All 20 Criteria
P1-A Structured Data — 1/5
No schema.org markup detected. Web search and page inspection returned no JSON-LD, Product, Offer, or AggregateRating schema on the homepage or pricing pages. The site has standard meta tags but nothing structured data agents can parse as authoritative offer signals. Standard SEO metadata only.
P1-B Machine-Readable Pricing — 3/5
Pricing is presented as a clean markdown table inside llms.txt — plans, credit amounts, and dollar values all machine-readable and in a single retrievable document. However, pricing is not tagged as schema.org/Offer or structured JSON, and the main pricing page appears to be a JavaScript-rendered page not accessible as raw structured data. Solid for LLM retrieval; not yet schema-grade.
P1-C llms.txt / Agent Layer — 5/5
tavily.com/llms.txt is a standout in this audit series. It contains: product descriptions and use-case decision trees ("when to use Tavily"), a full pricing table with credits and plan tiers, authentication instructions (Bearer token and OAuth 2.0), all five API endpoints with usage context, CLI install instructions, framework integrations (20+), cloud partnerships (Databricks, AWS, Azure, IBM, Snowflake), and structured external links. This is explicitly written for LLM consumption and is exactly what an agent needs to evaluate, configure, and use the product autonomously.
P1-D API / MCP Availability — 5/5
Full REST API with documented endpoints (Search, Extract, Crawl, Map, Research). Python and JavaScript SDKs (open-source on GitHub). Remote MCP server at mcp.tavily.com supporting OAuth 2.0 for seamless MCP client authentication. Agent Skills CLI for Claude Code, Codex, and Cursor. OpenAPI-spec referenced in docs. This is the most complete agent-integration stack seen in recent audits.
P1-E Discoverability (GEO) — 4/5
robots.txt explicitly names and allows major AI crawlers: ClaudeBot, GPTBot, PerplexityBot, Applebot-Extended, Google-Extended, anthropic-ai, CCBot. llms.txt is at the root domain (high discoverability). Strong developer content presence across LangChain, LlamaIndex, CrewAI, and OpenAI documentation ecosystems. Minor gap: no evidence of structured entity pages or deliberate GEO-specific content strategy beyond llms.txt.
P2-A Offer Completeness — 4/5
All five products (Search, Extract, Crawl, Map, Research), pricing tiers, credit costs, authentication, use cases, and integration options are findable in a single document (llms.txt). Rare for a company to achieve this in one authoritative source. Minor gap: the Research API pricing (async, polling-based) is less clearly defined than the others.
P2-B Scope & Limits — 4/5
Rate limits explicitly stated by environment: 100 RPM (Development, free), 1,000 RPM (Production, paid). Credit costs documented per operation type: 1 credit for basic search, 2 credits for advanced search. Plan credit quantities and monthly caps clearly stated per tier. Enterprise custom limits disclosed as custom. Well above average for API tooling.
P2-C Substitution & Fallback Rules — 1/5
No guidance found on service unavailability, fallback behavior, or what happens when an API call fails or returns empty results. Agents building production workflows cannot pre-validate Tavily's failure behavior without running the API. Standard 0/5 gap for this criteria across most audited companies.
P2-D Conditional Logic Transparency — 3/5
Enterprise plan conditions (SLA, security, custom pricing) require contacting sales — disclosed in llms.txt, not hidden. Development vs. Production environment distinction (100 RPM vs 1,000 RPM) is documented. Some conditions scattered across docs pages (credit costs by operation, plan upgrade paths). Not all conditions are machine-readable but most are disclosed.
P2-E Semantic Precision — 4/5
Specific, verifiable numbers throughout: "$0.008/credit," "100 RPM," "1,000 RPM," "1,000 credits/month free," "4,000 credits starting at $0–$99/month." Product names (Search, Extract, Crawl, Map, Research) are precisely differentiated with distinct use cases in llms.txt. Minimal vague claims — strongest at semantic precision of any recent audit.
P3-A Verifiable Performance Data — 3/5
status.tavily.com exists (dedicated status page). trust.tavily.com exists (security and compliance documentation). Insight Partners and Alpha Wave Global $25M Series A provides implicit third-party validation. No G2, Capterra, or Trustpilot presence found (expected for infrastructure/API product). Status page confirms real-time service monitoring, but no public uptime history SLA percentage found.
P3-B Scoped Permission Model — 3/5
OAuth 2.0 for MCP server is a meaningful step toward agent-scoped auth. API key via Bearer token. However, no explicit agent-scoped bounded permissions (time-bounded keys, spending-capped tokens, action-restricted scopes) documented. The OAuth 2.0 support is the strongest signal here — it enables client-specific auth without sharing master API keys.
P3-C Audit Trail / Transaction Log — 2/5
Credit usage API endpoint is documented (agents can query credit consumption). However, no machine-accessible transaction log, no audit log API, and no webhook system for consumption events found. Usage tracking is human-dashboard oriented based on available documentation.
P3-D Behavioral Consistency Signals — 2/5
Versioned API implied by docs structure. Terms of Service and Privacy Policy exist at tavily.com/terms and tavily.com/privacy. No published changelog, no explicit API version-control policy, no stated notice period for pricing or interface changes found. Platform is young (~2023 founding) — limited long-term stability track record.
P4-A Friction-Free Activation — 5/5
Free API key at app.tavily.com with no credit card required, 1,000 credits/month automatically. Two-command CLI setup: `pip install tavily-cli && tvly login`. MCP server connection is a single URL with the API key embedded. Agent Skills for Claude Code install automatically via CLI. No human gate at any point. This is textbook frictionless agent activation.
P4-B Agent Decision Signals — 4/5
Free tier provides a clear trial signal for autonomous evaluation. llms.txt explicitly lists "When to Use Tavily" — decision criteria written directly for agent consumption ("Your agent needs up-to-date information," "RAG pipelines that need grounded sources"). Minor gap: no explicit upgrade trigger signal (e.g., "upgrade when credits exhausted this month" webhook or API endpoint).
P5-A Integration Depth / Switching Cost — 4/5
Native integrations with LangChain, LlamaIndex, CrewAI, OpenAI Agents SDK, Anthropic Tool Calling, Google ADK, Pydantic AI, Vercel, n8n, Zapier, Flowise, LangFlow, Dify, Composio, and Agno. Enterprise cloud marketplace presence on Databricks, AWS, Azure, IBM watsonx, and Snowflake. Each integration embeds Tavily at the framework level, creating meaningful switching cost — removing Tavily requires reconfiguring search in every agent framework the team uses.
P5-B Agent Memory / Personalization Layer — 2/5
Credit usage is trackable via API (agents can query their own consumption history). Account-level usage dashboards exist. However, no persistent agent-specific memory layer, no context stored across sessions, no personalized search result tuning based on agent history. Interaction is stateless by design.
P5-C Programmatic Renewal Signals — 2/5
Paid plans auto-renew (standard SaaS billing). No documented API for checking subscription status, querying renewal date, or triggering plan upgrades programmatically. Agents cannot autonomously manage their own Tavily subscription without human intervention.
P5-D Compounding Value Signal — 2/5
The platform's index and retrieval quality improve over time, but no agent-readable signal communicates this. No documented API for accessing cached result quality, search relevance scores over time, or coverage improvement metrics. Value compounds silently — agents cannot detect or verify it.
Rubric v1 (April 2026). Scores reflect the company's state on the audit date and may have improved since.