Swatches follow the palette on screen — L switches between the dark and the light, projector-safe one.
Circle area is estimated monthly web sessions. Estimates are built from Tranco rank data calibrated to published visit figures; error grows as you go down the tail.
Positions are baked once, packed tight around each property's 2026 size. Any other year, the data-volume metric, or the agentic halos relax those positions to make room — pulling apart where something has grown, settling back in where it hasn't — always easing from the true baked layout, so turning everything off returns exactly to it.
Switching the metric to data volume re-weights every circle by a modeled bytes-per-session weight for its category. The numbers in tooltips and cards stay the measured session counts.
Measured. Circle sizes and the 2020–2026 series behind them, and the growth mode built from that series. Positions are measured only at 2026 with no layers on — the baked layout; everywhere else they're a relaxed presentation of that same data, not new data.
Modeled. The data-volume metric. Traffic-category halos, derived from Cloudflare Radar bot-share by content type and scaled across the timeline by Radar's long-run bot-share trend. Circle positions relax under either layer, or when scrubbing away from 2026, to make room — positions off that baseline are presentational, not data.
Modeled classification. The two visibility color modes and the coverage gap built from them. Both are flat recolorings: white is visible, dark gray is not, and the one hot accent is what only modern detection can see. Which properties keyword-based detection could read is a stated classification, not a measurement.
Illustrative. The dark-web cluster. Placeholder geometry and sizes, not measured.
Where a property is marked app-first on hover, web sessions understate real usage — most of all for TikTok.