesis). Read-only research pipeline — no production data-model changes made. Prompted by Vincent’s ask to run the retrievability-style decoding exercise on the browsable directory rank.