Keenable is a web-search service built for AI agents rather than for people scanning a results page. Its Search API returns ranked web pages with extracted text, a fetch operation returns clean markdown, and two newer surfaces, SELECT and Time Machine, extend search into structured extraction and historical lookups. The company’s headline numbers are an index of more than 100 billion documents and a p95 latency under 250 ms from US East. Both are Keenable’s own published claims. The more useful question for a developer is what a web-scale index costs to serve, and how Keenable says it keeps query time and cost under control.
What Keenable offers
Keenable describes itself as independent web-search infrastructure for AI labs and agents. It is software delivered through an API, SDKs and integrations, with cloud and, on the higher tier, on-premises deployment options. The company’s four product surfaces are summarized below.
As an Amazon Associate I earn from qualifying purchases.
| Surface | What it does | Status as published |
|---|---|---|
| Search API | Returns ranked web results with extracted page text, accessed through an API and SDKs | Developer access described on Keenable’s developer materials; no separate status qualifier given |
| Fetch | Returns the content of a page as clean markdown | Described in Keenable’s developer materials |
| SELECT | A SQL-like way to search web results, extract named fields, then filter, group and aggregate them into a table or report | Presented in Keenable’s SELECT essay as a product design direction |
| Time Machine | Point-in-time search over prior versions of pages, where the query sets the historical corpus and ranking | Labelled early access on Keenable’s official site |
The practical distinction is between retrieval and analysis. Search and fetch give an agent material to read. SELECT is aimed at questions whose answer is a property of a whole set of pages, and Time Machine is aimed at questions about what a page said at a given moment.
Pages or documents: what the index count measures
The title uses the phrase “100B-page index,” but Keenable’s own materials generally count “documents.” The two words are not interchangeable without a definition. A single URL can change over time, and a count of documents depends on how the company treats duplicates, versions and non-HTML content. Keenable’s published material does not explain its counting unit or method, so the figure should be read as a company-defined scale marker rather than a verified count of distinct web pages.
#1 Best Overall
The cost trade-off behind a large index
Keenable’s central architectural argument is that serving and scanning the whole web is expensive, so the system has to narrow the candidate set according to each query. CEO Andrey Styskin, co-founder and CEO of Keenable, put it this way in a TechCrunch report published August 25, 2026:
“If you do not fine-tune your index structures for a specific task, the cost of serving and scanning the whole internet is enormous because of the volume. That’s why you need to innovate on how you can narrow the search space based on your query very fast. This is what we are bringing to the table.”
This is a sound way to frame the problem, and it shifts the evaluation away from raw index size. Coverage tells you what could be found; the narrowing step determines what is found, how quickly, and at what cost. For a buyer, the metrics that follow from this trade-off are:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →- Corpus coverage and freshness for the specific domains your agent needs, not the global total.
- Candidate generation and ranking, meaning whether the top results are the ones your agent actually needs.
- Extracted-text quality, since an agent reasons over whatever text the API returns.
- Latency distribution, including p50 and p95 from your own region.
- Price per useful answer, which combines request price with the number of retries and follow-up fetches an agent needs.
These are analytical criteria drawn from the company’s stated trade-off. Keenable has not published an independent comparison showing how its narrowing method performs against other search APIs.
Published figures and what they establish
| Claim | Who published it | What it does and does not establish |
|---|---|---|
| More than 100 billion documents indexed | Keenable’s homepage; reported by TechCrunch on August 25, 2026 as Keenable’s claim | A company-stated scale. The counting unit and method are not published in the material reviewed. |
| p95 latency under 250 ms | Keenable’s homepage | States the percentile and the region (US East). The test setup, query mix, concurrency and client location are not given. |
| Quality comparison chart against a NEEDLE benchmark | Keenable’s homepage | A vendor chart. Quality is defined by Keenable as a seven-day mean fraction of pooled “ultimate” performance. The benchmark protocol and data have not been independently checked. |
| 46 researcher moves across 11 frontier foundation-model labs, January 2025 to August 2026 | Keenable’s SELECT essay | An example of SELECT output showing a structured workflow. It is not a general measure of search quality. |
No independent benchmark methodology for these figures was found in the published material. Treat the index scale and latency as claims to test, not as measured properties of the service in your deployment.
Latency: how to test the 250 ms claim
A p95 figure answers only one question: for 95 percent of measured requests, the response arrived within the stated time under the test conditions. Because those conditions are not published, a reader cannot tell whether 250 ms reflects a warm cache, a simple query or a production mix. A workable check looks like this:
- Run your own query set from the region where your agent is deployed, not from a US East test machine unless that is where you run.
- Record p50, p95 and p99 separately for search calls, and time fetch calls on the same URLs.
- Repeat at your expected concurrency, because tail latency tends to rise under load.
- Count failed or empty responses as well as slow ones, since a fast empty result does not help an agent.
This article did not benchmark Keenable. The steps above describe how a team would verify the claim.
Pricing as published
Keenable’s pricing page, as captured in sources available for this article (last crawled several weeks before early October 2026), lists the following. Pricing and eligibility terms can change, so confirm them on the live page before budgeting.
| Tier or offer | Published price | Access and conditions as listed |
|---|---|---|
| Free offer | 100,000 requests a month | Eligibility and terms are set by Keenable and should be checked on the pricing page |
| Agent Builder | $4 per 1,000 requests | Cloud-only, pay as you go |
| Frontier | $1 per 1,000 requests at 100 requests per second or more | Dedicated capacity for AI labs and inference platforms, with cloud and on-premises access; the threshold, eligibility and terms should be confirmed |
At the listed Agent Builder rate, one million requests a month would cost $4,000 before any tier-specific terms apply. Frontier pricing only makes sense at sustained high throughput, which is why the request rate is part of the price, not a side detail.
SELECT: from ranked links to structured answers
Keenable’s SELECT essay argues that ranked links and short snippets suit a person who opens one result. An agent answering a question about a set of pages may need the distribution across many results and specific fields pulled from each. SELECT makes that explicit by combining search, extraction, grouping and aggregation in one query.
The essay’s example, 46 researcher moves across 11 frontier foundation-model labs between January 2025 and August 2026, illustrates the type of answer. No single page holds that count, so it has to be assembled from many sources. The aggregate is only as complete as the pages the search found and the fields the extraction captured. Keenable’s essay is a product rationale, not independent evaluation, so completeness should be tested on your own question before relying on it.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsTime Machine: searching what a page said
Time Machine performs point-in-time search over prior versions of pages, and the query determines both the historical corpus and the ranking. That makes it useful for questions such as what a vendor’s documentation said in a given year. Keenable’s official site labels Time Machine as early access, so check whether your account can use it before designing a workflow around it. Snapshot coverage, meaning which dates and sites are represented, is not described in the published material.
Best Value
Integration paths
- Python and TypeScript SDKs: the SDK documentation describes keyless defaults, and an optional API key affects rate limits.
- LangChain integration: a published integration path for agents built on LangChain.
- MCP server: the repository documents hosted search and fetch tools and a keyless request cap. The numeric cap is not stated in the material reviewed.
These confirm the interfaces exist and how they are described. They do not establish reliability or how the developer experience compares with other search APIs.
Adoption and partnerships
In the TechCrunch report of August 25, 2026, Keenable said its API was in production at several AI labs and inference providers, for both training and runtime use. Those customers were not named. The report also names a partnership with the voice AI company Gradium for live information retrieval. These are Keenable’s statements as reported. They show that the service is being used, but they do not show that it outperforms competing APIs on any workload.
How to decide whether Keenable fits your agent
- Define the workload: live retrieval during an agent run, bulk collection for training data, or set-level analysis that SELECT targets.
- Build a query set from real tasks and compare relevance against the search API you use now.
- Measure latency from your deployment region, as described above.
- Calculate cost at your expected monthly volume, first against the free offer and then against the Agent Builder or Frontier rate that applies.
- Check extraction completeness on the specific fields you need, not on a sample of easy pages.
- If you need history, confirm Time Machine access on your account and test coverage for the dates and sites you care about.
- Confirm the deployment and data-control model. Cloud-only and on-premises options carry different operational obligations.
Keenable’s stated design premise is that an agent-first search system should narrow the search space quickly for each query. Whether it does that better than alternatives is an empirical question that the published material does not settle.
Free tools Windows power users keep installed
One-click scans. No signup required.
Keenable offers an agent-oriented search, fetch, extraction and historical-search stack with clearly tiered pricing. Its headline scale and latency figures are company claims, and the central question for buyers is how the query-time narrowing performs on their own workloads.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
