Exa vs Tavily

Agent access

1. ai-native userPoint an agent at llms.txt or agent-oriented docs

   weight 2 · round to Exa
   Exa✓fullprobed9/10

   Exa hosts a real llms.txt (verified live at exa.ai/llms.txt, HTTP 200) plus a full documentation index at /docs/llms.txt, and its docs pages have .md variants (e.g. search-api-guide.md) explicitly designed for agent consumption, with agent-skills.md teaching coding agents to use Exa's APIs. This is corroborated by a live probe, not just vendor claims. Missing for 10: independent third-party confirmation that agents actually consume/parse these files successfully in practice.

   • [probe] “PROBE llms.txt: HTTP 200 at https://exa.ai/llms.txt # Exa > Exa is a real-time knowledge index and retrieval platform built for AI agents. …”
   • [probe] “PROBE docs-md: HTTP 200 at https://exa.ai/docs/reference/search-api-guide.md > ## Documentation Index > Fetch the complete documentation ind…”
   • [claimed-docs] “Return query-relevant excerpts from Exa Search results while controlling context size and latency.”
   • [claimed-docs] “Use iterative search, reasoning, and grounded synthesis for complex research tasks.”
   • [claimed-docs] “Exa skills teach coding agents how to search, retrieve content, and build with Exa's APIs.”

   proof ↗⚑ flag

   Tavily✓fullprobed8/10

   Tavily hosts a working llms.txt (confirmed live via probe, HTTP 200) with agent-oriented CLI examples, plus an agent-setup SKILL.md and MCP server docs explicitly designed for agent consumption. missing for 10: independent third-party confirmation that agents successfully consume/act on the llms.txt content beyond Tavily's own docs.

   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.tavily.com/llms.txt # Tavily Docs > Tavily is the web layer for AI agents: real-time web search, c…”
   • [claimed-docs] “tvly search "latest AI news" tvly extract "https://example.com" tvly crawl "https://docs.example.com" --depth 2 tvly research "compare React…”
   • [claimed-docs] “Read and execute tavily.com/agent-setup/SKILL.md”
   • [claimed-docs] “Let your coding agent install and configure Tavily for you.”
   • [claimed-docs] “Compatible with both Cursor and Claude Desktop! Tavily MCP is also compatible with any MCP client.”
   • [probe] “official MCP server documented at https://docs.tavily.com/documentation/mcp”

   proof ↗⚑ flag

2. ai-native userRun the product headlessly / in CI for automation

   weight 2 · round to Tavily
   Exa✓fullprobed7/10

   Exa is an API-first product (exa-js SDK, REST endpoints, webhooks, streaming) that is naturally callable headlessly from scripts/CI without any UI, and supports async batch operations like Websets and deep research suited to automation pipelines. missing for 10: no explicit CI/CD example (e.g., GitHub Actions), no dedicated 'headless/automation' documentation section, and no independent report confirming CI usage.

   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [github] “Search streaming is available via `streamSearch(...)`, which yields OpenAI-style chat completion chunks.”
   • [claimed-docs] “Async deep research, list building, and enrichment.”
   • [claimed-docs] “Get real-time HTTP callbacks as items are added or enriched”
   • [claimed-docs] “A single command creates an Exa account and syncs an API key into your project.”
   • [probe] “PROBE docs-md: HTTP 200 at https://exa.ai/docs/reference/search-api-guide.md > ## Documentation Index > Fetch the complete documentation ind…”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily is a REST API with SDKs (Python/JS) and a CLI (tvly search/extract/crawl/research) that can be run non-interactively from scripts or CI, with API keys, rate limits, and usage/logs endpoints suited to automated pipelines. missing for 10: no explicit CI/CD pipeline example (e.g., GitHub Actions) or independent report confirming headless CI usage.

   • [claimed-docs] “pip install tavily-python”
   • [claimed-docs] “tvly search "latest AI news" tvly extract "https://example.com" tvly crawl "https://docs.example.com" --depth 2 tvly research "compare React…”
   • [claimed-docs] “The Python SDK allows for easy interaction with the Tavily API, offering the full range of our search functionality directly from your Pytho…”
   • [claimed-docs] “Tavily's JavaScript SDK allows for easy interaction with the Tavily API, offering the full range of our search and extract functionalities d…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”
   • [claimed-docs] “Retrieve per-request usage logs for the API keys under your account or organization.”
   • [claimed-docs] “tavily_client = TavilyClient(api_key="tvly-YOUR_API_KEY") response = tavily_client.search("Who is Leo Messi?")”

   proof ↗⚑ flag

3. ai-native userConnect an agent via an official MCP server

   weight 3 · round to Exa
   Exa✓fullprobed9/10

   Exa ships a well-documented official MCP server with setup guides for Claude Desktop, Cursor, VS Code, ChatGPT, Codex, Grok and any MCP client, exposing web search, fetching, Exa Agent, and Exa Connect tools; it's open source and requires no API key to start. This is a first-party, dedicated server product (not an agent-client scenario), directly satisfying the story. Missing for 10: independent hands-on community verification of the MCP server specifically (community evidence covers the web/search product but not MCP integration experience).

   • [claimed-docs] “Complete setup guide for Exa MCP Server. Connect Claude Desktop, Cursor, VS Code, and 10+ AI assistants to Exa's web search, fetching, Exa A…”
   • [claimed-docs] “Exa MCP connects AI assistants to Exa’s search capabilities, including web search, code search, Exa Agent, and Exa Connect.”
   • [claimed-docs] “Connect Claude Desktop, Cursor, VS Code, and 10+ AI assistants to Exa's web search, fetching, Exa Agent, and Exa Connect tools.”
   • [claimed-docs] “Connect ChatGPT, Codex, Claude, Grok, Cursor, and any other MCP client to Exa's web search, page fetching, Exa Agent, and Exa Connect tools.”
   • [claimed-docs] “No API key is required to get started. Exa MCP is open source and available on GitHub.”
   • [probe] “official MCP server documented at https://exa.ai/docs/reference/exa-mcp”

   proof ↗⚑ flag

   Tavily✓fullprobed8/10

   Tavily documents an official MCP server providing tavily-search and tavily-extract tools, explicitly compatible with Cursor, Claude Desktop, and any MCP client, confirmed by a live docs page probe. Missing for 10: independent/hands-on third-party confirmation of MCP server reliability and no mention of the full tool set (crawl/map/research) being exposed via MCP.

   • [claimed-docs] “The Tavily MCP server provides: Seamless interaction with the tavily-search and tavily-extract tools”
   • [claimed-docs] “Compatible with both Cursor and Claude Desktop! Tavily MCP is also compatible with any MCP client.”
   • [claimed-docs] “The Tavily MCP server provides: * Seamless interaction with the tavily-search and tavily-extract tools”
   • [claimed-docs] “Compatible with both [Cursor](https://cursor.sh) and [Claude Desktop](https://claude.ai/download)! Tavily MCP is also compatible with any MC…”
   • [probe] “official MCP server documented at https://docs.tavily.com/documentation/mcp”

   proof ↗⚑ flag

4. ai-native userUse an official CLI

   weight 2 · round to Tavily
   Exa—none0/10

   Evidence covers Exa's SDKs (exa-js), MCP server, dashboard, and API docs, but no official CLI tool is mentioned anywhere in the pack. Since API-first companies commonly ship a CLI, the axis applies, but there's no evidence Exa offers one.

   ⚑ flag

   Tavily~partialclaimed3/10

   The llms.txt docs excerpt shows CLI commands like `tvly search`, `tvly extract`, `tvly crawl`, `tvly research`, suggesting an official CLI named `tvly` exists, but there is no dedicated installation guide, README, or independent corroboration beyond this single snippet. missing for 10: dedicated CLI docs/install instructions, GitHub repo or package listing for the CLI, independent/hands-on confirmation of the tool.

   • [claimed-docs] “tvly search "latest AI news" tvly extract "https://example.com" tvly crawl "https://docs.example.com" --depth 2 tvly research "compare React…”

   proof ↗⚑ flag

5. ai-native userDrive the product through a documented public API

   weight 3 · round drawn
   Exa✓fullprobed9/10

   Exa provides a well-documented public API (search, answer, websets, structured extraction) plus official SDKs (exa-js) with concrete code examples for driving it programmatically, corroborated by llms.txt/docs probes and community hands-on trial usage. missing for 10: no discoverable OpenAPI/swagger spec (404s on all probed paths), which would round out machine-readable API documentation.

   • [claimed-docs] “from ~250 ms instant search to 12-40 second deep-reasoning search”
   • [claimed-docs] “Use `output_schema` with any search type to extract structured JSON from search results”
   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [github] “const { answer } = await exa.answer("What is the capital of France?");”
   • [claimed-docs] “Our API endpoints have default rate limits to ensure reliable performance for all users.”
   • [probe] “PROBE llms.txt: HTTP 200 at https://exa.ai/llms.txt # Exa > Exa is a real-time knowledge index and retrieval platform built for AI agents. …”
   • [probe] “PROBE docs-md: HTTP 200 at https://exa.ai/docs/reference/search-api-guide.md > ## Documentation Index > Fetch the complete documentation ind…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://exa.ai/openapi.json, https://exa.ai/swagger.json, https://exa.ai/api/openapi.json, https://e…”
   • [community] “Signed up for a trial account and I'm pretty impressed with the search API (haven't used Websets yet but looks cool). Suggested adding a 'cr…”

   proof ↗⚑ flag

   Tavily✓fullprobed9/10

   Tavily is fundamentally an API-first product with extensive documented REST endpoints (search, extract, crawl, map, research) plus official Python/JS SDKs, rate-limit docs, usage/logs endpoints, and keyless trial mode — all clearly aimed at programmatic/AI-native consumption. Missing for 10: a discoverable machine-readable OpenAPI/swagger spec (probe found 404s on standard paths), which would round out full API-native tooling.

   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”
   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [claimed-docs] “The Python SDK allows for easy interaction with the Tavily API, offering the full range of our search functionality directly from your Pytho…”
   • [claimed-docs] “Tavily's JavaScript SDK allows for easy interaction with the Tavily API, offering the full range of our search and extract functionalities d…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”
   • [claimed-docs] “Get API key and account usage details”
   • [claimed-docs] “Retrieve per-request usage logs for the API keys under your account or organization.”
   • [github] “You can try Tavily without an API key. Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Ta…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

6. ai-native userIssue scoped/least-privilege API credentials for an agent

   weight 2 · round drawn
   Exa—none0/10

   No evidence of scoped or least-privilege API key/credential issuance—docs mention API key sync into a project and rate limits, but nothing about granular permissions, scopes, or restricted-access tokens for agents.

   • [claimed-docs] “A single command creates an Exa account and syncs an API key into your project.”
   • [claimed-docs] “Our API endpoints have default rate limits to ensure reliable performance for all users.”

   proof ↗⚑ flag

   Tavily—none0/10

   Evidence shows API-key authentication, project IDs and session/user identifiers for usage attribution, but nothing about issuing scoped or least-privilege credentials (e.g., read-only keys, permission-restricted keys, per-tool key scoping) for agents. missing for 10: any documentation of restricted-permission API keys, role/scope-based key creation, or agent-specific least-privilege credential issuance.

   • [claimed-docs] “You can attach a Project ID to your client to organize and track API usage by project. This is useful when a single API key is used across m…”
   • [claimed-docs] “You can attach session and user identifiers to your client for attribution and analytics across multi-step interactions and agent workflows.”
   • [claimed-docs] “Get API key and account usage details”

   proof ↗⚑ flag

7. ai-native userBuild against official SDKs

   weight 2 · round to Tavily
   Exa✓fullclaimed8/10

   Exa ships an official JS/TS SDK (exa-js) on GitHub with documented methods like webSearch(), streamSearch(), answer(), and schema-constrained search, directly supporting AI-native SDK-based integration into agent loops. missing for 10: independent/hands-on community validation of SDK quality, and evidence of official SDKs beyond JS/TS (e.g., Python) are not shown in the pack.

   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [github] “Search streaming is available via `streamSearch(...)`, which yields OpenAI-style chat completion chunks.”
   • [github] “const { answer } = await exa.answer("What is the capital of France?");”
   • [github] “For type: "object", search currently enforces: max nesting depth: 2, max total properties: 10”
   • [github] “Search streaming is available via streamSearch(...), which yields OpenAI-style chat completion chunks.”
   • [github] “Get answers with citations const { answer } = await exa.answer("What is the capital of France?");”
   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”
   • [claimed-docs] “Use `output_schema` with any search type to extract structured JSON from search results”

   proof ↗⚑ flag

   Tavily✓fullprobed9/10

   Tavily ships official Python and JavaScript SDKs with rich, documented functionality (search, extract, crawl, map, research, QnA, RAG context) via pip/npm install and quick-start guides, plus reference docs covering auth, rate limits, project/session attribution, and proxy config. Missing for 10: independent/third-party corroboration beyond vendor docs and GitHub repo, and no official OpenAPI spec was found (probe returned 404s).

   • [claimed-docs] “pip install tavily-python”
   • [claimed-docs] “The Python SDK allows for easy interaction with the Tavily API, offering the full range of our search functionality directly from your Pytho…”
   • [claimed-docs] “Tavily's JavaScript SDK allows for easy interaction with the Tavily API, offering the full range of our search and extract functionalities d…”
   • [claimed-docs] “tavily_client = TavilyClient(api_key="tvly-YOUR_API_KEY") response = tavily_client.search("Who is Leo Messi?")”
   • [claimed-docs] “response = tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [claimed-docs] “response = tavily_client.map("https://docs.tavily.com")”
   • [claimed-docs] “response = tavily_client.research("What are the latest developments in AI?")”
   • [github] “Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Tavily API. Keyless mode supports `search…”
   • [claimed-docs] “You can attach a Project ID to your client to organize and track API usage by project. This is useful when a single API key is used across m…”
   • [claimed-docs] “You can attach session and user identifiers to your client for attribution and analytics across multi-step interactions and agent workflows.”
   • [claimed-docs] “If you would like to specify a proxy to be used when making requests, you can do so by passing in a proxy parameter on client instantiation.”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

8. ai-native userSubscribe to events via webhooks

   weight 2 · round to Exa
   Exa~partialclaimed6/10

   Exa's Websets API documents real-time HTTP callbacks ('webhooks') as items are added or enriched, giving AI-native users a way to subscribe to search/enrichment events rather than polling. However, this is scoped specifically to Websets and no broader webhook infrastructure (e.g., event types, signing/security docs, retries) is evidenced for other Exa products like plain search or Agent API. Missing for 10: platform-wide webhook documentation covering multiple event types, security/verification details, and independent/hands-on confirmation of webhook reliability.

   • [claimed-docs] “Get real-time HTTP callbacks as items are added or enriched”
   • [claimed-docs] “A Webset starts with a natural-language query and a target item count. Add criteria that every result must satisfy and enrichment fields to …”
   • [claimed-docs] “Find anything on the web, no matter how complex. Websets searches, verifies, and enriches results automatically.”

   proof ↗⚑ flag

   Tavily—none0/10

   Tavily's evidence pack covers search, extract, crawl, map, research APIs, SDKs, and an MCP server, but there is no mention of webhooks or event subscription mechanisms anywhere in the docs, GitHub, or probes; async research status must be polled via request ID rather than pushed via webhook.

   • [claimed-docs] “Retrieve the status and results of a research task using its request ID.”
   • [claimed-docs] “When using the Tavily Research API, you can stream responses in real-time by setting `stream: true` in your request.”
   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.tavily.com/llms.txt # Tavily Docs > Tavily is the web layer for AI agents: real-time web search, c…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

Agentic features

1. ai-native userOperate the product with natural-language commands

   weight 2 · round to Exa
   Exa✓fullcommunity8/10

   Exa's core search API accepts natural-language queries directly (e.g. 'agtech companies in the US that raised Series A'), Websets are built from a natural-language query plus criteria, and Exa's MCP server lets AI assistants like Claude/Cursor/ChatGPT operate it purely via NL commands without code. This is well documented across docs and GitHub examples. missing for 10: independent hands-on verification that NL commands reliably map to correct results (community reports show mixed accuracy, e.g. missed price enrichment and slow/stuck verification), and no first-party benchmark of NL-command success rate.

   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [claimed-docs] “A Webset starts with a natural-language query and a target item count. Add criteria that every result must satisfy and enrichment fields to …”
   • [claimed-docs] “Give it a query like "agtech companies in the US that raised Series A" and it will search, verify each result against your criteria, and enr…”
   • [claimed-docs] “Complete setup guide for Exa MCP Server. Connect Claude Desktop, Cursor, VS Code, and 10+ AI assistants to Exa's web search, fetching, Exa A…”
   • [claimed-docs] “Connect ChatGPT, Codex, Claude, Grok, Cursor, and any other MCP client to Exa's web search, page fetching, Exa Agent, and Exa Connect tools.”
   • [github] “Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.”
   • [community] “Searched 'lucid air touring models available for sale under 20,000 miles' and tried to add a 'sale price' column, but did not get the price …”
   • [community] “I gave it a try and my first search got one match, 14 misses, and all other results are 'Verifying...' but it seems stuck (been minutes)... …”

   proof ↗⚑ flag

   Tavily~partialprobed6/10

   Tavily's API accepts natural-language queries (search, research, crawl instructions) and its MCP server plus CLI (tvly search/extract/crawl/research) let AI-native users invoke it via natural-language-like commands rather than rigid syntax; instructions like crawl's 'Find all pages on the Python SDK' show NL-driven parameterization. missing for 10: no evidence of a true conversational/chat interface parsing arbitrary free-form user commands end-to-end, and no independent/hands-on verification beyond vendor docs and github readme.

   • [claimed-docs] “tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [claimed-docs] “response = tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [claimed-docs] “tvly search "latest AI news" tvly extract "https://example.com" tvly crawl "https://docs.example.com" --depth 2 tvly research "compare React…”
   • [claimed-docs] “The Tavily MCP server provides: Seamless interaction with the tavily-search and tavily-extract tools”
   • [probe] “official MCP server documented at https://docs.tavily.com/documentation/mcp”
   • [github] “Tavily Research Research lets you create comprehensive research reports on any topic, with aut”

   proof ↗⚑ flag

Api quality

1. ai-native userExplore an interactive API reference with runnable examples

   weight 2 · round to Tavily
   Exa!disputedcontradicted4/10

   Exa's docs claim a 'Get Code' feature that generates a 'complete, tested integration snippet' in under a minute (exa-docs-11), and the JS SDK provides runnable code examples (exa-gh-1-5), but there's no evidence of a true interactive API reference (OpenAPI/Swagger endpoints all 404'd per exa-probe-3), and a hands-on reviewer found the generated cURL example in 'Get Code' to be 'demonstrably wrong' (exa-comm-4), directly contradicting the 'tested' claim. missing for 10: a genuine try-it-out API console, working/tested generated snippets, independent confirmation the code examples run correctly.

   • [claimed-docs] “it generates a complete, tested integration snippet tailored to your exact stack and use case in under a minute”
   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [community] “Tried the GitHub repo search feature: it was about 30/70 on finding the things I asked for. The cURL example in 'Get Code' was demonstrably …”
   • [probe] “PROBE openapi: all candidate paths 404 (https://exa.ai/openapi.json, https://exa.ai/swagger.json, https://exa.ai/api/openapi.json, https://e…”

   proof ↗⚑ flag

   Tavily~partialprobed5/10

   Tavily has a documented API reference with endpoint pages (search, extract, crawl, map, research) and code snippets, but there's no confirmed interactive/runnable console—an openapi.json probe returned 404s across candidate paths, suggesting no standard interactive spec-driven explorer. missing for 10: evidence of an actual interactive 'try it' console or sandbox with live runnable requests, a discoverable OpenAPI/Swagger spec, and independent confirmation of interactivity beyond static code examples.

   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”
   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

2. ai-native userDownload a machine-readable API spec (OpenAPI or equivalent)

   weight 2 · round drawn
   Exa—none0/10

   A direct probe for an OpenAPI/Swagger spec at all standard locations (openapi.json, swagger.json, etc.) returned 404s, and while Exa offers llms.txt and markdown doc indexes, these are documentation aids, not a machine-readable API schema (endpoints/params/types) that a buyer could import into tooling like Postman or codegen.

   • [probe] “PROBE openapi: all candidate paths 404 (https://exa.ai/openapi.json, https://exa.ai/swagger.json, https://exa.ai/api/openapi.json, https://e…”
   • [probe] “PROBE llms.txt: HTTP 200 at https://exa.ai/llms.txt # Exa > Exa is a real-time knowledge index and retrieval platform built for AI agents. …”
   • [probe] “PROBE docs-md: HTTP 200 at https://exa.ai/docs/reference/search-api-guide.md > ## Documentation Index > Fetch the complete documentation ind…”

   proof ↗⚑ flag

   Tavily—none0/10

   Tavily has detailed API reference docs but an explicit probe for standard OpenAPI/Swagger spec locations (openapi.json, swagger.json, etc.) returned 404s, and no evidence pack item points to a downloadable machine-readable spec.

   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

3. ai-native userRely on versioned APIs with a documented deprecation policy

   weight 2 · round drawn
   Exa—none0/10

   No evidence pack item mentions API versioning scheme (e.g., v1/v2 endpoints) or any documented deprecation policy for older API versions; docs cover rate limits, pricing, and features but not version lifecycle management.

   ⚑ flag

   Tavily—none0/10

   No evidence of API versioning scheme (e.g., v1/v2 paths) or any documented deprecation policy; OpenAPI spec probes returned 404s and no changelog/deprecation docs appear in the evidence pack.

   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

1. ai-native userPerform bulk operations across many items at once

   weight 2 · round to Tavily
   Exa~partialcommunity7/10

   Exa's Websets product is explicitly built for bulk operations — 'build large lists', verify and enrich every match against criteria, async list building/enrichment, and scheduled searches for new items (exa-docs-7,8,9,10,15,28). However, hands-on community reports show real gaps: missing enrichment data (price columns), 'Verifying...' stuck states, and inconsistent match rates (exa-comm-2,3,4), indicating the bulk pipeline doesn't always complete reliably at scale. Missing for 10: independent benchmarks confirming reliability at large scale, and resolution of the noted enrichment/verification failures.

   • [claimed-docs] “You can also build websets visually in the [Dashboard](/docs/websets/dashboard/get-started), no code required.”
   • [claimed-docs] “Give it a query like "agtech companies in the US that raised Series A" and it will search, verify each result against your criteria, and enr…”
   • [claimed-docs] “Async deep research, list building, and enrichment.”
   • [claimed-docs] “Scheduled searches that surface new events on the web.”
   • [claimed-docs] “Find anything on the web, no matter how complex. Websets searches, verifies, and enriches results automatically.”
   • [claimed-docs] “A Webset starts with a natural-language query and a target item count. Add criteria that every result must satisfy and enrichment fields to …”
   • [community] “Searched 'lucid air touring models available for sale under 20,000 miles' and tried to add a 'sale price' column, but did not get the price …”
   • [community] “I gave it a try and my first search got one match, 14 misses, and all other results are 'Verifying...' but it seems stuck (been minutes)... …”
   • [community] “Tried the GitHub repo search feature: it was about 30/70 on finding the things I asked for. The cURL example in 'Get Code' was demonstrably …”

   proof ↗⚑ flag

   Tavily✓fullclaimed7/10

   Tavily explicitly supports bulk operations: Extract can batch up to 20 URLs per request, and Crawl/Map can traverse and explore hundreds of paths in parallel graph-based traversal, plus Research performs multi-search comprehensive reports with async status retrieval. These directly satisfy 'bulk operations across many items' for an AI-native/automation use case. Missing for 10: no documented bulk/batch search across many queries in one call, and no independent hands-on benchmark confirming throughput at scale beyond vendor docs.

   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”
   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [claimed-docs] “Retrieve the status and results of a research task using its request ID.”
   • [github] “Crawl lets you traverse a website's content starting from a base URL.”
   • [github] “Tavily Research Research lets you create comprehensive research reports on any topic, with aut”

   proof ↗⚑ flag

Citation fidelity

1. developerRely on results carrying canonical URLs, titles, and published dates so my app can cite sources accurately

   weight 3 · round to Exa
   Exa~partialclaimed5/10

   Exa's API supports date filtering (startPublishedDate) and domain/path targeting, and its answer/citation feature returns sourced URLs, implying results carry URLs and dates, but no evidence pack item explicitly documents the search result schema (title, canonical url, publishedDate fields) or addresses duplicate/canonicalization handling for citation accuracy. Missing for 10: explicit schema documentation showing title/url/publishedDate fields returned per result, evidence of canonical URL deduplication, and independent confirmation that dates/titles are consistently accurate.

   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”
   • [github] “Get answers with citations const { answer } = await exa.answer("What is the capital of France?");”
   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [claimed-docs] “Exa Contents returns clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically.”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence describes Tavily's search/extract/crawl endpoints and SDKs but never documents the actual result schema (URL, title, published date fields) returned by search results, so there's no evidence supporting reliable citation metadata for developers to build accurate citations.

   ⚑ flag

2. data engineerRead a documented stance on robots.txt, content licensing, and permitted use of results

   weight 2 · round drawn
   Exa—none0/10

   No first-party documentation addresses robots.txt policy, content licensing, or permitted use of search results; the only related evidence is a community tweet alleging Exa ignores robots.txt when indexing (exa-comm-9), which is a complaint, not a documented policy. Since Exa never publishes an explicit stance, this is an unmet, applicable axis rather than a vendor claim being disputed.

   • [community] “Exa.ai is indexing personal site data ignoring robots.txt”

   proof ↗⚑ flag

   Tavily—none0/10

   No evidence pack item documents Tavily's stance on robots.txt compliance, content licensing of crawled/extracted data, or permitted use of search/crawl results; only a general note about security/privacy validation layers (tavily-docs-18) touches trust but not this specific axis.

   ⚑ flag

Ai consumption

1. ai-native userHave an agent run a multi-step research flow — search, extract contents, and synthesize — through one API

   weight 2 · round drawn
   Exa✓fullcommunity8/10

   Exa's Agent API and Contents/Answer endpoints explicitly support a single-API multi-step flow: search, extract clean content, and synthesize a grounded answer with citations (exa-docs-25, exa-docs-26, exa-docs-24, exa-gh-3/8), and the search endpoint can even return contents inline (exa-docs-22, exa-docs-35). Community feedback shows mixed real-world reliability (e.g., stuck verification, incomplete enrichment) which tempers confidence. Missing for 10: independent hands-on validation of the full search→extract→synthesize loop working end-to-end without manual chaining, and clearer benchmarking of synthesis quality.

   • [claimed-docs] “Use iterative search, reasoning, and grounded synthesis for complex research tasks.”
   • [claimed-docs] “Enrich entities, build large lists, and run deep research from a single API with built in structured results and web citation”
   • [claimed-docs] “Exa Contents returns clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically.”
   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [github] “const { answer } = await exa.answer("What is the capital of France?");”
   • [github] “Get answers with citations const { answer } = await exa.answer("What is the capital of France?");”
   • [claimed-docs] “All contents features are also available in Exa Search for returned URLs, at no extra charge up to 10 results per search”
   • [community] “I gave it a try and my first search got one match, 14 misses, and all other results are 'Verifying...' but it seems stuck (been minutes)... …”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily's single API/SDK exposes search, extract, crawl/map, and a dedicated research endpoint that itself performs multiple searches, source analysis, and report synthesis (tavily-docs-4,13,26,27,29), all accessible through one client/API key. This directly matches a multi-step search→extract→synthesize research flow callable by an agent. Missing for 10: independent/hands-on third-party validation of the research endpoint's synthesis quality and no evidence of chaining search+extract+research in a single documented end-to-end example.

   • [claimed-docs] “tavily_client.research("What are the latest developments in AI?")”
   • [claimed-docs] “response = tavily_client.research("What are the latest developments in AI?")”
   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [claimed-docs] “"mini" is optimized for targeted, efficient research and works best for narrow or well-scoped questions. "pro" provides comprehensive, multi…”
   • [claimed-docs] “When using the Tavily Research API, you can stream responses in real-time by setting `stream: true` in your request.”
   • [claimed-docs] “Extract web page content from one or more specified URLs using Tavily Extract.”
   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [github] “Tavily Research Research lets you create comprehensive research reports on any topic, with aut”

   proof ↗⚑ flag

Page contents

1. developerRetrieve full page text or markdown for results, not just snippets

   weight 3 · round to Tavily
   Exa✓fullclaimed7/10

   Exa's Contents API explicitly returns 'clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically,' and this full-content extraction is also bundled into Search results (up to 10 per search) rather than being limited to snippets, with a token-efficient full-text mode described as '10x token efficient extracts.' Missing for 10: explicit documentation of markdown as an output format option, and independent/hands-on developer confirmation of full-text/markdown extraction quality.

   • [claimed-docs] “Exa Contents returns clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically.”
   • [claimed-docs] “All contents features are also available in Exa Search for returned URLs, at no extra charge up to 10 results per search”
   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [claimed-docs] “10x token efficient extracts of only the relevant tokens from a webpage. 4000 characters recommended.”
   • [claimed-docs] “Specialized model trained to extract the most relevant excerpts from the web”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily Extract explicitly retrieves full page content (not just snippets) from one or more URLs, batching up to 20 per request, and is documented as a distinct endpoint separate from Search snippets; SDKs and MCP server expose this extract capability directly. Missing for 10: explicit confirmation of markdown-formatted output vs raw text, and independent third-party verification of extraction fidelity.

   • [claimed-docs] “Extract web page content from one or more specified URLs using Tavily Extract.”
   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [github] “Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Tavily API. Keyless mode supports `search…”
   • [claimed-docs] “The Tavily MCP server provides: * Seamless interaction with the tavily-search and tavily-extract tools”

   proof ↗⚑ flag

2. developerGet structured output matching a JSON schema I define

   weight 2 · round to Exa
   Exa✓fullclaimed7/10

   Exa's docs explicitly describe an `output_schema` parameter usable with any search type to extract structured JSON from results, and the GitHub SDK documents concrete schema constraints (max nesting depth 2, max 10 properties) confirming real enforcement of a JSON schema. Missing for 10: independent/hands-on confirmation that schema-constrained output works reliably in practice and richer documentation of schema validation/error handling.

   • [claimed-docs] “Use `output_schema` with any search type to extract structured JSON from search results”
   • [github] “For type: "object", search currently enforces: max nesting depth: 2, max total properties: 10”

   proof ↗⚑ flag

   Tavily—none0/10

   No evidence in the pack mentions a JSON-schema-based structured output feature, response schema parameter, or Pydantic/type-safe output validation for search/extract/crawl/research endpoints; responses appear to be fixed JSON shapes rather than developer-defined schemas.

   ⚑ flag

3. data engineerCrawl subpages or entire sites starting from a search result

   weight 1 · round to Tavily
   Exa~partialclaimed5/10

   Exa's Contents API and domain-scoping options (includeDomains/excludeDomains with path prefixes, 'limit results to an entire domain or a specific section of a site') let engineers restrict search/content retrieval to a site or subsection, and Exa Contents cleanly extracts page content including JS-rendered pages and PDFs, which supports basic multi-page extraction from a domain. However there is no explicit documentation of a dedicated 'crawl subpages from a search result' or recursive site-crawl parameter, and no hands-on evidence of crawling an entire site tree. Missing for 10: explicit subpage/recursive-crawl API parameter, and independent verification that full-site crawling from a single result works reliably.

   • [claimed-docs] “Limit results to an entire domain or a specific section of a site.”
   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes such”
   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes”
   • [claimed-docs] “Exa Contents returns clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically.”
   • [claimed-docs] “All contents features are also available in Exa Search for returned URLs, at no extra charge up to 10 results per search”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily Crawl explicitly traverses a website's content starting from a base URL with graph-based parallel exploration and intelligent discovery, and Map generates full site maps similarly, both documented with SDK examples (tavily_client.crawl(...), tavily_client.map(...)). This directly satisfies crawling subpages/entire sites from a starting URL. Missing for 10: no independent/hands-on verification beyond vendor docs and GitHub README, and no explicit example of crawling starting directly from a search result URL (only from an arbitrary base URL).

   • [github] “Crawl lets you traverse a website's content starting from a base URL.”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [claimed-docs] “response = tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [claimed-docs] “tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [github] “Map lets you discover and visualize the structure of a website starting from a base URL.”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”
   • [claimed-docs] “response = tavily_client.map("https://docs.tavily.com")”

   proof ↗⚑ flag

Query filters

1. developerRestrict or exclude specific domains from my search results

   weight 2 · round to Exa
   Exa✓fullclaimed9/10

   Exa's API explicitly supports both includeDomains and excludeDomains parameters (accepting domains and path prefixes), documented in the search API guide and demonstrated in the exa-js GitHub README with a concrete includeDomains example. This directly delivers domain restriction/exclusion for developers via a first-party, well-documented parameter with SDK code example. Missing for 10: no independent/hands-on confirmation specifically of excludeDomains behavior in practice.

   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes such”
   • [claimed-docs] “Limit results to an entire domain or a specific section of a site.”
   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes”
   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence pack never mentions domain include/exclude filtering (no 'include_domains' or 'exclude_domains' parameters, or any similar filter) anywhere in the docs, SDK, or API reference excerpts provided; it only covers exact_match, crawl/map depth, research modes, and rate limits.

   ⚑ flag

2. developerRead per-result relevance scores to threshold what enters my pipeline

   weight 1 · round drawn
   Exa—none0/10

   No evidence pack item mentions a per-result relevance/similarity score field, score-based thresholding, or any API parameter to filter/sort results by a numeric relevance score. This is a reasonable axis for a search API, but nothing in docs, SDK, or community evidence documents such a score being returned or usable for pipeline thresholding.

   ⚑ flag

   Tavily—none0/10

   The evidence pack describes Tavily's search, extract, crawl, map, and research endpoints but never mentions a relevance/score field returned per result that a developer could use to threshold results. Missing for 10: any documentation of a numeric relevance score in the search response schema, and any guidance on filtering/thresholding by that score.

   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [github] “Use exact_match=True to only return results containing the exact phrase(s) inside quotes”

   proof ↗⚑ flag

Freshness

1. data engineerGet fresh results with date-range and recency filters backed by a frequently updated index

   weight 3 · round to Exa
   Exa✓fullprobed7/10

   Exa's API supports date-range filtering (startPublishedDate/endPublishedDate) as shown in the JS SDK example, and Exa positions itself as a 'real-time knowledge index' with scheduled searches that surface new events on the web, indicating recency-focused indexing and monitoring capabilities. missing for 10: explicit documentation of crawl/index refresh frequency or independent benchmarks validating recency/freshness claims beyond marketing language.

   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”
   • [claimed-docs] “Scheduled searches that surface new events on the web.”
   • [probe] “PROBE llms.txt: HTTP 200 at https://exa.ai/llms.txt # Exa > Exa is a real-time knowledge index and retrieval platform built for AI agents. …”
   • [claimed-docs] “Async deep research, list building, and enrichment.”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence pack covers Tavily's search, extract, crawl, map, and research endpoints, rate limits, and MCP integration, but contains no mention of date-range parameters, recency/freshness filters, or index update frequency. Missing for 10: explicit date-range/time_range search parameters, recency filter documentation, and any claim about index freshness or update cadence.

   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [github] “Use exact_match=True to only return results containing the exact phrase(s) inside quotes”
   • [claimed-docs] “tavily_client.search("Who is Leo Messi?")”

   proof ↗⚑ flag

2. data engineerQuery dedicated verticals like news, images, or finance through specific endpoints or category filters

   weight 2 · round to Exa
   Exa~partialclaimed4/10

   Exa's docs mention custom indexes for people, companies, and publications and domain/date filtering (includeDomains, startPublishedDate), suggesting some vertical-style filtering, but there is no explicit evidence of dedicated endpoints or category filters for news, images, or finance specifically. Missing for 10: explicit 'category' parameter documentation covering news/images/finance verticals, examples of querying these specific verticals, and independent confirmation of vertical-specific accuracy.

   • [claimed-docs] “Exa has custom indexes of 1B+ people, 50M+ companies, 350M+ publications (research papers, preprints, and more), and more.”
   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”
   • [claimed-docs] “Limit results to an entire domain or a specific section of a site.”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence pack describes Tavily's generic search, extract, crawl, map, and research endpoints, but nowhere mentions vertical-specific categories or filters such as news, images, or finance. Missing for 10: any documentation of a topic/category parameter (e.g., topic=news/finance), dedicated vertical endpoints, or image-specific search options.

   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [claimed-docs] “tavily_client.search("Who is Leo Messi?")”
   • [github] “Map lets you discover and visualize the structure of a website starting from a base URL.”

   proof ↗⚑ flag

Targeting

1. data engineerLocalize results by country, language, or location parameters

   weight 2 · round drawn
   Exa—none0/10

   The evidence pack documents domain include/exclude filters and date filters (e.g., includeDomains, startPublishedDate) but contains no mention of country, language, or geographic location search parameters anywhere in Exa's API or docs. Missing for 10: any documented country/locale/language filter parameter, geo-targeting options, or localization examples.

   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes such”
   • [claimed-docs] “Limit results to an entire domain or a specific section of a site.”
   • [claimed-docs] “Both `includeDomains` and `excludeDomains` accept domains, path prefixes”
   • [github] “includeDomains: ["nasa.gov", "space.com"], startPublishedDate: "2024-01-01",”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence pack documents search, extract, crawl, map, research, and MCP capabilities but never mentions country, language, or geo-location filtering/localization parameters for search results. missing for 10: any documented country/language/location parameter, evidence of localized result sets, or docs referencing geographic targeting.

   ⚑ flag

1. ai-native userDo everything through the API that I can do in the UI

   weight 2 · round to Tavily
   Exa~partialprobed5/10

   Exa's docs show that Websets, its dashboard feature, can be built visually with 'no code required' (exa-docs-7/29) and that the same functionality is exposed via the Websets API (exa-docs-8/15/28), suggesting rough UI/API parity for that feature, and search/contents/answer are all API-first with UI as a thin layer (exa-comm-7). However there is no explicit documentation stating full feature parity between the dashboard and API, and no OpenAPI spec was found (exa-probe-3) to verify completeness. Missing for 10: an explicit parity statement or changelog showing every UI capability (e.g. dashboard analytics, billing controls, websets monitoring UI) is mirrored in the API, and independent confirmation that no UI-only features exist.

   • [claimed-docs] “You can also build websets visually in the [Dashboard](/docs/websets/dashboard/get-started), no code required.”
   • [claimed-docs] “Give it a query like "agtech companies in the US that raised Series A" and it will search, verify each result against your criteria, and enr…”
   • [claimed-docs] “Find anything on the web, no matter how complex. Websets searches, verifies, and enriches results automatically.”
   • [claimed-docs] “A Webset starts with a natural-language query and a target item count. Add criteria that every result must satisfy and enrichment fields to …”
   • [claimed-docs] “You can also build websets visually in the Dashboard, no code required.”
   • [community] “Exa was originally just a search engine. They try to hide it these days to promote Websets, but you can still use it at exa.ai/search.”
   • [probe] “PROBE openapi: all candidate paths 404 (https://exa.ai/openapi.json, https://exa.ai/swagger.json, https://exa.ai/api/openapi.json, https://e…”

   proof ↗⚑ flag

   Tavily✓fullprobed7/10

   Tavily's core functionality (search, extract, crawl, map, research, Q&A, RAG context) is fully exposed through the API/SDKs, and CLI/MCP integrations mirror this, suggesting the API is the primary interface with no UI-only features documented. However, there's no explicit comparison confirming full UI-API parity, and no public OpenAPI spec was found (404 on all probed paths), leaving some doubt about completeness of self-service API discovery. Missing for 10: explicit UI feature list to compare against API, and a public OpenAPI/swagger spec for full endpoint parity verification.

   • [claimed-docs] “tavily_client.search("Who is Leo Messi?")”
   • [claimed-docs] “Extract web page content from one or more specified URLs using Tavily Extract.”
   • [claimed-docs] “tavily_client.crawl("https://docs.tavily.com", instructions="Find all pages on the Python SDK")”
   • [claimed-docs] “tavily_client.research("What are the latest developments in AI?")”
   • [github] “Map lets you discover and visualize the structure of a website starting from a base URL.”
   • [github] “context = tavily_client.get_search_context(query="What happened during the Burning Man floods?")”
   • [github] “answer = tavily_client.qna_search(query="Who is Leo Messi?")”
   • [claimed-docs] “Execute a search query using Tavily Search.”
   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”
   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [claimed-docs] “tvly search "latest AI news" tvly extract "https://example.com" tvly crawl "https://docs.example.com" --depth 2 tvly research "compare React…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

2. ai-native userExport all of my data in open formats and leave

   weight 3 · round drawn
   Exa—none0/10

   Exa's docs describe search, contents, websets, and MCP integration, but nowhere is there evidence of a data export feature (e.g., bulk export of websets/results in open formats like CSV/JSON) or account data portability for users to 'take their data and leave'. This is a fair axis for a data platform, but no evidence supports it.

   ⚑ flag

   Tavily—none0/10

   Tavily's docs show usage/logs endpoints (tavily-docs-30, tavily-docs-31) for tracking API usage, but there is no evidence of a comprehensive data export feature, open-format data dumps, or account data portability mechanism that would let a user export all their data and leave the platform.

   • [claimed-docs] “Get API key and account usage details”
   • [claimed-docs] “Retrieve per-request usage logs for the API keys under your account or organization.”

   proof ↗⚑ flag

3. ai-native userRead the product's source under an open license

   weight 2 · round to Exa
   Exa~partialclaimed3/10

   Exa's core search/retrieval engine and backend are proprietary SaaS; only peripheral components are open — the Exa MCP server is explicitly stated to be 'open source and available on GitHub' (exa-docs-31) and the JS client SDK lives on GitHub (exa-gh-1), but these are thin client/integration layers, not the product's actual search index or ranking source. missing for 10: no open license for the core search engine/index itself, no repository or license file covering the main product, no independent confirmation of license terms for the open pieces.

   • [claimed-docs] “No API key is required to get started. Exa MCP is open source and available on GitHub.”
   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [claimed-docs] “Exa MCP connects AI assistants to Exa’s search capabilities, including web search, code search, Exa Agent, and Exa Connect.”

   proof ↗⚑ flag

   Tavily—none0/10

   Tavily is a closed, hosted API/SaaS service; while a Python SDK client repo exists on GitHub, the evidence never states any open-source license or exposes the core Tavily service/backend source code. Missing for 10: explicit license grant, access to backend/service source, any statement of open-sourcing the core product.

   • [github] “Map lets you discover and visualize the structure of a website starting from a base URL.”
   • [github] “Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Tavily API. Keyless mode supports `search…”

   proof ↗⚑ flag

4. ai-native userSelf-host the core product

   weight 3 · round drawn
   Exa—none0/10

   Exa is presented as a hosted SaaS search/index platform with proprietary crawled indexes; only the thin MCP connector (exa-docs-31) is open source, not the core search/retrieval engine or index. No evidence anywhere of a self-hostable core product, on-prem deployment, or open-source release of the underlying search infrastructure.

   • [claimed-docs] “No API key is required to get started. Exa MCP is open source and available on GitHub.”
   • [probe] “PROBE llms.txt: HTTP 200 at https://exa.ai/llms.txt # Exa > Exa is a real-time knowledge index and retrieval platform built for AI agents. …”
   • [claimed-docs] “Exa has custom indexes of 1B+ people, 50M+ companies, 350M+ publications (research papers, preprints, and more), and more.”

   proof ↗⚑ flag

   Tavily—none0/10

   Tavily is offered exclusively as a hosted API/SaaS product (API keys, rate limits, pricing tiers, keyless mode against the 'public Tavily API') with no evidence of a self-hostable core engine, on-prem deployment option, or open-source server code.

   • [claimed-docs] “1,000 API credits / month ... No credit card required”
   • [github] “Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Tavily API. Keyless mode supports `search…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”

   proof ↗⚑ flag

Pricing clarity

1. developerRead documented rate limits and concurrency caps with a self-serve upgrade path

   weight 2 · round to Tavily
   Exa~partialclaimed4/10

   Exa has a dedicated rate-limits doc page (exa-docs-16) confirming default per-endpoint limits exist, and self-serve signup/API-key provisioning with credits is documented (exa-docs-33, exa-docs-34), implying a self-serve path to get and fund an account. However, no evidence shows actual numeric rate limits, concurrency caps, or a documented tier-based upgrade mechanism (e.g., paid plan comparison unlocking higher throughput). Missing for 10: concrete rate-limit/concurrency numbers, documented upgrade tiers and how limits scale with them, and any independent confirmation of these limits in practice.

   • [claimed-docs] “Our API endpoints have default rate limits to ensure reliable performance for all users.”
   • [claimed-docs] “A single command creates an Exa account and syncs an API key into your project.”
   • [claimed-docs] “$20 credits on sign-up with $10 credits every month. No payment method required”

   proof ↗⚑ flag

   Tavily~partialclaimed6/10

   Tavily's docs explicitly state RPM caps for Development (100) vs Production (1,000) environments, describe 429/retry-after behavior, and the pricing page shows a self-serve free tier and usage-based paid plans with a usage/logs API to track consumption — together implying a self-serve upgrade path from dev to production limits. However, there is no explicit concurrency-cap documentation or a clearly described step-by-step upgrade workflow beyond the environment distinction. Missing for 10: explicit concurrency limits, a documented step-by-step self-serve upgrade/billing flow.

   • [claimed-docs] “the API returns a `429 Too Many Requests` response with a `retry-after` header indicating the number of seconds to wait before making anothe…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”
   • [claimed-docs] “`Production` 1,000”
   • [claimed-docs] “1,000 API credits / month No credit card required”
   • [claimed-docs] “Get API key and account usage details”
   • [claimed-docs] “Retrieve per-request usage logs for the API keys under your account or organization.”
   • [claimed-docs] “1,000 API credits / month ... No credit card required”
   • [claimed-docs] “You get 1,000 free API Credits every month. **No credit card required.**”

   proof ↗⚑ flag

2. developerPrototype against the API on a free tier or free credits

   weight 2 · round to Tavily
   Exa✓fullcommunity8/10

   Exa's pricing docs explicitly state $20 free credits on sign-up plus $10/month recurring credits with no payment method required, directly enabling free prototyping, and community evidence confirms users signing up for trial accounts and testing the API without payment friction (exa-comm-1). missing for 10: no independent long-term verification of ongoing $10/month credit renewal or explicit free-tier rate limits beyond generic rate-limit doc.

   • [claimed-docs] “$20 credits on sign-up with $10 credits every month. No payment method required”
   • [claimed-docs] “A single command creates an Exa account and syncs an API key into your project.”
   • [community] “Signed up for a trial account and I'm pretty impressed with the search API (haven't used Websets yet but looks cool). Suggested adding a 'cr…”
   • [claimed-docs] “Our API endpoints have default rate limits to ensure reliable performance for all users.”

   proof ↗⚑ flag

   Tavily✓fullclaimed9/10

   Tavily offers 1,000 free API credits/month with no credit card required, plus a keyless SDK mode for search/extract to try before signing up, and clear rate limits for development tier prototyping. missing for 10: no independent/community corroboration of the free-tier experience beyond vendor docs.

   • [claimed-docs] “1,000 API credits / month ... No credit card required”
   • [claimed-docs] “1,000 API credits / month No credit card required”
   • [claimed-docs] “You get 1,000 free API Credits every month. **No credit card required.**”
   • [github] “Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Tavily API. Keyless mode supports `search…”
   • [github] “You can try Tavily without an API key. Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Ta…”
   • [github] “You can try Tavily without an API key. Instantiate `TavilyClient()` with no arguments and the SDK runs in keyless mode against the public Ta…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”

   proof ↗⚑ flag

3. developerSee clear per-request or per-credit pricing without talking to sales

   weight 3 · round drawn
   Exa~partialclaimed6/10

   Exa's pricing page (exa-docs-34) shows self-serve credit-based pricing ($20 signup credit, $10/month, no payment method required) and a dedicated /docs/reference/pricing page exists (exa-docs-9, exa-docs-10) listing feature tiers, indicating pricing is public and requires no sales contact. However, the evidence pack doesn't show a detailed per-endpoint/per-request cost table or explicit dollar-per-call rates, so the granularity a developer needs to estimate exact costs is unclear. Missing for 10: explicit per-request/per-token cost breakdown, confirmation that all endpoints (websets, agent, MCP) are covered in the same public pricing page, independent confirmation that pricing is fully self-serve without enterprise gating.

   • [claimed-docs] “$20 credits on sign-up with $10 credits every month. No payment method required”
   • [claimed-docs] “Async deep research, list building, and enrichment.”
   • [claimed-docs] “Scheduled searches that surface new events on the web.”

   proof ↗⚑ flag

   Tavily~partialclaimed6/10

   Tavily's public pricing page confirms a self-serve, no-sales-required model: 1,000 free API credits/month with no credit card required and a 'pay only for what you use' model. However, the evidence pack does not show an explicit per-request/per-credit cost table or endpoint-specific credit pricing breakdown, so the exact cost transparency is only partially demonstrated. Missing for 10: a detailed per-endpoint credit-cost table, tier pricing breakdown beyond the free tier.

   • [claimed-docs] “1,000 API credits / month ... No credit card required”
   • [claimed-docs] “Pay only for what you use”
   • [claimed-docs] “1,000 API credits / month No credit card required”
   • [claimed-docs] “You get 1,000 free API Credits every month. **No credit card required.**”

   proof ↗⚑ flag

Scale throughput

1. data engineerSubmit batched or asynchronous jobs for high-volume query workloads

   weight 2 · round to Exa
   Exa✓fullclaimed7/10

   Exa's docs describe Websets for async list-building/enrichment at scale, async deep research, scheduled searches, and webhook callbacks for job status — directly supporting asynchronous high-volume workloads, backed by rate-limit documentation for query volume management. Missing for 10: an explicit bulk/batch query submission endpoint (vs. Websets-specific async jobs), and independent/hands-on benchmarks confirming reliable performance at high volume.

   • [claimed-docs] “Async deep research, list building, and enrichment.”
   • [claimed-docs] “Scheduled searches that surface new events on the web.”
   • [claimed-docs] “Get real-time HTTP callbacks as items are added or enriched”
   • [claimed-docs] “A Webset starts with a natural-language query and a target item count. Add criteria that every result must satisfy and enrichment fields to …”
   • [claimed-docs] “Our API endpoints have default rate limits to ensure reliable performance for all users.”

   proof ↗⚑ flag

   Tavily~partialclaimed6/10

   Tavily's Research endpoint supports async job submission with request-ID based status polling and streaming, and Extract supports batching up to 20 URLs per request; rate-limit docs show RPM tiers and 429/retry-after handling for high-volume workloads. However there's no dedicated bulk/batch job queue for Search itself, no webhook/callback pattern for large offline jobs, and no explicit high-volume batch pricing guidance beyond per-credit usage. Missing for 10: a true batch-submission API for search queries, webhook/callback support for async job completion, and documented guidance/pricing for large-scale batch workloads.

   • [claimed-docs] “Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detail…”
   • [claimed-docs] “Retrieve the status and results of a research task using its request ID.”
   • [claimed-docs] “When using the Tavily Research API, you can stream responses in real-time by setting `stream: true` in your request.”
   • [claimed-docs] “Extract content from one URL or batch up to 20 URLs in one request. Successful URLs appear in results; per-URL failures appear in failed_res…”
   • [claimed-docs] “the API returns a `429 Too Many Requests` response with a `retry-after` header indicating the number of seconds to wait before making anothe…”
   • [claimed-docs] “Environment Requests per minute (RPM) Development 100 Production 1,000”

   proof ↗⚑ flag

1. ai-native userChoose where my data is stored (region/residency)

   weight 2 · round drawn
   Exa—none0/10

   No evidence anywhere in the pack mentions data residency, region selection, or storage location options for Exa's API, MCP server, or Websets; this is an applicable privacy axis for an AI-native data/search platform but no documentation addresses it.

   ⚑ flag

   Tavily—none0/10

   No evidence in the pack mentions data residency, regional storage options, or geographic control over where data is processed/stored; only security/privacy filtering (PII blocking) is mentioned, not data residency choice. Missing for 10: any mention of region selection, data residency guarantees, or geo-specific deployment options.

   ⚑ flag

2. ai-native userPrevent my data from being used to train AI models

   weight 3 · round drawn
   Exa—none0/10

   The evidence pack contains no documentation, policy statement, or setting related to opting out of data being used for AI model training; the only tangential item concerns Exa scraping third-party websites (ignoring robots.txt), not user-submitted data protection. No opt-out mechanism, privacy policy language, or enterprise data-handling guarantee is present.

   ⚑ flag

   Tavily—none0/10

   No evidence pack item addresses training-data opt-out or a policy preventing customer data from being used to train AI models; only PII/security filtering is mentioned, which is unrelated to model-training data usage.

   ⚑ flag

3. ai-native userControl data retention and deletion

   weight 2 · round drawn
   Exa—none0/10

   No evidence pack items address data retention policies, deletion controls, or privacy/data-handling settings for user data or query history; one community post alleges Exa ignores robots.txt when indexing web content, which is unrelated to user data retention controls. Missing for 10: any documentation of data retention periods, user-initiated deletion/export mechanisms, or privacy controls over stored queries/results.

   ⚑ flag

   Tavily—none0/10

   No evidence in the pack addresses data retention policies, deletion controls, or user ability to manage stored data/logs beyond usage logging mentions; nothing documents a deletion/export mechanism.

   ⚑ flag

4. ai-native userOpt out of telemetry and usage tracking

   weight 2 · round drawn
   Exa—none0/10

   No evidence in the pack addresses telemetry, usage tracking, or opt-out settings for Exa; there's even a community report of Exa itself indexing data disregarding robots.txt, but nothing about the product providing an opt-out for tracking of its own users.

   ⚑ flag

   Tavily—none0/10

   No evidence Tavily documents any telemetry/usage-tracking opt-out mechanism; the only related docs (tavily-docs-33, tavily-docs-34) describe optional session/user/project identifiers for analytics attribution, not a way to disable tracking. Missing for 10: any documented telemetry opt-out setting, privacy policy language on data collection, or configuration flag to disable analytics.

   • [claimed-docs] “You can attach a Project ID to your client to organize and track API usage by project. This is useful when a single API key is used across m…”
   • [claimed-docs] “You can attach session and user identifiers to your client for attribution and analytics across multi-step interactions and agent workflows.”

   proof ↗⚑ flag

Ai consumption

1. ai-native userPlug the API into an agent as a ready-made tool, with function-calling schemas and agent-framework integrations designed for direct tool-call consumption

   weight 3 · round drawn
   Exa✓fullprobed8/10

   Exa ships an official JS/TS SDK with `webSearch()` designed explicitly for use as a tool in OpenAI/Anthropic agent loops, plus an open-source MCP server connecting Claude Desktop, Cursor, VS Code, ChatGPT, Codex, Grok and other MCP clients, and structured output_schema/JSON extraction for direct tool-call consumption. missing for 10: no official OpenAI/Anthropic function-calling JSON schema published directly (only inferred via SDK helper), no independent hands-on validation of agent-framework integration reliability, and no public OpenAPI spec (probe returned 404s).

   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop. Call `webSearch()` with no arguments to get Exa's recommended settings for ag…”
   • [github] “Use Exa as a `web_search` tool in an OpenAI or Anthropic loop.”
   • [claimed-docs] “Complete setup guide for Exa MCP Server. Connect Claude Desktop, Cursor, VS Code, and 10+ AI assistants to Exa's web search, fetching, Exa A…”
   • [claimed-docs] “Exa MCP connects AI assistants to Exa’s search capabilities, including web search, code search, Exa Agent, and Exa Connect.”
   • [claimed-docs] “Connect ChatGPT, Codex, Claude, Grok, Cursor, and any other MCP client to Exa's web search, page fetching, Exa Agent, and Exa Connect tools.”
   • [claimed-docs] “No API key is required to get started. Exa MCP is open source and available on GitHub.”
   • [claimed-docs] “Use `output_schema` with any search type to extract structured JSON from search results”
   • [github] “For type: "object", search currently enforces: max nesting depth: 2, max total properties: 10”
   • [probe] “official MCP server documented at https://exa.ai/docs/reference/exa-mcp”
   • [claimed-docs] “Exa skills teach coding agents how to search, retrieve content, and build with Exa's APIs.”

   proof ↗⚑ flag

   Tavily✓fullprobed8/10

   Tavily ships SDKs (Python/JS) designed explicitly for LLM/agent consumption (qna_search 'perfect for usage by LLMs', get_search_context for direct RAG feeding), plus an official MCP server compatible with any MCP client for tool-call integration, and clear API reference/OpenAPI-style endpoint docs for search/extract/crawl/map/research. missing for 10: no discoverable OpenAPI/JSON schema spec (probe found 404s) and no explicit named framework integrations (e.g., LangChain/LlamaIndex tool classes) cited in evidence.

   • [github] “This is how you get accurate and concise answers to questions, in one line of code. Perfect for usage by LLMs!”
   • [github] “you now have a context string that you can feed directly into your RAG Application”
   • [claimed-docs] “The Tavily MCP server provides: * Seamless interaction with the tavily-search and tavily-extract tools”
   • [claimed-docs] “Compatible with both Cursor and Claude Desktop! Tavily MCP is also compatible with any MCP client.”
   • [claimed-docs] “The Python SDK allows for easy interaction with the Tavily API, offering the full range of our search functionality directly from your Pytho…”
   • [claimed-docs] “Tavily's JavaScript SDK allows for easy interaction with the Tavily API, offering the full range of our search and extract functionalities d…”
   • [probe] “official MCP server documented at https://docs.tavily.com/documentation/mcp”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.tavily.com/openapi.json, https://docs.tavily.com/swagger.json, https://docs.tavily.com/…”

   proof ↗⚑ flag

Llm ready results

1. developerGet a sourced, LLM-generated answer to a query in a single API call

   weight 2 · round drawn
   Exa✓fullclaimed8/10

   Exa's `/answer` endpoint (exposed via exa-js `exa.answer()`) returns an LLM-generated answer with citations in a single call, as shown in exa-gh-3 and exa-gh-8 ("Get answers with citations"). This is backed by first-party SDK code, though there's no independent hands-on validation of answer quality specifically for this endpoint. Missing for 10: independent/community verification of the answer API's accuracy and citation quality, and deeper docs beyond SDK snippets.

   • [github] “const { answer } = await exa.answer("What is the capital of France?");”
   • [github] “Get answers with citations const { answer } = await exa.answer("What is the capital of France?");”
   • [claimed-docs] “Enrich entities, build large lists, and run deep research from a single API with built in structured results and web citation”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily's qna_search (and its underlying Q&A search_depth) returns a single, sourced, LLM-generated direct answer in one API call, explicitly described as 'accurate and concise answers to questions, in one line of code. Perfect for usage by LLMs.' The search endpoint also supports an include_answer option returning a synthesized answer alongside sources. Missing for 10: no independent third-party benchmark of answer quality/accuracy, and no explicit example showing the 'answer' field returned alongside source URLs in the same response payload.

   • [github] “answer = tavily_client.qna_search(query="Who is Leo Messi?")”
   • [github] “This is how you get accurate and concise answers to questions, in one line of code. Perfect for usage by LLMs!”
   • [claimed-docs] “tavily_client.search("Who is Leo Messi?")”
   • [claimed-docs] “Execute a search query using Tavily Search.”

   proof ↗⚑ flag

2. developerGet results sized and ranked for direct use as LLM context in a RAG pipeline

   weight 3 · round drawn
   Exa✓fullcommunity8/10

   Exa's docs explicitly target RAG-context sizing: '10x token efficient extracts' with a recommended 4000-character cap, `output_schema` for structured JSON, clean page content extraction handling JS/PDFs, and citation-backed `answer()` calls — all clearly aimed at feeding LLM pipelines directly. Community feedback (exa-comm-1, exa-comm-8) corroborates the RAG-focused design ('SERP is not the right way to let an LLM interact with the web'), though one user requested a 'cruft cleaner' suggesting content isn't always perfectly trimmed. Missing for 10: independent benchmarks quantifying token-efficiency/relevance gains, and no evidence addressing edge cases like very large result sets consistently staying within context budgets.

   • [claimed-docs] “10x token efficient extracts of only the relevant tokens from a webpage. 4000 characters recommended.”
   • [claimed-docs] “Use `output_schema` with any search type to extract structured JSON from search results”
   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [claimed-docs] “Return query-relevant excerpts from Exa Search results while controlling context size and latency.”
   • [claimed-docs] “Exa Contents returns clean page content from URLs, handling JavaScript-rendered pages, PDFs, and complex layouts automatically.”
   • [github] “const { answer } = await exa.answer("What is the capital of France?");”
   • [community] “Signed up for a trial account and I'm pretty impressed with the search API (haven't used Websets yet but looks cool). Suggested adding a 'cr…”
   • [community] “I am a BIG Exa fan. NewsCatcher and Exa aren't direct competitors but share the same insight — SERP is not the right way to let an LLM inter…”

   proof ↗⚑ flag

   Tavily✓fullclaimed8/10

   Tavily's get_search_context explicitly produces a context string sized for direct RAG use ('you now have a context string that you can feed directly into your RAG Application'), and qna_search is described as tuned for concise LLM-consumable answers; docs also describe content validation layers improving result quality. missing for 10: no independent/third-party benchmark of ranking quality or context sizing tradeoffs beyond first-party claims.

   • [github] “you now have a context string that you can feed directly into your RAG Application”
   • [github] “context = tavily_client.get_search_context(query="What happened during the Burning Man floods?")”
   • [github] “This is how you get accurate and concise answers to questions, in one line of code. Perfect for usage by LLMs!”
   • [claimed-docs] “Requests pass through security, privacy, and content validation layers that block PII leakage, prompt injection, and malicious sources.”

   proof ↗⚑ flag

Retrieval modes

1. developerFind pages similar to a URL I already have

   weight 1 · round drawn
   Exa—none0/10

   The evidence pack covers Exa's keyword/neural search, contents extraction, Websets, MCP integration, and agent tooling, but nowhere mentions a 'find similar' or URL-based similarity search endpoint. No documentation, GitHub snippet, or community report in this pack references finding pages similar to a given URL.

   ⚑ flag

   Tavily—none0/10

   Tavily's evidence shows search (query-based), extract, crawl, and map (both traverse a site's own link structure from a base URL) but no capability to find other pages on the web that are semantically/topically similar to a given URL, which is the actual ask here. Crawl/Map only discover pages within the same site's domain, not similar external content.

   • [github] “Crawl lets you traverse a website's content starting from a base URL.”
   • [claimed-docs] “Tavily Crawl is a graph-based website traversal tool that can explore hundreds of paths in parallel with built-in extraction and intelligent…”
   • [github] “Map lets you discover and visualize the structure of a website starting from a base URL.”
   • [claimed-docs] “Tavily Map traverses websites like a graph and can explore hundreds of paths in parallel with intelligent discovery to generate comprehensiv…”

   proof ↗⚑ flag

2. developerChoose between keyword, neural, or hybrid retrieval modes per query

   weight 2 · round to Exa
   Exa~partialclaimed4/10

   Docs mention 'custom search types with appropriate latency-quality profiles' (instant vs deep-reasoning search), implying multiple selectable search modes, but the evidence never explicitly names 'keyword', 'neural', or 'hybrid' as distinct per-query parameters as described in the story. missing for 10: explicit documentation of keyword/neural/hybrid mode names, a code example showing a 'type' parameter switch, and independent confirmation that all three modes are selectable per query.

   • [claimed-docs] “from ~250 ms instant search to 12-40 second deep-reasoning search”
   • [claimed-docs] “custom search types with appropriate latency-quality profiles, from ~250 ms instant search to 12-40 second deep-reasoning search”
   • [claimed-docs] “Exa has custom search types with appropriate latency-quality profiles, from ~250 ms instant search to 12-40 second deep-reasoning search.”

   proof ↗⚑ flag

   Tavily—none0/10

   The evidence pack shows Tavily's search API supports exact-phrase matching (exact_match=True) but there is no documentation of selectable retrieval modes such as keyword vs neural vs hybrid search per query.

   • [github] “Use exact_match=True to only return results containing the exact phrase(s) inside quotes”

   proof ↗⚑ flag

3. developerRun meaning-based semantic search that finds results keyword engines miss

   weight 3 · round to Tavily
   Exa!disputedcontradicted5/10

   Exa's docs describe a purpose-built neural/semantic search model that returns 'clean, relevant page content' in natural language rather than keyword matching, and one HN commenter explicitly notes Exa's insight that 'SERP is not the right way to let an LLM interact with the web.' However, hands-on community reports contradict the quality claim: one user got '1 match, 14 misses' with results stuck 'Verifying...', and another found GitHub repo search only '30/70' accurate at finding requested items. Missing for 10: independent benchmark comparisons against keyword search, resolution of the reported miss-rate issues, and broader corroboration of search accuracy beyond a single anecdotal thread.

   • [claimed-docs] “Search the web in natural language and get clean, relevant page content in one request.”
   • [claimed-docs] “Specialized model trained to extract the most relevant excerpts from the web”
   • [claimed-docs] “Exa has custom search types with appropriate latency-quality profiles, from ~250 ms instant search to 12-40 second deep-reasoning search.”
   • [community] “I am a BIG Exa fan. NewsCatcher and Exa aren't direct competitors but share the same insight — SERP is not the right way to let an LLM inter…”
   • [community] “I gave it a try and my first search got one match, 14 misses, and all other results are 'Verifying...' but it seems stuck (been minutes)... …”
   • [community] “Tried the GitHub repo search feature: it was about 30/70 on finding the things I asked for. The cURL example in 'Get Code' was demonstrably …”

   proof ↗⚑ flag

   Tavily~partialclaimed5/10

   Tavily's search API accepts natural-language queries (e.g., 'Who is Leo Messi?') and the existence of an `exact_match=True` flag implies the default behavior is not strict keyword matching, suggesting some meaning-based ranking. However, no evidence explicitly documents embeddings, semantic ranking, or a comparison showing it surfaces results keyword search misses. Missing for 10: explicit documentation of semantic/embedding-based ranking methodology, benchmark or example showing recall beyond keyword search, and independent corroboration of search-quality claims.

   • [claimed-docs] “tavily_client.search("Who is Leo Messi?")”
   • [github] “Use exact_match=True to only return results containing the exact phrase(s) inside quotes”
   • [github] “answer = tavily_client.qna_search(query="Who is Leo Messi?")”
   • [github] “you now have a context string that you can feed directly into your RAG Application”

   proof ↗⚑ flag