> ## 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: "Resend: the Smallest Corpus With the Highest Score" description: Resend benchmarks 92.3 on Context7 — the highest score in our probes — on a small corpus, with an agent-addressed llms.txt and .well-known discovery. slug: /agentic-discovery/case-studies/resend series: The Agentic Discovery Playbook — Case Study last_verified: 2026-06-11 --- # How Resend Scored 92.3: the Smallest Corpus With the Highest Score > **The lesson:** Resend posts the highest retrieval-quality score in our category probes — 92.3 on Context7 — with one of the smallest corpora we audited. Every page is runnable markdown, the llms.txt addresses agents directly, and discovery manifests live at `.well-known` URIs. Density and addressability beat bulk; a young product can out-score corpora many times its size. ## At a glance | | | |---|---| | Category | Transactional email API | | Context7 benchmark (`websites/resend`) | **92.3** — the highest single score in our category probes (2026-06-11) | | Corpus size (`websites/resend`) | 3,765 snippets — small by winner standards | | SDK entry (`resend/resend-node`) | just 33,355 tokens / 712 snippets, benchmark 80.2, updated 2 weeks | | Agent-skills installs | ~17.6K — order of magnitude only; telemetry-based | ## What they built Resend's agent surface is small, complete, and pointed entirely at machines. The llms.txt doesn't merely exist for agents — it talks to them: > "For AI agents and automation, use the tools below" > — resend.com/llms.txt And the "tools below" include something we found nowhere else in 18 audits — discovery manifests at well-known URIs: > "- [MCP Discovery](https://resend.com/.well-known/mcp.json)" · "- [Skills Discovery](https://resend.com/.well-known/agent-skills/index.json)" > — resend.com/llms.txt — the only well-known-URI agent discovery we found in the wild | Surface | What Resend ships | |---|---| | llms.txt | Fully agent-addressed; links MCP + skills discovery manifests and the OpenAPI spec | | `.well-known` manifests | `resend.com/.well-known/mcp.json` and `.well-known/agent-skills/index.json` — unique in our survey | | Markdown mirrors | Sitewide `.md`, including `pricing.md` — even the commercial pages are agent-readable | | OpenAPI | Spec served at the site root | | MCP server | Official, at github.com/resend/resend-mcp | | Agent Skills | Official skills repo; ~17.6K installs (order of magnitude) | | Docs style | Quickstart-dense; every page self-contained and runnable | The docs style is the part that moves the score. Resend's pages are quickstart-shaped: install command, imports, code, expected result — the exact snippet anatomy retrieval rubrics reward. There's no sprawling conceptual layer for a scorer to wade through, because the product itself is narrow: send email, manage domains, check delivery. The corpus is small because the product surface is small, and every snippet earns its place. The `pricing.md` detail is worth pausing on. Markdown-mirroring docs pages is now common among agent-forward products; mirroring the *pricing page* signals a different assumption — that an agent may be the one comparing vendors, not just implementing one. Resend builds for the evaluation visit, not only the integration visit. ## The receipts All figures observed 2026-06-11; single-day metrics carry ±10% error bars. **The headline:** querying Context7 for "resend" returns `/websites/resend` at benchmark **92.3** — among the highest we observed anywhere, and the top single score in our category probes. Where that sits against the field: | Entry | Snippets | Benchmark | |---|---|---| | `websites/resend` | 3,765 | **92.3** | | `websites/stripe` (largest corpus we found anywhere) | 265,284 | 84.5 | | Drizzle (`drizzle-team/drizzle-orm`) | 440 | 82.8 | | Resend SDK (`resend/resend-node`) | 712 | 80.2 | | Polar (`polarsource/polar`) | 2,297 | 64.7 | Resend beats the largest docs corpus on the internet on quality with roughly 1.4% of its volume. Even its SDK entry — just 33K tokens — outscores Polar's corpus three times its size. **Quality over size is the pattern, not the exception.** The same audit batch shows Drizzle at 82.8 on 440 snippets while Polar scores 64.7 on 2,297 — the pairwise proof that the scoring rubric measures density, not mass. Resend is the extreme point on that curve: corpus engineered down to runnable essentials, score near the ceiling. Our correlation study across 17 indexed entries found benchmark barely tracks corpus mass (ρ≈0.24–0.30) — the benchmark is an eval you can study for, and Resend studied. **The honest gap:** Resend wins its *brand* query, but the *category* query "transactional email api" has no coherent winner — the results are a jumble of Mailchimp Transactional SDKs, Listrak, ForwardEmail, and ElasticEmail, with `/resend/react-email` (benchmark 84.7) the only high-benchmark entry present. The highest quality score in the category has not yet claimed the category's task query. That's a description-engineering gap, not a docs gap. **The unique artifact:** the `.well-known/mcp.json` + `.well-known/agent-skills/index.json` pair makes Resend the only product we observed using well-known-URI conventions for agent discovery. Cost to replicate: roughly an afternoon. **Why early discovery surfaces matter more than they look:** our pilot experiment E4 found that when an agent *learns* a vendor ships agent tooling, that fact alone flipped vendor choice 2/2 (single model, pilot scale). Agent-operability has become a selection criterion — and `.well-known` manifests plus an llms.txt that lists your MCP are precisely how an agent learns it at decision time. ## What to copy - [ ] Address your llms.txt to agents explicitly and link your MCP, skills, and OpenAPI from it — make it a tool index, not a sitemap. ([Play 5](/agentic-discovery/llms-txt)) - [ ] Mirror *every* page as markdown, including pricing — agents evaluating vendors read commercial pages too. ([Play 6](/agentic-discovery/markdown-docs-for-ai-agents)) - [ ] Publish `.well-known/mcp.json` and `.well-known/agent-skills/index.json` — costless, standard-shaped, and you'd be among the first. ([Play 3](/agentic-discovery/mcp-server-distribution)) - [ ] Make every docs page a self-contained runnable unit: install + imports + code + expected output. This is what 92.3 is made of. ([Play 7](/agentic-discovery/code-snippets-for-ai-agents)) - [ ] Ship the official MCP and skills repo even while small — presence compounds. ([Play 3](/agentic-discovery/mcp-server-distribution), [Play 4](/agentic-discovery/agent-skills-and-agents-md)) - [ ] Then claim the *task* query, not just your brand query — check who wins your category's words today. ([Play 2](/agentic-discovery/ai-agent-registries-and-directories)) ## What NOT to over-copy - **Small corpus works because the product is small.** A transactional email API genuinely fits in a few thousand snippets; a database or framework doesn't. Copy the density discipline, not the corpus size. - **The 92.3 was measured on Resend's brand query.** It demonstrates retrieval *quality*, not category *demand* — Resend does not appear in our top-50 fetch-share table, and its category query remains unclaimed. A high benchmark is necessary, not sufficient. - **`.well-known` manifests are early, not proven.** Resend is the only product shipping them; we have no traffic evidence they're being read yet. Ship them because they're cheap and standards-shaped, not because we measured payoff. - **Survivorship and snapshot error.** One product, one day, ±10% error bars; benchmark scores are recomputed continuously. Install counts (~17.6K) are opt-out telemetry — order of magnitude only. ## FAQ **Why does Resend score 92.3 on Context7?** Because its corpus is engineered for the scoring rubric: quickstart-dense pages where every snippet is self-contained and runnable, no duplication, no repo noise. Context7's benchmark has a product's snippets answer common developer questions and scores the answers — small-but-complete beats large-but-padded, and Resend (92.3 observed 2026-06-11) is the cleanest example we found. **What is resend.com/.well-known/mcp.json?** A discovery manifest at a well-known URI telling agents where Resend's MCP server lives, paired with `.well-known/agent-skills/index.json` for skills. Resend is the only product we found using well-known-URI conventions for agent discovery — both files are linked directly from its llms.txt. **Does a small docs corpus hurt AI agent discovery?** Not on quality metrics — our 17-entry correlation study found benchmark scores barely track corpus size (ρ≈0.24–0.30), and Resend's small corpus posts the highest score in our probes. Size only matters insofar as it covers the questions developers actually ask; padding beyond that triggers duplication penalties. **Does Resend win the "transactional email api" query?** Not yet, as of 2026-06-11 — that category query returns no coherent winner, with `/resend/react-email` (benchmark 84.7) the only high-benchmark entry present. Resend's 92.3 lives on its brand query; claiming the task query is a description-engineering move it hasn't made. --- *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: [How DodoPayments Owns a Category Query](/agentic-discovery/case-studies/dodopayments) · Next: [Bun's CLAUDE.md Growth Hack](/agentic-discovery/case-studies/bun) → > **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)