This page sets out how AI Gear Stack works: how products get selected, how recommendations are reached, what our commercial relationships are, and what we do when we get something wrong.
It is written to be checked against. If our published work does not match what is described here, that is a failure on our part and we would like to hear about it.
How products are selected
Products enter consideration because they are plausible answers to a question a reader is actually asking — not because they are new, not because they are heavily marketed, and never because a commercial arrangement exists.
For each guide we define the buyer first: their workload, their budget band and their hard constraints. We then assemble the candidate set from:
- current and recent hardware that meets the technical requirements of that workload
- widely available products, from manufacturers with a track record of firmware and driver support
- previous-generation hardware where it remains genuinely competitive, which in AI work it often does
Products are excluded when they cannot do the job, when availability is unreliable, or when the specifications cannot be verified from the manufacturer.
A guide is allowed to conclude that nothing fits. If a budget will not buy a machine capable of the stated workload, we say so rather than recommending the least-bad option.
How recommendations are reached
Our review methodology sets out the scoring rubric in detail. In summary:
- Establish what the workload actually demands — memory capacity first, memory bandwidth second, compute third.
- Assess each candidate against those requirements, not against a general performance ranking.
- Score across nine fixed criteria so that scores are comparable between reviews.
- Write the recommendation for a specific reader, and say plainly who should not buy it.
Every recommendation on this site names the person it is for. A pick with no stated buyer is not a recommendation, it is a ranking.
Basis of assessment
Every review and guide states, on the page, what the assessment rests on. There are exactly three categories and we do not blur them:
- Specification analysis — assessed from manufacturer specifications and published architectural detail. We have not physically tested the unit.
- Third-party benchmark synthesis — specifications combined with independently published benchmark data, cited inline. We have not physically tested the unit.
- Hands-on testing — tested in-house against our published methodology.
We will never imply hands-on testing that did not happen. Where a figure is a manufacturer claim, it is described as one.
Affiliate relationships and independence
AI Gear Stack earns commission when readers buy through links on the site. Full detail is on our affiliate disclosure page. Three commitments follow from it:
Commission rates never influence recommendations. Editors do not see commission rates when writing. Products that pay nothing — including manufacturer-direct purchases and retailers we have no relationship with — are recommended on identical terms.
We do not accept payment for coverage. No paid reviews, no paid placements inside editorial content, no paid scores. Any sponsored content, should we ever publish it, will be labelled as such at the top of the page and excluded from editorial recommendations.
Commercial links are visually distinct from editorial links. Purchase buttons look like purchase buttons. Links to our own analysis look like text links. A reader should never click a recommendation and be surprised to arrive at a retailer.
Use of AI in our editorial process
We use AI tools in research and drafting — for summarising specification documents, checking arithmetic, and drafting structure.
We do not publish AI-generated content unreviewed, and we do not operate any pipeline that turns generated text into a live page automatically. Every published URL has passed through fact check and affiliate review as separate, human stages, and our publishing system will not allow a page to go live before it has.
Responsibility for everything published here sits with the named author and reviewer, not with a tool.
Data sources
Specifications come from manufacturer documentation wherever possible. Where we rely on third-party measurements, the source is named and linked in the sources list at the foot of the article.
We do not reproduce retailer product descriptions or rewrite manufacturer marketing copy.
Pricing and availability
We do not publish prices we cannot verify from an authorised source at the time you read the page. A price captured at writing is wrong within days, and a wrong price is worse than none.
Instead we classify products into broad price bands and link out so you can see the current figure yourself. We never claim a price is the lowest available, never suggest stock is limited, and never manufacture urgency.
Updates and corrections
Hardware guidance decays. Guides are reviewed when new hardware lands in the category, when a recommendation stops being true, or when a reader tells us something is wrong.
Where an update changes a recommendation, the page carries an updated date. Cosmetic edits do not reset that date — an “Updated” badge that means nothing is worse than no badge.
Material errors are corrected openly and logged on our corrections page.
Who is responsible
Every article names its author. Where a second editor has checked the work, that reviewer is named too. Author pages set out relevant background.
We do not invent contributors, and we do not publish under fabricated expert personas.