Growth is easy to compute and easy to fake.

1. We do not start from a list

A trend tool that only charts terms someone already thought of is a search box with extra steps. Every run reads the complete Wikimedia pageview dump for English Wikipedia — every article, every month — and scores the entire corpus. Roughly two million articles clear the traffic floor and get scored. Nothing is on a watchlist; the corpus proposes the candidates.

2. Seasonal control happens before anything expensive

The candidate pass compares each article's recent quarter against the same quarter one year earlier, and against what that article did between those same two quarters last year. An article that climbs every July does not get credit for climbing this July. That one comparison removes most of what would otherwise look explosive.

3. Then we measure properly, on daily data

Survivors get their full daily history pulled and put through five transformations:

• Normalisation. Every series is divided by that day's sitewide English Wikipedia traffic. Holiday dips, outages and platform changes move every article at once; after normalisation they move none of them.
• Weekly smoothing. A seven-day centred window removes the weekday/weekend sawtooth without lagging the series the way a trailing average would.
• Annual decomposition. We estimate a multiplicative day-of-year profile per topic from its own prior years, smooth it, and divide it out. Christmas is a 4× seasonal index on Christmas-related articles, so dividing by it leaves a flat line — which is the correct answer.
• Robust slope. Growth rate is a Theil–Sen median-of-pairwise-slopes fit on the log of the smoothed series. A single outlier day cannot swing it.
• Baseline. Every topic is compared to its own level 12–24 months ago, not to other topics. "Big" and "growing" are different questions.

4. A spike is not a trend

This is the test that matters most, and it is a test of the floor, not the peak. Take the lowest 28-day median in the last 90 days. If that trough is still well above the old baseline, attention has permanently relocated — the thing is genuinely bigger now. If the trough has fallen back, something happened and then stopped happening.

Alongside it we measure concentration — how much of the recent excess attention landed in the single busiest week — and decay, how far the series has already fallen from its peak. A death, an award, a court ruling or a viral post scores high on concentration, high on decay, and flat on the floor test. Those are classified spike and never enter the index.

5. Bots do not read on Saturdays

Wikimedia already strips known crawlers, but a lot of automated traffic gets through as "user". Human attention has a weekly rhythm. We compute the weekend-to-weekday ratio and the day-to-day coefficient of variation; a series that is implausibly flat, or that has no weekly pattern at all, is dropped regardless of how good its growth looks.

6. One story, one entry

Wikipedia will hand you a drug's generic name, its brand name and its mechanism as three separate rising topics. Two entries are merged only when their week-over-week shapes correlate above 0.86 and they share vocabulary or three or more Wikipedia categories. Shape alone would merge unrelated things that happened to rise in the same month.

7. Stage, not just direction

Position on the curve is derived from three numbers: level against baseline, current slope, and where the topic sits relative to its own all-time high. "Up 300%" means something very different for a topic at 40% of its record than for one setting a new record this week.

What we deliberately do not use

• Google Trends. Relative-only (no absolute volume), aggressively rate limited, and the numbers move between identical queries. Fine to sanity-check a single term by hand; not something to build an index on.
• Reddit search. Unauthenticated access is now blocked or throttled to the point of being unreliable. It returns when there is an API key behind it.
• Product Hunt. Needs OAuth; on the list, not in the pipeline.
• Social engagement counts. Trivially inflated, and the platforms that matter no longer expose them freely.

What we do use for corroboration

Hacker News post volume for the term (via Algolia), this quarter against the same quarter last year; npm and PyPI monthly downloads when the topic has an exact package-name match; GitHub stars when a token is configured. Corroboration raises a topic's confidence score. It is never required — most consumer, health and culture trends leave no trace in developer data, and demanding a second source would quietly turn this into a software-only index.

Honest limits

• Wikipedia lags the very earliest phase. Something has to be notable enough to have an article before we can see it, so the truly pre-emergent week or two is invisible here.
• Topics with under about 130 views a day are excluded — below that, daily noise swamps the signal and the seasonal profile cannot be estimated.
• Articles created within the last 200 days are flagged and scored conservatively: the direction is real, the size of the move is not yet trustworthy.
• The index is English Wikipedia only, so it is skewed toward the English-speaking world.