How we evaluate web data tools
serp.fast catalogues web data infrastructure for AI product builders. The directory currently spans eight categories – AI search APIs, SERP data APIs, web scraping APIs, browser infrastructure, web indexes, agentic extraction, open-source frameworks, and benchmarks. A tool qualifies for inclusion when it is operational, has a working website with public pricing or pricing-on-request, and serves a use case that is recognisable to an AI builder shipping a product. Vaporware, abandoned projects, and tools without active development are excluded. We also do not list traditional proxy-first vendors in the main directory; that audience is served well by other publications.
What we evaluate
Every tool review combines four inputs. The first is vendor documentation: feature lists, supported endpoints, rate-limit language, and the public pricing page. The second is hands-on testing where the vendor offers a free tier or trial credits, with a focus on whether the documented behaviour matches the product. The third is third-party benchmark data when it is published independently and recently – ClawBench is one current example. The fourth is qualitative signals from the vendor's public presence: changelog cadence, support quality, and GitHub activity for open-source projects. We weight criteria toward what AI builders actually need in production: reliability, clarity of pricing at scale, integration ergonomics with model providers and agent frameworks, and whether the vendor's positioning matches the actual product. We weight less heavily things that matter more to other audiences, such as multi-language SDK breadth or on-premise deployment options.
Where we have not tested a tool ourselves, we say so directly in the review. Where an assessment depends on vendor claims rather than first-party observation, the page makes that clear. We do not currently publish numerical scores. Many comparable sites do; the tradeoff is that scores compress nuance into a single digit, and we would rather a reader take 90 seconds to read a verdict than five seconds to glance at a rating. If we add a scoring rubric later, we will publish the full rubric and recompute every tool on the same basis, with the date the rubric was applied.
Data sources and testing
For tools we have tested directly – Exa, Tavily, Firecrawl, Browserbase, and others where the vendor offered a usable trial – the review draws on real request logs, not marketing copy. The how-we-test page describes the specific metrics we record (latency P50/P95, success rate against a fixed set of real-world targets, normalised cost per thousand requests, and JS rendering coverage on JavaScript-heavy pages). For tools we have not yet tested, the review is explicit about that, and the assessment is restricted to what we can verify from the documentation, public benchmarks, and the vendor's observable trajectory. We do not extrapolate performance numbers we have not measured.
The directory is editorially curated and is not a syndication of vendor press releases. Tools cannot pay to be added. We do read press releases and funding announcements, but a launch without a working product does not qualify for a directory listing.
Update cadence and corrections
Every tool page is reviewed at least once per quarter. In practice we refresh more often than that when the underlying landscape moves – pricing changes, major releases, acquisitions, and significant outages all trigger an unscheduled review. Each tool page records when it was last reviewed. If a vendor or reader spots an inaccuracy, we correct it. Email with the page, the specific claim, and a source we can verify; we respond within five business days and update the page with a visible timestamp. Material corrections are noted on the page so a reader can see the history rather than just the final state.
Conflicts of interest
Some outbound links in the directory are affiliate links. They are marked as such on the page they appear on, and they carry the relationship attribute Google asks for on commercial referrals. Affiliate revenue is not material to whether a tool is listed, where it sits in a category, or what the editorial verdict says. We have walked away from affiliate programmes from vendors we could not recommend, and we have written critical assessments of tools we earn referrals on. Newsletter and news-feed sponsorships are accepted but always labelled, and a sponsor cannot edit any other surface of the site. The full revenue position is on the about page; the editorial principles that underpin all of this live on the editorial standards page.