Methodology
183 companies were assessed; 88 are published and 95 are listed as not published with a reason code. Each published entry was collected from the vendor's own pages and re-checked by a separate verifier.
Who is included
A company is listed when it tests AI products for clients and has its own public page describing that service. The page must name at least one AI product type, such as LLM applications, AI agents or computer vision.
Tool vendors (test automation platforms, evaluation SDKs, red-teaming software, observability products) are outside the scope of this directory unless the company also runs a testing services practice with its own service page.
Research institutes and nonprofit evaluators without commercial testing services are not listed.
How an entry is built
- 01
Collect
A collector opens the vendor's website, finds the AI testing page and records the product types, services and engagement models the page names. Quotes are capped at 15 words.
- 02
Second source
HQ country, founding year, size band, delivery locations and certifications need an independent source that agrees: a registry, a review platform profile, a press release or a marketplace listing.
- 03
Verify
A separate verifier re-opens every source URL, checks each value and quote, and sets the verification date. Only records with that date are published.
Statements and figures
Product types, services, engagement models and the stated scope are vendor statements. They are labelled "as stated by the vendor" with the date.
Figures need two independent sources. When only one source is found, the field stays blank. When two sources disagree, the field stays blank and both values are logged.
Headcount is printed as a band, never as an exact number. Prices are not printed. The only pricing field records whether the vendor publishes a price on its own site.
Services in the taxonomy that a vendor's AI page does not mention are shown as "Not stated on the vendor's AI page". This is not a statement that the vendor does not offer them.
Allowed sources
Vendor site, company registry, review platform company profile (never ratings or review counts), press release, marketplace listing, regulator, conference or standards body. Aggregator top-10 lists, social posts and competing vendors' blogs are not used as sources.
Publication thresholds
- Entry: website resolves, AI testing page found and names an AI product type, HQ country confirmed by two sources, at least 3 facet fields filled, not a tool-only vendor.
- Product type, service, company type and country pages: at least 3 published entries.
- City pages: at least 3 entries headquartered in the city.
- Product type in country pages: at least 3 entries.
- Comparison pages: both entries published, same company type, and at least 5 of 9 attributes filled for both.
Everything below a threshold is listed on the not-published page with a reason code.
How comparison pairs are chosen
Each comparison shows two published entries of the same company type. Pairs come from a fixed list of planned pairs and, for every entry, its three most similar entries of the same type, measured by the overlap of the AI product types and services each states. Equal similarity is broken alphabetically.
A pair is published only when at least 5 of 9 attributes are filled for both entries. Planned pairs that miss the rules are listed on the not-published page.
How the ranking works
Lists are ordered by a coverage score. The score is the sum of three counts, taken from the verified record of each entry:
- AI product types stated on the vendor's AI page: 0 to 9
- Testing services stated on the vendor's AI page: 0 to 12
- Figures confirmed by two sources (founded, size band, delivery locations, certifications): 0 to 4
The maximum is 25. Entries with equal scores share a rank and are listed in alphabetical order.
The score measures how much AI testing a vendor states on its own AI page and how many of its figures two independent sources confirm. It does not measure quality, results or client satisfaction, and a vendor that lists more services gets a higher score.
The formula is the same for every entry. No position can be bought. Readers can re-sort any list by name, HQ country, size band or founding year.
Terms used
Service pages use standard terms such as red teaming, defined on first use.
red teaming: Deliberate attempts to make a system produce harmful or unintended output.Refresh cadence
Figures older than 12 months and service statements older than 6 months are re-checked before they are shown again. Changes are dated in the changelog.
Independence
This directory accepts no payment for listing, placement, badges or removal. Vendors cannot buy a position: every list is ordered by one published coverage score, computed the same way for every entry from verified data. Entries are built from each vendor's own published pages and from independent sources named next to every figure.
The site is funded by its publisher. It carries no advertising and no affiliate links. If this changes, the change is dated in the changelog and this statement is updated.
No star ratings, review counts or client testimonials appear on this site. Review platforms are used only as a second source for headquarters, founding year and size band.
Corrections
Anyone can request a correction by email with the URL of the entry, the line in question and a public source. Corrections are checked against the source, applied within the next build, and dated in the changelog. Requests to remove a verified fact are declined; requests to remove an entry are honoured only when the company no longer offers the service or no longer exists.
Questions
What is the difference between an AI agent testing service and an agent testing tool?
A testing service is work done by the vendor's people for a client: scoping, test design, runs and a report. A tool is software the client runs itself. This directory lists services only; software-only vendors are listed on the not-published page with the code tool-only.
How is testing an AI product different from traditional software testing?
The same input can give different outputs, so tests assert on properties and score ranges instead of exact values. Quality is measured on evaluation sets, and new risks such as hallucination, prompt injection and bias need their own tests. Testing continues after release because model behaviour can drift.
What is the difference between testing AI products and using AI tools to test software?
Testing AI products means checking a model-based product itself: its answers, safety and bias. Using AI tools to test software means a QA team uses AI to write or run tests for ordinary applications. This directory lists only the first; vendors whose AI page covers only the second are listed as not published with the code no-ai-page.
What is the difference between automated red teaming tools and human-led red teaming services?
Tools run large sets of known attack prompts quickly and repeatably. Human-led services design new attacks for the specific product and judge subtle failures. This directory lists services; tool-only vendors are on the not-published page.
What is the difference between a voice AI testing service and a voice testing platform?
A service is testing done by the vendor's people for a client. A platform is software the client operates. This directory lists services; voice testing platforms without a services practice are listed as tool-only on the not-published page.