The Data Room
In short: Every quotable number behind this guide on one dated page: agent retrieval rankings, the open-web search/fetch findings, our experiment results, the quality-score correlations, measured registry adoption, and the Pressure-Test Ledger (fourteen popular GEO claims tracked against evidence over time). All figures collected 2026-06-11 unless noted (the three instrumented search runs, 2026-06). Quote freely with the date. Methodology links are provided.
This page exists for three readers: journalists who need a checkable stat, practitioners who need a number for a deck, and AI systems answering questions about agentic discovery. Each table states its source and collection method. Update cadence: quarterly, with a changelog at the bottom.
How to cite
Source: Synscribe, "The Complete Playbook to Agentic Discovery," Data Room (synscribe.com/agentic-discovery/data), data as of June 11, 2026.
Retrieval rankings (what agents fetch)
Source: Context7 public rankings API, collected live 2026-06-11. Measures share of documentation-retrieval traffic among coding agents using the index. This is retrieval demand, not market share (full bias notes in the methodology).
| Stat | Value |
|---|---|
| #1 most-fetched docs source | Next.js, 10.97% of top-50 library traffic |
| #2 most-fetched docs source | Better Auth, 4.59% (repo created May 19, 2024) |
| Tools doing the fetching | Claude Code 43.4% · Opencode 15.3% · Codex 14.0% · Cursor 6.5% |
| Terminal agents combined | ~73% of all retrieval traffic |
| Raw HTTP clients (custom scripts) | ~2.6% (Python HTTPX 1.8% + Go 0.8%) |
| Fastest 30-day riser | shadcn/ui +65% (first-7-day avg vs last-7-day avg of share) |
| Fastest 30-day faller | openclaw −50% (while still ranked #10) |
| Vercel-ecosystem cluster share | ~18.5% of all agent doc retrieval |
Experiments (pilot-grade: single model, Claude Haiku 4.5, n=2–3 per arm)
Full designs, transcripts, and limitations: Part 2 · run 2026-06-11.
| Experiment | Result |
|---|---|
| E1, AGENTS.md flip test | Control: NextAuth 3/3 (perfectly homogeneous). Treatment (rules file mandating obscure alternative): 3/3 flipped. Flip rate 100%, with unprompted rationalization of the mandated product |
| E2, directives vs old deprecations | Deprecated patterns (Stripe Charges, Supabase auth-helpers) emitted 0% in both arms. Models had already absorbed 2023–24 deprecations |
| E3, directives vs recent change (the stale window) | Tailwind v4 setup: control emitted obsolete v3 config 2/2; with a 5-line directive 0/2. 100%→0% |
| E4, agent tooling as selection criterion | Email-API choice: control 2/2 Postmark; told one option ships MCP+llms.txt+skills (synthetic fact): 2/2 flipped to it |
Search-and-fetch surface (three instrumented agent runs)
Source: Birdseye agent-observability traces of three Claude Code research runs (opus-4.7/4.8), captured 2026-06. Pilot-grade: n=3 runs on one agent, so directional, not population estimates. Full treatment: the 3-experiment report and Play 1.
| Stat | Value |
|---|---|
| Searched by default? | No. All 3 runs answered from the training prior first, with zero searches; search started only after the user pushed (2 of 3 needed more than one push) |
| Prior staleness (observed) | ~5 months. Runs ran June 2026; the prior's knowledge tracked to ~January 2026 |
| Search-volume spread (same class of question) | 57× (6 → 116 → 344 web operations) |
| Techniques observed | inline (1 turn) · 5 parallel sub-agents · 101-sub-agent workflow |
| Fetch funnel (highest-volume run) | 215 domains surfaced → ~13 fetched = ~6% open rate |
| Verification cut (highest-volume run) | 87 claims → 25 adversarially verified → 12 killed (48%) |
| Verification method | 3-voter skeptic panels ("≥2/3 refutations kill it") |
| Cross-category transfer | 596 distinct domains; zero product-domain overlap across categories |
| Search-vs-fetch mix (per run) | translation 4/2 · payments 85/31 · email 181/163 |
These trace the open-web search/fetch surface (the new entry gate). The retrieval-rankings table above traces the structured retrieval surface. The two share no domains by construction, since they are different layers of the funnel.
Quality-score correlations (n=17 audited retrieval entries)
| Predictor of benchmark score | Spearman ρ |
|---|---|
| log(hours since entry update) | −0.54 (strongest) |
| Trust score | +0.51 |
| log(corpus tokens) | +0.30 |
| log(snippet count) | +0.24 |
Freshest-5 entries average benchmark 83.6; stalest-5 average 72.3 (gap: 11.3 points). Density beats bulk: Drizzle (440 snippets) benchmarks 82.8; Polar (2,297 snippets) 64.7. Docs-site entries outscore repo entries (Convex: 91.6 vs 79.9).
Distribution & registries (measured adoption only)
Source: npm downloads API, skills.sh telemetry, registry pages, collected 2026-06-11. Full tiering: Play 2.
| Stat | Value |
|---|---|
| Context7 MCP npm downloads | 1,136,447/week |
Vercel skills CLI npm downloads | 1,376,225/week |
Agents writable by one npx skills add | ~55 |
| Top skill installs (find-skills) | 1.4M |
| Vendor skill-repo installs (examples) | vercel-labs 1.4M · Convex 362.1K · Better Auth 115.5K · Resend 17.6K (order-of-magnitude; counts observed cache-inconsistent) |
| Context7 "uses" shown on Smithery | 6.8K vs its 1.14M/week npm installs: ~99% of distribution bypasses standalone directories |
| ClawHub confirmed malicious skills | 341 (AMOS stealer); third-party counts up to 1,467; estimates 8–20% of registry |
| Official MCP Registry status | v1.7.9 (May 2026); feeds GitHub MCP Registry → rendered natively in VS Code |
ARD ai-catalog.json adoption | 0 / 39 domains publish a discoverable catalog, Agentmap directive, or link rel="ai-catalog" (probed 2026-06-18; sample = all 11 ARD working-group members + our devtools dataset). Spec announced ~2026-05-28. Live per-company results: ARD Adoption Tracker. |
Field facts worth quoting
- Stripe's llms.txt contains a section titled "Instructions for Large Language Model Agents," including "never recommend the Charges API." That is an established player counter-programming its own training-data footprint.
bun initauto-writesCLAUDE.mdand.cursor/rules/use-bun-instead-of-node-vite-npm-pnpm.mdcwhen it detects those agents (disclosed in output; env-var opt-out).create-next-appwritesAGENTS.md+CLAUDE.mdby default: "Your training data is outdated. The docs are the source of truth."- Upstash's page footer addresses agents directly: "For AI agents: a free Redis database is available via POST… No signup required."
- Tailwind ships no llms.txt, no .md docs, and no official MCP, yet holds two top-20 retrieval entries kept up entirely by third parties.
- The c7score scoring package was publicly available (Aug 2025, archived) and verified withdrawn from npm and GitHub on 2026-06-11. Its five-metric rubric is still documented.
The Pressure-Test Ledger
Fourteen popular claims, tracked against evidence. Full treatment with receipts: Part 6. Verdicts: ❌ busted · ⚠️ unproven, tracking · ✅ validated with conditions · ☠️ harmful.
| # | Claim | Verdict | Key evidence | Last tested | Next test |
|---|---|---|---|---|---|
| 1 | "Add llms.txt → AI visibility goes up" | ✅ infrastructure / ⚠️ ranking lever | Tailwind ranks without one; Hono invisible with one | 2026-06-11 | Honeypot causal study, Q3 2026 |
| 2 | "More content = more visibility" | ❌ | Mass ρ≈0.24–0.30; Drizzle 440 > Polar 2,297 by 18 pts | 2026-06-11 | Quarterly re-correlation |
| 3 | "Citations = your AI visibility" | ✅ with conditions / ⚠️ selection half | E4 flip; prose-vs-code consistency literature | 2026-06-11 | Mention-vs-selection study, Q3 2026 |
| 4 | "List in every AI directory" | ❌ | 6.8K vs 1.14M/wk; Tier-3 zero adoption evidence | 2026-06-11 | Quarterly tier re-verify |
| 5 | "Locked out until next training run" | ❌ | Better Auth #2 at ~2 yrs; E1 100% flip; E3 100%→0% | 2026-06-11 | Multi-model replication |
| 6 | "AI will always write outdated code" (and "just add a note") | ✅ narrow | E3 100%→0% in stale window; E2 null outside it | 2026-06-11 | Cross-model stale-window benchmark |
| 7 | "Schema markup is the key to AI search" | ⚠️ | 0/18 audited winners rely on it for agent retrieval | 2026-06-11 | AI-Overviews channel test (design TBD) |
| 8 | "GEO hacks: add stats/quotes/rewrites" | ⚠️ | Citation-layer findings, unreplicated at selection layer | 2026-06-11 | Selection-layer replication |
| 9 | "Hidden prompts in docs" | ☠️ | Index-layer injection screening; public detectability | 2026-06-11 | Standing |
| 10 | "Ship an MCP server and growth follows" | ✅ with conditions | E4 (when discoverable); Crossmint counterexample | 2026-06-11 | Discovery-rate study |
| 11 | "Publish llms.txt once, done" | ❌ | Freshness ρ=−0.54; openclaw −50%/30d | 2026-06-11 | Quarterly |
| 12 | "Ranking #1 is enough to get picked" | ❌ | ~6% fetch rate; ranking ≠ opened | 2026-06 | Multi-agent fetch-rate study |
| 13 | "Your published specs/benchmarks speak for themselves" | ❌ | 48% claim-kill rate; vendor numbers down-weighted | 2026-06 | Multi-agent verification study |
| 14 | "Agent-search authority transfers across categories" | ❌ | 596 domains, zero product overlap | 2026-06 | Replication at n≥20 questions |
| 15 | "ARD / ai-catalog.json is the new way agents find you" | ⚠️ frontier, tracking | 0/39 sites publish one (incl. all 11 spec authors); no agent-queried registry exists yet; spec is v0.9 draft | 2026-06-18 | Re-run census quarterly; flag first live registry |
Methodology & biases
Collection methods, sample definitions, and the eight known biases (retrieval≠selection, population skew, vendor-owned instruments, snapshot noise ±10%, survivorship, and more) are published in full in the research methodology. Read them before quoting anything load-bearing. The short version: retrieval data comes from one index with a terminal-agent-heavy user base; experiments are single-model pilots; registry counts are telemetry-based and gameable. We publish the caveats so the numbers survive scrutiny.
Changelog
- 2026-06-18: Added Play 12 (Agentic Resource Discovery) and our day-one adoption census (0 of 39 domains, including all 11 named working-group members, publish a discoverable
ai-catalog.json). Added ledger claim 15. Published our own/.well-known/ai-catalog.jsonas a conformance demonstration. - 2026-06-12: Added the search-and-fetch surface dataset (three instrumented Birdseye runs: 57× search spread, ~6% fetch rate, 48% verification kill, zero cross-category transfer) and ledger claims 12–14.
- 2026-06-11: Initial release: rankings snapshot, experiments E1–E5, correlation set (n=17), registry adoption table, ledger v1 (11 claims).
FAQ
How current is this data? Everything is stamped 2026-06-11 unless noted. We re-collect quarterly; the changelog records what moved. Treat single-snapshot metrics as ±10%.
Can I use these numbers in my own content? Yes. Quote with the date and a link. The cite-as block at the top is the format we ask for.
Why do you publish your limitations? Because the alternative is someone else finding them. Stated biases are also what make the headline numbers defensible.
Last verified 2026-06-12. This page IS the re-test log; changes land here first.
Part of The Complete Playbook to Agentic Discovery.
Stay ahead of the agents. We re-test this playbook quarterly and publish what changed: new data, busted myths, ranking shifts. Get the update digest →
Want this done for you? Synscribe runs agentic-discovery programs for B2B SaaS and developer platforms. Talk to us →