How Vibedex Tests Creative AI Platforms

What we measure

Most AI tool round-ups make subjective claims that don't survive a second look. We restrict ourselves to facts a prospective buyer can verify before opening their wallet:

• Capability presence — does the platform actually offer this thing? Yes / no / claimed / unknown.
• Free-tier access — no signup, signup required, free credits, paid only, or unknown.
• Export status — clean export, watermark, blocked behind paywall, or not tested.
• Paid floor — the cheapest paid tier that unlocks the capability.
• Commercial-use rights on free output — yes, restricted, no, or unknown.
• Named-feature bundles — when a platform markets a single-named workflow (e.g. "Product Photography"), we check whether the page actually demonstrates the steps it claims.
• Trust signals — G2, Reddit themes, App Store ratings, billing/cancellation complaint share.

What we deliberately don't measure

We don't subjectively rate output quality across platforms. Most platforms put their best output behind paywalls, so any honest cross-platform quality comparison would require paying for every tier of every platform on identical prompts. We don't have that access, and pretending we do produces the same generic-listicle output Vibedex was built to replace.

We also don't score specific failure modes (e.g. "hands", "character consistency", "lip sync") cross-platform. When these matter for a specific use case, we mention them qualitatively with sources — never as a number.

Evidence labels — every claim is graded

Every cell in our dossier carries an evidence label so readers can tell apart what we actually saw from what a platform claimed in marketing:

Trust signals — and what we don't use

Trust verdicts are derived from a fixed rule, never from gut feel:

• G2 reviews — primary score when sample size is ≥ 50. The most useful single signal for B2B-adjacent creative platforms.
• Reddit themes — top 3 positive + top 3 negative qualitative themes per platform. We flag when the top 2 negatives are both trust-related (billing, account, scam).
• App Store ratings — only used for mobile-first platforms (Fotor, Picsart, CapCut), and only when sample is ≥ 100.
• Trustpilot — used only as a risk flag for billing/cancellation complaint share. We do not publish Trustpilot scores. Our own cross-channel analysis found TP scores systematically run 2–3 stars below G2 averages for creative-AI platforms, because angry users write reviews and happy users don't.

A platform is rated Pass when G2 ≥ 4.0 with n ≥ 50, billing risk is at worst cautionary, and Reddit negatives don't dominate trust themes. Auto-fail requires G2 < 3.0 (n ≥ 50) and high-severity billing risk. Everything else is Caution — usable with reservations stated in the article.

Verdicts — computed, not authored

For each use case (e.g. "product hero shots for catalog listings"), each platform gets a verdict computed deterministically from the dossier — not written by an editor:

• Specialized — a verified named workflow covers ≥ 80% of required capabilities, no required-capability export is blocked, and trust isn't auto-fail.
• Has the parts — ≥ 80% required capabilities present, no required-capability export blocked, but no verified named workflow.
• Partial — 40–79% required capabilities present.
• Not for this — < 40% required capabilities, OR trust is auto-fail, OR a required-capability export is blocked.

Editorial judgment lives in the prose around the verdict — "strong on this, weak on that" — never in the verdict label itself. If an editor disagrees with a verdict, we change the rule, not the cell.

Scope

The v2 framework was locked on 2026-05-26 with the following scope:

Platforms are tested in waves: e-commerce + image-heavy first, then video + audio, then editors and remaining suites. Articles ship as their underlying dossier data lands. Each published article carries its own "last updated" date so you know when Vibedex last visited the live platforms behind its claims.

Reproducibility and refresh

Every ingest appends a snapshot to an audit log, so we can reconstruct what a platform's dossier looked like on any past date. When a platform updates pricing, free-tier mechanics, or named features, we re-run the collection protocol and the new state overwrites the current dossier — the old state stays in the audit log.

If you see a claim in a Vibedex article that no longer matches the live platform, the platform has updated since our last visit. The article's last updated date is the answer to "when did Vibedex last check this".

What this isn't

This is not a model benchmark. We don't score Midjourney vs Flux vs Nano Banana on identical prompts — that's the Vibedex image-model benchmark, which is a separate dataset using Gemini 3 Pro as judge across 200 prompts × 18 models. The Creative AI Platform benchmark described here is about platforms — the apps and product surfaces buyers actually subscribe to.

It is also not a leaderboard. A single ranked list across creative platforms is misleading because platforms specialize in different use cases. Our verdicts are always per-use-case, never platform-overall.