Research methodology
How information reaches a page, and how you can tell what strength each claim has. The short version: every field carries its own status, "unconfirmed" is a legitimate answer, and nobody gets to call something verified without saying who verified it.
The four information statuses
Most directories treat a claim as true because a vendor said it. We label the strength of the evidence behind each field instead, so you can decide how much weight a claim deserves in your own evaluation.
| Status | What it means |
|---|---|
| Vendor Claimed | Published by the vendor. We have not confirmed it independently. |
| Source Confirmed | We located the claim in a public source. We did not test it ourselves. |
| Independently Verified | We tested or confirmed this ourselves against primary evidence. |
| Not Yet Confirmed | We have not been able to confirm this yet. Absence of evidence, not evidence of absence. |
Evidence strength is not the same as verification
An evidence grade says how strong a claim is. It does not say who checked it. Those are different questions, and conflating them is how a directory ends up displaying a green check that nobody ever earned.
So every record carries a second value: who verified it. The rule is absolute — without a verifier, nothing can be shown as verified. A record whose material has not been independently checked is shown as vendor claimed, however strongly its own source file describes it.
| Verifier | What it means |
|---|---|
| Checked by our research team | We located the source, read it, and confirmed the claim ourselves. |
| Checked by a third party | An independent organisation outside this platform verified this. We did not repeat the check ourselves. |
| No independent verification yet | Nobody has independently checked this yet. Whatever strength a source states, it remains a claim. |
Where a record's own grade is higher than its verification supports, we say so on the record page rather than quietly showing the lower grade. The original value stays on file for internal tracking, and the difference between the two is counted on the data status page.
Record status
Record status describes where a record sits in the research process, as opposed to how strong any one claim on it is. A record can be pending overall while individual fields carry vendor-claimed grades — those are different questions about the same record.
| Status | Definition |
|---|---|
| Verified | Every material field on this record has been checked against a named source by our research team. |
| Partially verified | Most fields are sourced. Some are not confirmed, and the page says which. |
| Pending verification | The record exists, but the claims on it have not been independently checked yet. |
| Needs re-check | This record was checked some time ago and may no longer reflect the current state. Treat it as a starting point, not a current fact. |
| Archived | The product or company is no longer active in the market. Kept for reference, not for selection. |
Records carry a next review date, set by interval: 180 days for verified records, 90 for partial, 60 for pending, 30 for those needing a re-check. A review that has passed its date is flagged as overdue on the record rather than being quietly left as current.
Source types
A vendor's own documentation and a regulator filing are not the same kind of object, and a directory that does not distinguish them gives the reader no way to tell.
| Type | What it means |
|---|---|
| Vendor documentation | Published by the vendor themselves. |
| Third-party publication | Published by an organisation with no commercial interest in the claim. |
| Our own test or evaluation | Produced by this platform's research team. |
| Public record | Company register, regulator filing, procurement disclosure, or comparable public document. |
| Editorial review | Assessed by our editorial team from the material available, without a named external source. |
Field-level source mapping
Evidence is recorded per field, and each field is linked to a specific named source. Where a field has a grade but no linked source, we say so rather than attaching a plausible-looking citation — a wrong link is worse than a missing one, because it makes an unchecked field look checked.
Sources that use a reserved example domain are shown as placeholders. Those addresses never resolve to a live site, so presenting them as citations would be presenting a non-source as a source.
What we list: four different kinds of thing
A directory that mixes these together produces confident nonsense — a reseller described as the product author, a consulting firm described as the technology owner, a bespoke engagement listed as a product anyone can buy. We keep them separate, and each has its own page type.
| Kind | What it is | How it is shown |
|---|---|---|
| AI Agent | A deployable piece of software with a name, a vendor and a licence. | An agent detail page, with deployment, pricing, capabilities and integrations. |
| AI Agent Platform | Infrastructure for building, running or evaluating agents — you compose agents yourself rather than adopting a finished one. | A platform record, compared on orchestration, evaluation and monitoring rather than on business integrations. |
| Enterprise AI Solution | An engagement or delivery-shaped offering scoped to one organisation, delivered by a solution provider. | A solution record, with its nature stated: available product, vendor solution, reference implementation or conceptual approach. |
| Company | A legal entity: an original vendor, a solution provider, a reseller, a consultancy, an implementer. | A company page listing roles, capabilities and corporate relationships. |
Original vendor vs enterprise solution provider
These are two separate relationships and we never infer one from the other.
The original vendor is the party that developed and publishes the product. The enterprise solution provider is the party that delivers that product to an enterprise — through consulting, implementation, integration, private deployment or resale. The same agent can have one original vendor and several solution providers. A solution provider does not become the vendor by virtue of reselling or implementing the product, and we do not change the original vendor because another company sells it.
Agent types
Three classifications are first-class in the data model and are used consistently across discovery, search, filters, agent pages and comparison.
- Specialized AI Agent — built around one defined workflow, with narrow scope.
- General-Purpose Enterprise Agent — broad capability, able to support multiple business tasks. We use this term rather than "all-round agent", which reads as marketing rather than a category.
- Custom / Bespoke Enterprise Agent — an engagement, not a product listing. Described by underlying technology, customization level, deployment options and business requirements.
Type is a classification, not a quality score. A general-purpose agent is not automatically the better choice, and a custom engagement is not automatically more capable — it is a different procurement shape with a different cost and risk profile.
On-Premise is not Private Cloud
These are treated as distinct deployment values, because they carry different obligations for a buyer.
| Value | What it means |
|---|---|
| On-Premise | Software runs on infrastructure the enterprise controls, inside its own environment. |
| Private Cloud | Dedicated infrastructure operated by the provider, typically in a single-tenant cloud arrangement. |
| Vendor-hosted Cloud | Shared multi-tenant infrastructure operated by the provider. |
| Hybrid | A combination of the above across components or environments. |
| Unknown / Not Confirmed | The vendor has not stated which applies. |
Because the difference matters, On-Premise is surfaced in discovery paths, filters, agent cards, agent detail pages, comparison and enterprise procurement information — not buried in a technical specification table.
Why "unknown" is a first-class answer
When a vendor does not publish whether something is supported, the honest answer is "we do not know". Recording that as unsupported is a small lie that compounds: across a shortlist of ten products, it turns several unknowns into apparent weaknesses, and the product that never documents anything looks like the weakest option in the market.
So the comparison interface distinguishes Not Supported (confirmed absent) from Unknown and Not Disclosed (the vendor has not said either way). They are visually different and never collapse into "No".
Industry classification
Industries are a controlled vocabulary with stable identifiers, so a classification change never breaks a URL or orphans a record. Where a record's primary market has not been determined, we say "applies to" rather than inventing a primary/secondary split the data does not support.
Ordering and ranking
Search results are ordered by how closely an agent matches the business task you described. Where two agents match equally, the tie breaks on information status, then alphabetically.
We do not publish an overall score, an editorial ranking or a "best of" list. A single number across deployment, pricing, security and evidence would quietly encode our judgement of what matters in your procurement, and we are not in a position to know that.
Third-party ratings
Where a rating exists, it comes from an external source and is labelled with the number of ratings behind it. Where no rating exists, the page says "No third-party rating yet" rather than showing a neutral-looking placeholder. A vendor with no ratings is not a vendor with bad ratings.
Sponsored placement
Sponsored blocks are labelled, and they are structurally isolated from search: placement cannot enter ordering, scoring or filtering, and evidence status cannot be bought. The full policy, including what the guarantee does and does not cover, is on the sponsorship page.