Methodology · AI Trend Market

AI Trend Market is a forecast desk, not a news feed. Every trend is scored from dated, sourced evidence so you can act before consensus. Here's how the numbers work.

Trend Score (0–100)

A weighted blend of signals, normalized 0–100:

Velocity — how fast evidence and search demand are accelerating (a 7-to-90-day delta, not lifetime volume).
Freshness — how recent the evidence mass is.
Source breadth — how many independent source classes agree (GitHub + HN + a company blog beats one random post).
Novelty — how early and non-generic the topic is.
Business ROI — whether there's a clear, monetizable wedge.
Authority & engagement — credibility and real interest.
Saturation (inverted) — low saturation lifts the score.

Low-saturation-but-rising trends get a deliberate boost — those are the gems worth catching early.

Saturation bands

0–30% Unsaturated — room to be early.
31–60% Competitive — viable with a sharp angle.
61–80% Crowded — high demand, hard to stand out.
81–100% Overheated — likely past the opportunity.

The stages

Trending — highest current momentum, ranked by score.
New — first spotted in the last 7 days.
Still Hot — older than a week but still rising.
Gems — low saturation + rising demand.
Majors — moves from major labs (OpenAI, Anthropic, Google, Microsoft, NVIDIA, xAI, DeepSeek, Perplexity).
Open Source — open platforms, models, and repos.

Skill Hot Score (0–100)

Skills are scored from real signal, not editorial guesses — the same proof-first stance as trends. A skill's Hot Score blends:

outranks a 10k one (no flat cap).

repo sinks; a live one rises. This is the "active repo" test.

The skill is matched to its best active trend, and that trend's score lifts it.

Adoption — repo stars and forks, log-scaled so a 150k-star project genuinely
Activity — how recently the repo was actually pushed. A popular-but-abandoned
Maintenance — open-issue load relative to audience and recent activity.
Trend momentum — whether the skill rides a hot, rising trend on the board.

The blend is gated by a confidence factor: a skill with a live, API-verified repo scores at full confidence; a link-only entry with no repo and no trend match is discounted so it can't headline on numbers it can't back up.

We only list genuine reusable skills — Claude skills, MCP servers, Cursor rules, prompt packs, agents, and workflows. General ML frameworks, model/checkpoint repos, and standalone chat apps are filtered out.

Date integrity

We separate three timestamps and never conflate them:

published_at — when the world saw it (drives freshness and momentum).
spotted_at — when we first detected it.
collected_at — when our last refresh fetched it.

Undated sources (some web-search results) are treated as supporting evidence only — they never inflate velocity.

Sources