Editorial Policy

Purpose

AIInsider exists to help people who build with AI make better decisions faster. That means accuracy over speed, specificity over hedging, and usefulness over engagement. If a piece would get more traffic by overstating what a model can do, it gets published understated instead.

Accuracy and sourcing

Claims are checked against primary sources — model cards, official documentation, published papers, release notes, pricing pages, source repositories, or our own testing. Secondary coverage is a lead, not a source.
We do not invent numbers. Benchmark scores, prices, context-window sizes, funding amounts, parameter counts, version numbers and dates are either verified and attributed, or they do not appear. If we cannot attribute a figure, we describe the situation qualitatively instead — "materially cheaper per token" rather than a made-up dollar figure.
Benchmark results are reported with their source and, where it matters, the conditions they were produced under. Numbers from different harnesses are not presented as directly comparable.
Pricing is usually described as tiers — free, individual, team, enterprise — because exact figures go stale within weeks. Where we do cite a price, it carries the date it was checked.
Vendor marketing claims are labeled as vendor claims until they are independently verifiable.
Where something is genuinely uncertain, contested or moving, we say so in the text. Hedged accuracy beats confident invention.
We do not fabricate quotes or attribute statements to named people who did not make them.

Durability over volatility

Version numbers and prices change monthly; architecture, evaluation methodology and failure modes do not. We deliberately weight articles toward the durable half of a subject, so a piece stays useful after the next model release rather than becoming wrong.

How we test

Reviews and comparisons are based on hands-on use, not feature tables copied from a pricing page. Our approach:

Tools are tested on realistic tasks in their category — an AI coding assistant against an actual codebase, an agent framework against a multi-step task with real tool calls, a RAG stack against a corpus with awkward documents in it.
Competing tools in the same comparison get the same tasks, the same prompts where applicable, and comparable effort spent on configuration.
We use the tier a normal reader would use. Where an enterprise tier changes the answer, we say which tier we tested.
We record what broke. A review that finds no weaknesses has not been done properly.
Code in tutorials is run before publication. If a snippet depends on a specific SDK version or model, that is stated.

Ratings

Where a tool carries a rating, it is an editorial judgment made against consistent criteria within its category — capability on the core job, output quality, reliability, developer experience and documentation, integration surface, transparency about data handling, and value for the price. Categories are rated against their own peers: a specialized tool is not marked down for lacking features it was never meant to have.

Ratings are set by the desk that tested the tool. They are never set by, negotiated with, or sold to a vendor. Ratings can move up or down when we retest, and a retest is triggered by a major release or by credible reader reports that a tool has changed.

How tools get into the directories

Entries in the AI tool directory and theAI agent directory come from three routes: our own tracking of the space, reader recommendations, and vendor submissions via thecontact page. All three are assessed the same way.

To be listed, a tool must:

Be publicly available and usable — not a waitlist page or an unreleased demo.
Do something meaningfully distinct within its category.
Publish enough about pricing and data handling for a reader to evaluate it.
Be functional when we check it. Abandoned and broken tools get removed.

Listings are free, and inclusion cannot be purchased. Ordering reflects editorial assessment and relevance, never payment. Some outbound links are affiliate links, which does not affect inclusion, ordering or rating — the full position is in thedisclaimer. Entries are reviewed periodically and removed when a product is discontinued or no longer meets the bar.

Independence and conflicts of interest

We do not accept payment for coverage, placement, ratings or favorable framing.
No vendor, sponsor or advertiser has pre-publication review rights.
Advertising and sponsorship, where present, are handled separately from editorial and are visually distinguishable from articles.
Anyone contributing to AIInsider must declare a financial interest, employment, consulting relationship or equity position in a company they would be writing about. Where a material conflict exists, that person does not write or edit that piece; where coverage is unavoidable, the relationship is disclosed in the article.
Free trials, review licenses, API credits and briefing access are accepted so we can test properly, and are disclosed where relevant. They never buy an outcome.
Where the interests of a vendor, an advertiser or our own traffic conflict with the interests of a reader trying to make a decision, the reader wins.

AI assistance in our own process

We cover AI, so we should be direct about using it. AI tools may be used for research support, summarizing source documents, drafting, editing and code review inside our workflow.

Every published article is reviewed by a human who is accountable for its factual and technical accuracy.
No article is published on the strength of a model's output alone, and no claim originating from a model is published without being traced back to a primary source.
Model output is never used to generate benchmark numbers, prices, quotes or citations. Those come from the source or they do not run.
Code and configuration in tutorials is executed by a human before publication, regardless of what produced the first draft.
We do not accept unsolicited, fully AI-generated article submissions.

Corrections

Factual errors are corrected as soon as they are confirmed.
Substantive corrections — anything that changes a conclusion, a rating or a recommendation — are noted on the article with the date of the change.
Typographical and formatting fixes are made silently.
Articles revised to reflect new model releases or changed pricing carry a visible updated date.
Where a claim turns out to be unsupportable, we remove it and say that we removed it, rather than quietly softening it.

Report an error via the contact page or by emailinginfo@riocloudsolutions.com with the article URL and the specific claim.

Unpublishing

We do not remove accurate published reporting on request, and we do not delete a negative review because a vendor has since fixed the problem — we retest and update instead. Where a legal or safety concern is raised, we review it, and any resulting change is recorded on the article rather than made silently.

Authorship

Articles are attributed to the editorial desk responsible for them — Editorial for news and industry analysis, Research for model and lab coverage, Engineering for tutorials and agent work, Reviews for tools and comparisons. Each desk owns technical review in its area. Attribution is never used to imply credentials that are not held. The desks are described on the about page.

Feedback on this policy

This policy is a commitment, not a formality. If you think we have broken it, say so —info@riocloudsolutions.com. Related pages: about, disclaimer and theterms and conditions.