> ## Documentation Index > Fetch the complete guide index at: https://www.synscribe.com/agentic-discovery/llms.txt > Use this file to discover all pages before exploring further. --- title: "Drizzle Case Study: 440 Snippets That Beat 2,297" description: Drizzle benchmarks 82.8 on just 440 Context7 snippets while Polar's 2,297 score 64.7 and Hono misses the top 50. Density beats surface checklists. slug: /agentic-discovery/case-studies/drizzle series: The Agentic Discovery Playbook — Case Study last_verified: 2026-06-11 --- # Drizzle's 440 Snippets Beat Corpora 5× Bigger (and What Polar and Hono Teach) > **The lesson:** Drizzle ships almost no agent surface — an llms.txt and one flawed llms-full.txt — yet benchmarks 82.8 on 440 snippets, beating Polar's 2,297-snippet full-stack surface (64.7) and Hono's exemplary three-tier llms.txt (63.3, absent from the top 50). Index-entry quality and freshness beat checkbox count. Density is the rubric. ## At a glance | | | |---|---| | Category | TypeScript ORM | | Context7 traffic rank | **#19** (`/drizzle-team/drizzle-orm-docs`, 1.22% of top-50 share, 2026-06-11) | | 30-day momentum | **+14%** (first-7-day vs last-7-day share) | | Audited entry (`drizzle-team/drizzle-orm`) | 440 snippets / 96,732 tokens, benchmark **82.8**, updated 1 week | | The contrast pair | Polar: 2,297 snippets, benchmark 64.7, 1 month stale · Hono: 48 snippets, benchmark 63.3, absent from top 50 | ## What they built Almost nothing — that's the point. Drizzle's deliberate agent surface is an llms.txt plus a 94KB llms-full.txt. And the llms-full.txt isn't even clean: it's raw MDX with build artifacts leaking through, including framework imports like `import Npm from '@mdx/Npm.astro'` sitting in the middle of what's supposed to be an agent-readable document. No `.md` mirrors. No MCP server. No agent docs pages. By the checklist in our [readiness audit](/agentic-discovery/agent-readiness-audit), Drizzle fails most rows. What it has instead is docs *content* that happens to be exactly what retrieval rubrics reward: terse, self-contained permutation pages — "Get started with Drizzle and X" repeated across roughly 60 database/runtime combinations, each one complete with install, imports, config, and code. No conceptual sprawl, no duplicated patterns, nothing that needs another page to run. The permutation structure matters more than it looks. An agent rarely asks "how does Drizzle work" — it asks "Drizzle + Turso" or "Drizzle + Supabase + Bun," because it's holding a specific stack. Sixty pages that each answer one combination completely map one-to-one onto the questions retrieval systems generate and developers actually type. The form *is* the optimization, whether or not it was intended as one. The instructive part is the triangle with two products that made the opposite trade: | | Drizzle | Polar | Hono | |---|---|---|---| | llms.txt | yes (+ flawed llms-full) | yes + llms-full | **exemplary**: three tiers incl. ``-wrapped llms-small.txt | | `.md` mirrors | no | yes, with agent banners | no | | MCP | no | **production + sandbox** servers, 8-client recipes | no | | Context7 entry | 440 snippets, **82.8**, 1 week fresh | 2,297 snippets, **64.7**, **1 month stale** | repo parsed to 48 snippets / 3,267 tokens, **63.3** | | Top-50 traffic rank | **#19, +14%/30d** | — | absent | Hono's surface even includes the most stylish llms.txt artifact we found: > "``This is the tiny developer documentation for Hono.``" > — hono.dev/llms-small.txt, a docs file that opens by impersonating a system prompt And it doesn't matter — because Hono's Context7 *repo* entry parsed to almost nothing (48 snippets, 3,267 tokens), nobody curated it, and the framework is nearly invisible at the retrieval layer despite being beloved and fast-growing. ## The receipts All figures observed 2026-06-11; single-day metrics carry ±10% error bars, and benchmark scores are recomputed continuously. **The density result.** Drizzle: 440 snippets, benchmark 82.8 — the highest in its audit batch on the fewest snippets. Polar: 2,297 snippets, benchmark 64.7. That's the pairwise illustration of our wider finding: across 17 indexed entries, benchmark tracked corpus mass at only ρ≈0.24–0.30. The published c7score rubric explains why — it scores question coverage, snippet uniqueness (duplication penalized), formatting, metadata noise, and initialization (install + imports present). Sixty self-contained "Drizzle + X" pages are a near-perfect fit; 2,297 snippets that repeat patterns actively lose points. **The freshness result.** The strongest correlate of benchmark score in our study wasn't size or surface — it was recency: | Correlate of benchmark score (n=17 entries) | Strength | |---|---| | log(hours since update) | Spearman **−0.54** — the strongest we measured | | Corpus mass (tokens / snippets) | ρ≈0.24–0.30 — barely matters | | Freshest 5 entries, average benchmark | **83.6** | | Stalest 5 entries, average benchmark | **72.3** | Drizzle's entry was 1 week fresh at snapshot; Polar's was 1 month stale. Polar's problem isn't its surface — production *and* sandbox MCPs, full Mintlify pipeline, agent banners on every `.md` page — it's that nobody is watching the index entry. The Mintlify pipeline even stamps each page with the right instinct: > "## Documentation Index — Fetch the complete documentation index at .../llms.txt — Use this file to discover all available pages before exploring further." > — banner atop every `.md` docs page at Polar (and Bun, Crossmint, x402, and others) — agent-grade by platform default, stale by neglect **The parse result.** Hono's failure mode is different again: its surface is fine, but the entry agents actually retrieve through parsed to 48 snippets. Surface ≠ checkbox; the index entry has to actually parse well, and someone has to look. **Demand confirmation:** `/drizzle-team/drizzle-orm-docs` holds rank #19 in top-50 Context7 fetch share (1.22%) and grew +14% over 30 days — minimal surface, real and rising agent demand. ## What to copy - [ ] Write permutation pages for your golden path — "{Product} + {their stack}" ×N, each fully self-contained — agents query the combination. ([Play 7](/agentic-discovery/code-snippets-for-ai-agents)) - [ ] Make every snippet init-complete (install + imports + runnable code) and demonstrate each pattern exactly once. ([Play 7](/agentic-discovery/code-snippets-for-ai-agents)) - [ ] Treat your Context7 entry as the product: claim it, watch the "Updated" field, re-parse within 7 days of every release. Polar's 64.7 is mostly a staleness tax. ([Play 2](/agentic-discovery/ai-agent-registries-and-directories)) - [ ] Verify what the parser actually extracted — Hono's exemplary llms.txt coexists with a 48-snippet index entry. Audit the entry, not your site. ([Play 2](/agentic-discovery/ai-agent-registries-and-directories), [Play 5](/agentic-discovery/llms-txt)) - [ ] Lint your llms-full.txt as an agent would read it — Drizzle ships MDX build artifacts (`import Npm from '@mdx/Npm.astro'`) in its agent file and wins anyway, but that's luck, not a plan. ([Play 5](/agentic-discovery/llms-txt), [Play 6](/agentic-discovery/markdown-docs-for-ai-agents)) - [ ] Score yourself with the five-metric rubric before the index does. ([Play 7](/agentic-discovery/code-snippets-for-ai-agents), [Play 11](/agentic-discovery/ai-evals-and-leaderboards)) ## What NOT to over-copy - **Minimal surface is not the recommendation.** Drizzle wins *despite* skipping `.md` mirrors, MCP, and agent docs, not because of it — it leaves Plays 6, 3, and 4 entirely on the table. The lesson is that content density is the prerequisite the other plays sit on, not a substitute for them. - **Don't ship the flaw with the pattern.** The MDX artifacts in llms-full.txt are a defect that happens not to be fatal here. Clean output costs nothing. - **Category demand is doing some lifting.** ORMs are high-retrieval-demand, boilerplate-heavy territory; a low-demand product with identical docs would score well and still see little traffic. - **The triangle is three data points.** Drizzle/Polar/Hono illustrate mechanisms our 17-entry correlations support, but each pairwise comparison is n=1 vs n=1, snapshot ±10%, and Polar's staleness could resolve with a single re-parse — these positions are not permanent. ## FAQ **Why does Drizzle score so high with only 440 snippets?** Because the scoring rubric measures density, not mass: self-contained, init-complete, non-duplicated snippets that cover common developer questions. Drizzle's ~60 terse "Get started with Drizzle and X" pages fit that shape exactly, benchmarking 82.8 (2026-06-11) while Polar's 2,297 snippets score 64.7. **Is an llms.txt enough for agentic discovery?** Drizzle suggests llms.txt plus genuinely dense docs can carry a product to rank #19 in agent fetch share — but Hono proves the inverse: a best-in-class llms.txt with an uncurated index entry (48 parsed snippets) leaves you invisible. The deciding layer is the retrieval index entry, not the files on your domain. **Why is Hono missing from the Context7 top 50?** Its repo entry parsed to almost nothing — 48 snippets and 3,267 tokens, benchmark 63.3 — so agents retrieving through Context7 find barely anything to fetch. Nobody curated the entry or submitted the docs site as the canonical source; the fix is index ownership, not more docs. **Does shipping an MCP server raise your retrieval ranking?** Not directly — Polar ships production and sandbox MCP servers and still benchmarks 64.7 with no top-50 presence, while MCP-less Drizzle sits at #19. MCP solves a different problem (live capability and usage depth); retrieval ranking is decided by index-entry quality and freshness. --- *Snapshot date 2026-06-11; single-day metrics carry ±10% error bars. Part of [Case Studies](/agentic-discovery/case-studies) · [The Complete Playbook to Agentic Discovery](/agentic-discovery).* ← Previous: [Clerk: Docs Written as Prompts](/agentic-discovery/case-studies/clerk) · Next: [Glossary of Agentic Discovery](/agentic-discovery/glossary) → > **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 →](/agentic-discovery#updates) > > **Want this done for you?** Synscribe runs agentic-discovery programs for B2B SaaS and developer platforms. [Talk to us →](/contact)