Editorial policy

This page describes how content on RunLocalAI is created, reviewed, and maintained. We publish it for transparency and as a quality contract with our readers.

Sources of truth

Every factual claim on the site comes from one of these sources:

  • The model or tool's official documentation, repository, or model card
  • Published research papers (cited)
  • Benchmarks we ran ourselves on the hardware listed in About
  • Named community benchmarks, with the source URL on every published row
  • Manufacturer specifications for hardware

Pages tag every benchmark row with its source: owner (we measured it), community (with citation), or official (vendor-published).

AI-assisted content

We use AI to help draft and structure pages — writing introductions, generating FAQs from structured data, and summarizing model cards. A human editor reviews every page before publication for technical accuracy, citation correctness, and licensing claims.

We do not auto-publish brand-new top-level pages without human approval. Routine maintenance updates (refreshing GitHub stars, adding a new benchmark to an existing page, fixing a broken link) can be applied without human review and are logged in our audit trail.

Verification cadence

Each page shows a "Last verified" date. Pages with active traffic are reviewed at least every 90 days. Stale pages (no traffic for 6+ months) are flagged for editorial decision — either refreshed or deindexed, never silently abandoned.

Corrections

If you spot an error — a wrong VRAM number, an outdated GitHub stat, a misattributed benchmark — please email us via the contact page. We correct verified errors within 7 days and add a brief correction note to the page.

Independence

Affiliate relationships and ad networks do not influence our reviews. We have never accepted payment in exchange for coverage. When we recommend a tool, model, or piece of hardware, it is because we use or have tested it. See How we make money for full disclosure.

Content removal and deindexing

We may deindex pages that:

  • Become substantially duplicative of another page (similarity threshold 75%)
  • Cover models or tools that have been deprecated or abandoned
  • Have not generated traffic for 12 months

Deindex actions are logged in our audit trail with the reason and date.

What we refuse to do

A handful of patterns are common in the AI-content world that we explicitly do not adopt — they erode reader trust and we treat them as bright lines, not tradeoffs.

  • No fake aggregate ratings. We never publish a "9.2/10 from 142 reviews" line that isn't real. The Pros/Cons schema we ship is drawn from the actual editorial whoIsItFor / whoShouldSkip arrays — never invented for SERP appearance.
  • No fabricated benchmarks. Every tok/s, watt, and dB figure on the site comes from either our own testing harness or a cited community / vendor source. Pages that need a number we don't have describe the directional truth ("runs hot under load") rather than invent a value.
  • No fake urgency. No "limited time," no countdown timers, no "buy now before prices change" framing. Hardware prices do change; we note that with a freshness label, not with manufactured pressure.
  • No sponsored content presented as editorial. If a vendor ever pays for a placement (they have not), it will be labeled sponsored and excluded from buyer-guide rankings.
  • No AI-generated reviews of products we haven't actually used. AI helps us draft and structure prose; the editorial recommendation always comes from a human who has tested or operated the thing.

How buyer-guide rankings get decided

Hardware buyer guides rank picks by operator-grade leverage at the buyer's actual price tier, not by manufacturer wattage / TFLOPS / generation alone. The decision lens for any "best GPU for X" page:

  1. VRAM ceiling for the workload (the gate — if the model doesn't fit, speed doesn't matter)
  2. Memory bandwidth (the speed limit once VRAM fits)
  3. $/GB-VRAM at current market prices (where the value actually lives)
  4. Runtime ecosystem support (CUDA, ROCm, MLX — does day-zero new-model support arrive on this card?)
  5. Power, noise, heat envelope (operational reality, not a spec sheet)

If a card wins on rule 1-3 but loses on rule 4-5 in a way that matters for the audience, the verdict says so. Our job is the buyer's decision, not the vendor's marketing.

Why we recommend used silicon

Used 24 GB cards (notably the RTX 3090) remain the highest-leverage AI buy in 2026 by a meaningful margin. We say so. Some publications avoid used- market recommendations because they're harder to monetize via affiliate links. We accept lower per-click revenue on the used recommendation in exchange for keeping the editorial leverage honest — that's the trade we made when we wrote the editorial policy, and we maintain it.