Keenable sells two fast primitives on a 100B-document index; Parallel and Exa sell the whole research factory

Asked:
Can you make a document really digging deep what you kind explaining how the product works for Keenable, how it was made, and the diffrent offering for example I understnad how Mcdonalds works I really want to understand every part of how the business. Then can you compare that to its main competitors for now lets focus on Parallel and Exa

Researched as of 5 September 2026 from 240 official product and documentation pages across the three vendors plus 757 independent and company reporting extracts. All three companies sell web intelligence to developers and AI agents; product claims about quality, latency and coverage below are vendor claims unless attributed to independent reporting.

The value chain at a glance

Six stages, three companies. A solid block is a documented product; an outlined amber block is a capability reached only indirectly or still announced; an empty dotted cell is not offered publicly. Hover any block for detail.

documented offeringindirect / announcednot offered
Keenable advertises the lowest headline price in the group — from $1 per 1,000 requests — but the rate is conditioned on dedicated capacity at 100+ requests per second; independent coverage notes a $4 per 1,000 “Agent Builder” tier for smaller users. quasa.io
Parallel’s menu spans a 2,400x internal price range on one axis alone: Task API runs cost $5 to $2,400 per 1,000 depending on the processor tier — evidence that a “request” is not one product but many. docs.parallel.ai
Exa is the ecosystem incumbent of the three: it reports 5,000+ companies and 400,000 developers, with Databricks, AWS, Vercel and Cursor cited as customers, against Keenable’s unnamed AI labs. chierhu.medium.com
Capital tells the strategy: Keenable has raised $26M (seed, Accel), Parallel $230M ($2B valuation), Exa $357M+ ($2.2B). Keenable is betting a 15-engineer team can win on index economics rather than product breadth. techcrunch.com

1 · Executive summary

Keenable is an independent, agent-oriented web search and page-fetch layer built on its own 100B+ document index, sold as two low-latency primitives (search and fetch) with under-250ms p95 claimed in US East. Parallel Web Systems is the broadest explicitly modular “programmatic web” stack — Search, Extract, Task, Responses, Monitor, FindAll, Entity Search and Chat — founded by former Twitter CEO Parag Agrawal. Exa is a semantic/neural search and data platform with Search, Contents, Agent, Websets and Monitors, plus proprietary people, company and publication indexes.

Where each wins: Keenable for high-volume, latency-sensitive grounding at frontier scale and for keyless prototyping; Parallel when the buyer wants one vendor for the whole workflow from cheap retrieval to long-running cited research and monitoring; Exa for entity-heavy work (people, companies, papers), neural retrieval quality, and the most mature customer ecosystem.

2 · Category map

“Web intelligence” is really five distinct jobs, and the lead visual is organised on them: SERP/search (ranked results and excerpts for a query), fetch/extract (turning one URL into clean model-ready text), deep research (multi-hop, cited synthesis), entity/list building (exhaustive enumeration of companies or people matching criteria), and monitoring (standing queries pushed on change). Keenable competes in the first two; Parallel and Exa ship products across all five. Do not flatten the units: a search request, an extracted URL, a Task run, a Responses call, a matched entity and a monitor check are priced and consumed as different products.

3 · Keenable from first principles — the McDonald’s walk-through

Think of Keenable as a kitchen that pre-cooks the entire web. Raw inputs: public web pages, crawled and refreshed continuously. The production line: parse, clean and index those pages into a store of 100B+ documents. The menu: exactly two core items — search (ask a question, get ranked results with snippets and timestamps) and fetch (hand over a URL, get the page back as clean markdown, from the indexed copy or fetched live). Ordering channels: REST API, a remote MCP server for agents (tools search_web_pages and fetch_page_content), a CLI, and roughly 26 framework integrations from LangChain and LlamaIndex to Vercel AI SDK and NVIDIA NeMo. The counter with no cashier: a keyless shared public tier (1,000 requests/hour, 10/second per IP) lets anyone taste the product; an API key removes the hourly cap and switches on the revenue meter — credits per operation under SKUs search.realtime, search.pro, fetch and fetch.live, with 100,000 free requests each month. Distinctive menu options: point-in-time search (query_time re-bases the whole query to a past instant), date/site filters, live fetching, and an LLM extraction instruction on fetch.

The operating pipeline — a public-source reconstruction

Solid = documented by Keenable; dashed amber = inferred from public evidence, standard for any web-scale search system; dotted grey = not publicly disclosed. Keenable has not published its crawler design, ranking models, hardware or margins, and none are invented here.

inferred1 · Crawl & refresh

Discover and re-visit pages at web scale. Required by a 100B+ index and live fetch, but crawler design is undisclosed.

inferred2 · Parse & normalize

Strip pages to clean text/markdown. Implied by markdown fetch output; method undisclosed.

documented3 · Index

100B+ documents with acquisition and publication timestamps (ISO 8601 in API responses).

documented4 · Query interpretation

Natural-language query plus site, date and point-in-time filters; realtime vs pro modes.

not disclosed5 · Retrieval & ranking

Ranking models, hardware and index structures are proprietary; founders cite task-tuned index design as the cost lever.

documented6 · Fetch layer

Return the indexed copy or fetch live (live=true); optional LLM extraction instruction on the page.

documented7 · Package for models

Titles, URLs, snippets, timestamps, markdown with truncation flags — sized for a context window, not a human SERP.

documented8 · Expose

REST, remote MCP, CLI, SDK/framework integrations, agent connectors (Claude, ChatGPT plugins, OpenClaw, ketch…).

documented9 · Meter

Credits per authenticated operation by SKU; keyless tier unbilled; 402 when credits run out.

4 · Keenable as a business

Users vs buyers. The user is an agent or a developer; the economic buyer is an AI lab, inference provider or agent-platform team — Keenable’s own hiring page says its most important customers are “the teams building frontier models and the products on top of them”, ranging from a solo builder doing hundreds of queries a day to a lab doing millions. Acquisition loop. Keyless public tier → framework integrations put the tool inside every popular agent stack → free 100k monthly requests convert tinkerers → usage-based credits and dedicated-capacity contracts monetize scale. Costs (analytical inference, not disclosed figures): crawling and refreshing the web, index storage and serving compute are the dominant variable costs; the CEO calls index building “painfully expensive”. Fixed costs are small — roughly 15 engineers across the US and Europe. Unit-economic drivers (inference): requests served, share of live fetches vs cached copies, index refresh rate, and compute per query. Partnerships. One disclosed: voice-AI company Gradium, for live retrieval within a spoken-conversation latency budget. Defensibility. An owned index at 100B+ documents is the moat claim; the risk section below tests it.

5 · Origin — how it was made

Keenable was founded by Andrey Styskin (CEO; previously led Yandex’s search, AI and cloud division, later a director at Amazon AGI) and Matthias Petri (Chief Scientist; Principal Applied Scientist at Amazon AGI, where he worked on web-grounding infrastructure for Alexa). PitchBook and several reports state a 2025 founding and a $26M seed dated 5 November 2025; the company emerged from stealth on 25 August 2026 — treat the exact founding date as reported, not confirmed by the company. The seed was led by Accel with Conviction Partners, Brightwing Capital, ScOp Venture Capital and angels from Google and Amazon. The thesis: consumer search engines were optimized for human clicks and are being closed to third-party API access, so agents need an independent index tuned for high-frequency, low-latency machine queries. Production use is reported at several unnamed AI labs and inference providers, during both training and runtime.

6 · Full menu — all three companies

Keenable

  • Search API — realtime and pro SKUs; site/date/point-in-time filters. From $1/1k at 100+ RPS (dedicated); $4/1k reported for smaller tiers.
  • Fetch API — indexed copy or live=true; markdown out; optional extraction instruction. SKUs fetch, fetch.live.
  • Remote MCP + CLI — search_web_pages, fetch_page_content; billing metadata via MCP _meta.
  • Keyless public tier — 1,000 req/hr, 10 rps per IP, unbilled; 100k free authenticated requests/month.
  • Integrations — ~26 frameworks and agents (LangChain, LlamaIndex, Haystack, Mastra, Vercel AI SDK, NeMo, n8n, Dify, Convex, Pipecat…). These are channels, not separate products.
  • Web Query Language — announced multi-source query product; no concrete public form yet.

Parallel

  • Search — turbo/fast $1/1k, basic/advanced $5/1k; objective alongside the keyword query.
  • Extract — $1/1k URLs; JS-heavy pages and PDFs.
  • Task — deep research runs, $5–$2,400/1k by processor (lite→ultra8x).
  • Responses — OpenAI-compatible cited answers, $10–$250/1k by reasoning effort.
  • Monitor — $3–$10/1k checks; snapshots and event streams.
  • FindAll — $0.25/run + $0.03/match (base); verified entity lists.
  • Entity Search — synchronous, $5/1k requests.
  • Chat — research models, $5–$25/1k.
  • MCP + agent payments — free keyless Search MCP; MPP/x402 pay-per-request for agents.

Exa

  • Search — instant→deep-reasoning; $7/1k with contents included for 10 results; Deep $12, Deep-Reasoning $15/1k.
  • Contents — $1/1k pages per content type; text, highlights, dynamic highlights, summaries, subpage crawl.
  • Exa Agent — async runs; fixed efforts $0.012–$1.00/request or metered ACUs at $0.10.
  • Websets — entity list building; Free, Core $49/mo, Pro $449/mo, Enterprise.
  • Monitors — scheduled searches, webhook delivery, deduplication.
  • Specialized indexes — 1B+ people profiles, 70M companies, 350M publications (vendor claims).
  • MCP + agent payments — remote MCP, free plan; x402/MPP pay-per-request.

7 · Head-to-head matrix

Prices are current public list prices from official documentation as of 5 September 2026 and can change. Units are not equivalent across rows or vendors.

DimensionKeenableParallelExaSources

Archetypal buying decisions

8 · Competitive strategy and moat

All three own an index, so the moat contest moves to second-order factors. Index scale and freshness: all claims are self-reported and none independently audited. Retrieval quality: each vendor publishes benchmarks in which it wins; treat all as marketing until reproduced. Agent-native output is table stakes — dense excerpts, markdown, structured output. Workflow breadth favors Parallel, entity data favors Exa, price/performance at frontier volume is Keenable’s wedge. Distribution: Keenable has moved fastest per dollar — 26 framework integrations on a $26M seed — while Exa has the deepest installed base and Parallel the strongest enterprise channel (Google Gemini Enterprise grounding). Capital intensity cuts both ways: Keenable’s founders argue task-tuned index design makes web search 10x cheaper; if true it is the moat, if not, the $331M funding gap to its rivals is decisive.

9 · Risks and diligence questions for Keenable

These are open questions raised by the public record, not allegations.

10 · Conclusion

Today Keenable is a fast, cheap, narrow search-and-fetch layer on an owned web-scale index, one year old, with $26M and 15 engineers. To become a durable large business it must convert unnamed production use into named, expanding contracts, prove its cost and latency claims independently, and either stay the best-price primitive as volume explodes or grow a second act (Web Query Language). Choose Keenable for high-volume grounding and fetch where latency and unit cost dominate; Parallel when one vendor must cover search through deep research, entity discovery and monitoring; Exa for neural retrieval quality, people/company/publication data and the most proven ecosystem.

11 · Methodology and source notes

Product details come from each vendor’s official documentation and blog (80 pages per vendor in this research set) and are therefore vendor claims — including all quality, latency, coverage and benchmark statements. Company facts (funding, founders, customers, risks) come from independent reporting: TechCrunch, SiliconANGLE, Pulse2, Quasa, PitchBook and others, drawn from 757 company-reporting extracts. Where sources conflict — for example Keenable’s founding date and seed timing — the conflict is stated rather than resolved. Prices are public list prices as of 5 September 2026 and can change.

Built from 240 official Keenable, Parallel and Exa product/documentation rows (80 each, one source URL per row) and 757 independent/company reporting extracts, researched as of 2026-09-05; the research first scanned 341 Keenable, 631 Parallel, 746 Exa and 757 company search-result extracts before narrowing. Prices are vendor list prices per stated unit (requests, URLs, runs, matches, checks) and are not equivalent across products. The matrix condenses each dimension to one line per vendor; secondary integrations and minor endpoints were cut for space.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable SELECTAsk your own question
Made with Keenable SELECT