What the internet is actually paying attention to

How the number is built

Each source series is smoothed with a time-aware EMA, then divided by its own trailing 90-day median. The baseline is a median, not a mean — attention data is all spike and no symmetry, and one viral day under a mean-based baseline raises the bar enough to hide the next real move.

Sources combine by weight, renormalised over whichever actually reported, so a missing source shifts the mix rather than silently scoring zero.

What the window changes

The window redraws the charts and the percentage beside each reading. That figure is the move between the first and last points actually plotted — it describes the line you are looking at, and nothing else.

Which is why it can sit a fraction of a point away from the fixed 7d column below. Those columns are exact-date lookups: 7d means the reading seven calendar days ago or nothing at all. The window figure starts wherever the topic's history actually starts inside the window, so on a market with a gap the two span slightly different intervals. The table is the reference; the panel describes its own chart.

Each panel's y-scale is refitted to the visible slice, which is why a chart can look dramatic at 7 days and flat at a year. The dashed line is 1.00 in every panel and at every window — that is the part which never rescales.

Why you can trust the shape

The index was never told about the calendar, and it found it anyway. Star Wars peaks on 2026-05-06 and 2026-05-25 — May the 4th, and the anniversary of A New Hope. Pokémon peaks on 2026-03-01; Pokémon Day is 27 February.

Both lag their event by about two days, which is exactly what a five-day EMA half-life should do. That is the check that makes the rest of these numbers worth reading.

The two numbers are not the same scale

The index is a ratio to a topic's own normal. The price is in degrees, and each market's absolute level is its own history — ebola at 97° is not "worth more" than pokemon at 12°, any more than an index of 1.28 outranks one of 1.02 in popularity.

So the meaningful comparison is between changes, never levels — which is why the price sits in its own strip and stays off the chart, and why its change follows the same window as the index beside it. Noise publishes a 24-hour change and nothing longer, so every other window is computed from our own snapshots, which reaches only 3.3 days back. Where the record is shorter than the window, the cell reports the span it really covers and is labelled with it — a four-day move under a "1y" heading would be a quiet lie.

Where a reading is provisional

A market marked provisional has fewer than 120 days of history, so its "normal" is built from a short and often unsettled stretch. Cursor is the worked example: 78 days, nearly all of it decay from the spike when its Wikipedia article was created, which produces a truthful-looking −80% at 30 days that says more about the article's age than about attention.

The reading is still shown, because waiting for a full window would mean months of blank panels. It is marked because a three-week baseline and a settled year are not the same claim.

What it cannot see

Instagram and TikTok have no usable free API, so topics that live there — looksmaxxing above all — are systematically under-read. That is a missing input, not a tuning problem.

90s-nostalgia is a diffuse mood rather than an entity and has no clean query; it is kept as an honest floor for what this method can do.