Methodology

1. Where mentions come from

We track public content from creators across YouTube (official Data API; public captions), podcasts (RSS + transcription), X (official API v2) and newsletters (a subscription inbox). Each item runs through an extraction pipeline that identifies tool mentions with their position or timestamp. We store full transcripts privately for processing only — publicly we show quotes of ≤120 characters, always linked to the original source.

2. Sponsored vs. organic

Sponsorship detection runs in two layers. A deterministic rule layer always runs: #ad / #sponsored / #werbung / #anzeige hashtags, sponsor-segment phrases (“this video is sponsored by”, “thanks to today’s sponsor”, “Werbung”/“Anzeige” disclosures), and affiliate-coded links in descriptions (?ref=, ?via=, known network domains). A model-based layer classifies subtler cases. Rules always win: if a rule fires for a tool, its mentions in that item are labeled sponsored regardless of the model output.

• Organic — no payment signal detected for this tool in this content.
• Sponsored — explicit disclosure or affiliate-coded placement detected.
• Unclear — signals conflict or are insufficient; we say so instead of guessing.

The authenticity meter shown on tool and creator pages is exactly this split over their public mentions.

3. Trending score

Each public mention contributes a weight w = w_time · w_reach · w_auth · w_usage:

• w_time: exponential decay with a 14-day half-life — recent mentions matter most.
• w_reach: log-scaled creator reach, capped at 1.0 around one million followers.
• w_auth: organic 1.0 · unclear 0.5 · sponsored 0.25 — organic counts 4× more than paid.
• w_usage: “uses” 1.2 · “recommends” 1.0 · passing mention 0.6 · criticism and switching away count negative.

The trending score sums these weights over the last 90 days; weekly snapshots produce the ▲/▼ deltas on /trending. Category rankings on /best pages sort by the count of creators actively using a tool (mentioned within 120 days), with the trending score as tie-breaker.

4. The money firewall

ActuallyUses earns affiliate commissions on outbound links (always marked ↗ with a disclosure). That revenue system is structurally separated from scoring: the code that computes rankings has no access to affiliate tables, and changes to that rule are treated as a product-breaking bug. Money never changes organic rankings or authenticity labels. Vendors can never buy a position; sponsorships of the newsletter or trending page (if any) are labeled and never affect ordering.

5. Corrections & quality

Every data point carries a report link; corrections are versioned and audited. Creators can claim their page and hide false mentions — each change is logged in an audit trail, and hidden mentions leave scores too. Extraction quality is enforced by a versioned, labeled eval set: a change to prompts or models ships only if precision stays ≥95% (recall ≥80%). AI-assisted summaries are labeled as such; the underlying claims always link to their sources. Data freshness is stamped on every page.