Research and grading methodology

Where the data comes from

Vendor records are built from primary sources, in this order: the vendor's own site, documentation, pricing page, status page, security or trust page, changelog and public API reference; then press releases and funding announcements; then company registries; then reputable industry press; then a client firm's own site confirming the relationship.

Three things are never a source: a competitor's comparison table, an anonymous claim in a forum or Discord, and another model's output. Marketing copy is not a source either — "bank-grade security" is not a certification, "lightning fast" is not a latency figure, and "trusted by hundreds of firms" is not a client count. A field with no specific, checkable statement behind it stays blank.

The highest-risk fields — security certifications, uptime guarantees, named clients, every pricing number, company registration and funding — are only filled from an explicit statement we can point at. A blank field is correct. A guessed field is damage.

How prices are made comparable

Vendors bill monthly, quarterly, annually, per active account, as a share of revenue, or as a one-off licence, and most sell several of those at once. Put side by side, none of those numbers mean the same thing.

So every recurring tier is normalised to a monthly figure — quarterly divided by three, annual by twelve — and the cheapest such tier becomes the comparable number. Estimated first-year cost is that figure times twelve, plus any published one-off setup fee.

Per-active-account fees and revenue share are deliberately excluded from it and reported on their own rows. Their cost is a function of your funded book, not the vendor's price list, and at any real volume they routinely exceed the flat fee. Rolling them into one number would produce a figure that describes nothing, which is the exact failure this site exists to avoid.

How the disclosure grade is computed

Each vendor is scored out of 100 across eleven published-or-not checks. The weights are a judgement, and they are here rather than buried in the code: price weighs most because an unpriced vendor cannot be shortlisted at all, and contract terms and security follow because those are where finding out late is most expensive.

• A public price — 18 points
• What the price covers — 7 points
• Contract length and notice period — 12 points
• Data ownership and export on exit — 8 points
• Named security certifications — 12 points
• An uptime figure or status page — 10 points
• Support hours and response times — 6 points
• A time to go live — 8 points
• How and where it is hosted — 7 points
• A public API surface — 5 points
• Firms powered or named clients — 7 points

What the grade is not. It is not a quality score, a recommendation, or a safety rating. A vendor that publishes a weak SLA scores better than one that publishes none — deliberately, because a buyer can act on a weak number and cannot act on a blank. Across the vendors we track the median grade is 6/100, which tells you more about the market than about any vendor in it.

What the comparisons will and will not decide

A head-to-head page tints a cell only where the measure has a defensible direction from the buying firm's side: a lower cost, a shorter lock-in, a shorter notice period, a published figure against an unpublished one. Deployment model, hosting region, terminal support and asset classes carry no winner, because which is better is a fact about your firm rather than about the vendor.

Rows where neither vendor publishes anything are dropped from the page entirely rather than shown as two empty cells, and the gaps are collected instead into an Ask on the first call list — which is the part of a comparison a vendor's own site can never give you.

Independence

Nobody pays for placement in a comparison, for a position in an alternatives list, or for a disclosure grade. Featured is a paid label on a listing and changes nothing else. Some outbound vendor links are affiliate links; that is disclosed on-site and has no effect on any figure, ranking or grade on this page or any other.

The full normalised dataset, with every caveat attached, is published at /llms-full.txt so anyone — including an AI assistant — can check our arithmetic rather than take it on trust.