Advance Keenable to full diligence: strong founder-market fit and an owned 100B-document index, but the case turns entirely on undisclosed unit costs at a $1–$4 per 1K price point

This memo, dated 6 September 2026, draws on 40 sourced evidence rows on Keenable (company docs, the lead investor, mainstream and industry news, developer communities and directories) and 20 rows covering the competitive search-API landscape, all as of 2026-09-06. Units are USD; "/1K" means per 1,000 requests. Reported facts are marked reported and analyst constructions analyst assumption throughout; Keenable has disclosed no revenue, retention, COGS, cash balance or contract terms.

1 · Recommendation

Advance to full partner diligence. Do not issue an unconditional investment recommendation. Keenable presents unusually strong founder-market fit, a genuinely owned web index built for machine-scale retrieval, reported sub-250 ms p95 latency, integrated search plus clean-content fetch, point-in-time retrieval ("Time Machine"), a planned Web Query Language, and reported commercial contracts with multiple AI labs. But price, revenue, retention, customer concentration and measured cost per 1,000 requests are all undisclosed. The central question is whether an expensive independent index can sustain attractive gross margins while direct API prices compress toward $1–$5/1K and hyperscalers bundle grounding into model calls.

2 · Why now

reported Agent workloads issue search at machine volumes and need model-ready content, not ten blue links. Keenable claims agent search "an order of magnitude more cost-efficient than traditional search infrastructure" (pulse2.com), and Google/Microsoft API access is constrained for AI use (ventureburn.com). analyst assumption The window is real but closing: incumbents are already repricing grounding, and the seed sizes in this category imply the land-grab is on now.

3 · Company and product

reported Founded 2025 by Andrey Styskin (former CEO of Yandex Search, later director at Amazon AGI) and Matthias Petri (former Principal Applied Scientist at Amazon AGI who built web grounding for Alexa and designed a trillion-token-scale index) — pulse2.com, note.com. The product is a Search API (search_web_pages, fetch_page_content) over an owned, continuously crawled index of 100B+ documents, served via REST, CLI and MCP, plus Time Machine point-in-time retrieval. The API is reported in production at several AI labs and inference providers for both training and runtime — siliconangle.com.

4 · Product differentiation — a stack, not a slogan

1. Owned index reported — continuously crawled and ranked rather than SERP resale (keenable.ai). Demonstrated: the index exists and is served publicly.
2. Agent-optimized retrieval structures reported — index structures that narrow the search space fast to avoid scanning the web per query; sub-250 ms p95 claimed (techcrunch.com). Partially demonstrated; latency figures are self-reported.
3. Search plus clean fetch in one platform reported — via REST/CLI/MCP, up to 50 results per request (docs.keenable.ai). Demonstrated in public docs.
4. Point-in-time retrieval (Time Machine) reported — search the web as of a past date (dealroom.co). Shipped per docs changelog; depth of history unverified.
5. Web Query Language reported/roadmap — multi-source reasoning when no single page answers; announced, "to be released soon" (opentools.ai). Roadmap claim, not demonstrated.
6. Learning flywheel analyst assumption — an index that "learns continuously from being used" by agents (keenable.ai). Plausible mechanism, zero disclosed evidence of realized quality gains.

analyst assumption NEEDLE, Keenable's open live benchmark that rebuilds its query set hourly (marktechpost.com), is useful evidence of engineering seriousness but is Keenable's own benchmark — not independent proof of quality leadership.

5 · Market and demand drivers

reported Buyers are AI labs, inference providers and agent developers; demand spans training-time corpus access and runtime grounding. Keenable reports production usage in both — siliconangle.com. analyst assumption Machine query volumes can exceed human search volumes by orders of magnitude per customer, so a small number of labs can carry the model — which is also the concentration risk in §11.

6 · Competition

analyst assumption The chart above normalizes 20 competitor evidence rows into 12 directional $/1K positions. Three clusters: SERP resellers (Serper ~$1/1K, DataForSEO, Bright Data, SearchAPI) that are cheap but depend on Google remaining scrapeable; independent or AI-native indexes (Brave ~$5/1K, Exa ~$7/1K, Tavily ~$0.008/credit, Linkup ~$5.50) that are Keenable's direct comparables; and bundled answer/grounding products (Perplexity Sonar $5/1K plus token charges, Google grounding $14–$35/1K grounded prompts) that compete for the same budget with distribution advantages. Keenable's $1–$4/1K undercuts every independent-index comparable — aggressive if clean fetch and 50 results are truly included, but the $1 dedicated tier creates gross-margin risk unless its own-index cost curve is genuinely superior.

7 · Business model and unit economics

reported Published SKUs: free tier of 100,000 requests/month; $4/1K pay-as-you-go; $1/1K dedicated at 100+ RPS; launch promo free through end of September (docs.keenable.ai, x.com).

analyst assumption — illustrative, COGS undisclosed Revenue = annual requests / 1,000 × realized price per 1K. Gross margin = 1 − variable cost per 1K ÷ realized price per 1K. At the $4 price, variable cost of $0.25 / $0.50 / $1.00 per 1K yields 93.8% / 87.5% / 75.0% gross margin. At the $1 dedicated price, variable cost of $0.10 / $0.25 / $0.50 yields 90% / 75% / 50%. The $1 tier is where margin risk lives: a $0.50 true cost halves the margin.

analyst assumption Contract math: 100 RPS × 86,400 s × 30 days = 259.2M requests/month → $259,200 MRR / $3.11M ARR at $1/1K fully utilized; ~$0.93M ARR at 30% utilization. A 10M req/month pay-as-you-go customer at $4/1K = $40,000 MRR / $480,000 ARR.

8 · Three-scenario financial model

analyst assumption Bottom-up constructions from the pricing evidence — not management guidance, not forecasts. Formulas shown.

Break-even and runway frame analyst assumption

If 30 employees cost $250K fully loaded each ($7.5M) and non-payroll infrastructure/GTM/G&A totals $5.5M, annual operating expense is about $13M. At zero revenue, $26M equals roughly 24 months of funding before financing costs and timing effects — not actual runway, since the financing date, current cash, prior spend and revenue are unknown. Break-even revenue is about $15.3M at 85% GM or $17.3M at 75% GM: roughly five to six fully utilized 100-RPS workloads at $1/1K, or roughly 16–19 such workloads at 30% utilization.

9 · Valuation and return framework

analyst assumption The financing valuation is undisclosed (pitchbook.com lists the round without terms), so this is mechanics, not a fair value. For an illustrative $5M check: at $100M post-money, ownership is 5%; a 10x gross return needs a $1B exit before dilution, or ~$2B if subsequent dilution halves ownership. At $150M post, ownership is 3.33%; 10x needs $1.5B before dilution, or ~$3B after 50% dilution. Entry price is decisive; these are mechanical examples, not knowledge of Keenable's terms.

10 · Risks, ranked

1. Unknown COGS / capital intensity (high severity, high probability) — crawling, storing and re-ranking 100B+ documents is expensive; no cost disclosure exists.
2. Price compression and bundling (high, high) — Google, Microsoft, OpenAI and Perplexity can bundle grounding; direct API prices already compress toward $1–$5/1K.
3. Customer concentration (high, medium) — a few labs likely dominate revenue; no concentration data disclosed.
4. Vertical integration by model providers (high, medium) — the largest customers are the likeliest builders of in-house retrieval.
5. Quality/freshness gap vs larger indexes (medium, medium) — 100B documents trails Google-scale coverage; NEEDLE is self-run evidence.
6. Content rights, robots and publisher litigation (high, medium) — crawling for AI use is contested terrain.
7. Roadmap dependence (medium, medium) — Web Query Language and the learning flywheel are undemonstrated.
8. Free-tier abuse (low, high) — 100K free requests/month plus keyless access invites arbitrage despite per-IP caps.
9. Sales cycle and reliability bar (medium, medium) — labs demand SLA-grade uptime; no SLA history disclosed.
10. Benchmark independence, security and data provenance (medium, medium).
11. Oversized seed (medium, medium) — $26M raises burn tolerance and next-round expectations.

11 · Diligence requests and kill criteria

• Last six months of requests and revenue by customer and cohort; NRR and churn; top-five concentration.
• Contracted minimums and actual utilization on dedicated capacity; gross margin by SKU; variable cost decomposition (serving, crawl, storage, bandwidth) per 1K requests.
• Freshness, coverage and latency measured against independent query sets, not NEEDLE alone; crawl/recrawl cadence and storage capex.
• Content-rights and robots policy, litigation exposure; uptime/SLA history; pipeline; current cash and burn.
• Kill criteria: measured variable cost above ~$0.50/1K at the $1 tier; any single customer above 50% of revenue with no diversification path; independent quality tests losing broadly to Brave/Exa/Perplexity on agent workloads; evidence contracts are trials rather than committed spend.

12 · Milestones and investment gates

1. Production gross margin of at least 75% at the $1 dedicated tier under realistic utilization.
2. At least three unaffiliated production customers with referenceable retention or binding minimums.
3. No customer above 35% of revenue at scale, or a credible diversification plan.
4. Independent quality/freshness win in the target agent workloads.
5. At least 18 months runway post-investment under a downside plan.
6. Entry valuation capable of returning the fund under realistic dilution and a $2B–$3B success case.

13 · Investment conclusion

Keenable is a rare team building the hard, capital-intensive layer of the agent stack, with real public product evidence and reported commercial traction at launch. The bet is not on demand — machine-scale web retrieval demand is visible everywhere in the competitive set — but on whether an owned index can hold 75%+ gross margins at $1/1K while incumbents compress price. That is an empirical question the data room can answer. Advance to full diligence; invest only if the six gates above clear.

Method: venture IC memo built from two supplied result sets as of 2026-09-06 — 40 sourced evidence rows on Keenable (company docs, lead investor, news, developer communities, directories; syndicated repeats treated as corroboration only when independently reported) and 20 rows of competitor/API pricing evidence across multiple independent hosts. All financial scenarios, unit economics, runway and valuation figures are analyst assumptions marked as such; Keenable has disclosed no revenue, retention, COGS, cash or contract terms. $/1K price positions are directional normalizations since billing units differ. Duplicate syndicated URLs and per-row publish dates were cut for space. Units USD.