Independence and conflicts of interest
Pickbench is an independent research publisher. What is never sold, how we make money, how we measure, and how we disclose conflicts of interest.
Last updated: October 17, 2026
What we are
We measure how AI models choose between brands and publish the results. We publish research about brands whether or not they asked to be measured and whether or not they are our customers, and we decide what to measure and publish ourselves.
Never sold
No one can pay for any of these:
- a rank or tier, or being included or left out;
- a verified status or level;
- placement or priority in what we serve to AI agents;
- influence on battle or Consilium answers, or on which scenarios count in headline results;
- another brand's private data, or any visitor's personal information;
- sponsored slots or badges;
- a correction, a brand statement, or faster handling of either.
How we make money
We sell category reports, data subscriptions, agency plans, custom research and related services to businesses, including brands that appear in our published results. That is a possible conflict of interest, and the rules on this page are how we manage it.
If a brand in a published result pays us, or an agency pays us for work on behalf of that brand as its client, we list the relationship in our conflicts register and mark it next to that result. This covers report, subscription, agency-plan and custom-research customers.
No fee depends on what a public result shows. Fees for paid work are fixed in writing before the work starts; any performance-based part uses only private metrics agreed in that contract, never a public rank, tier or result.
How it is enforced
The code that computes rankings reads only pre-registered plans, stored model results it cannot change and judge records; it never reads billing, sales, customer or affiliate-commission data. Separate database roles and automated tests check this in our code.
When we add shop links, they will be attached after ranking, and shops will be sorted by price, never by commission.
In sales conversations, we quote only a brand's consensus tier with its interval, never one model's result picked out, best or worst.
Panel and gateway
Our official editions currently run on our fast panel: current AI models from several vendors, called through their APIs via the OpenRouter gateway, with data collection by providers denied. Each edition names its panel, the models, the conditions and the date. We never automate consumer chat apps.
We do not use model outputs to train or improve any AI model.
How results are stated
- Every result page names the API models, the conditions and the date, never a consumer app. The choices are made by those AI models; we only aggregate them.
- Shares come with confidence intervals, and brands are grouped into tiers rather than strict ranks. Brands are ordered only by the measured choice share and its interval, and Consilium offers by price; no payment or business relationship changes either order.
- The analysis plan is fixed, and its hash recorded, before a run. The buyer scenario is the unit we count.
- Results are published as signed snapshots that anyone can check against our public keys.
Corrections
We correct errors and list every correction, with its date and reason, in our public corrections log. The same process and evidence standard apply to every brand, and payment never affects it. The corrections policy in the terms
Conflicts register
No brand, agency, AI model vendor or investor currently pays us or holds a financial interest in us, so this register has no entries. If that changes, we will list each relationship here and next to the affected results.
Ownership and funding: Pickbench is operated independently; operator legal entity and funding details will be published on this page when available.
What we do not measure
We do not measure or publish results in these categories, where a recommendation could cause particular harm: prescription or medical treatments; specific credit, investment and insurance products; firearms; gambling; and political content.
Limits we state
- API results are not what consumer chat apps or real shopping agents do.
- Visitor votes are an open poll of self-selected visitors, not a representative survey.
- A shift in AI preference is not a shift in sales, so we make no traffic or sales claims.
- Pages marked Demo data use fictional brands and simulated choices, and are never mixed with real results.
Changes
When this page changes, we post the new version here with its date, and the list below says what changed.
- October 17, 2026: final text published. Earlier versions were never in effect.
Contact
This site is operated by Pickbench (the operator; its legal entity details will be published on this page).
Postal address: will be published on this page when available.
Legal requests, corrections and questions about these pages: [email protected]
Privacy requests: [email protected]
To request access, correction or deletion (including Pick Bench records), email [email protected] and include the client id shown at /en/legal/privacy/my-id/.
Operator legal details will be added here when available.