US midterms 2026 · Jevlin's record · loading

Election night, as 156 search sessions

On election night, every race was one Jevlin session. Each polling round was a turn, and each turn one call that fanned out to six search engines. Below is what Jevlin kept about those calls: when they ran, what came back, which route answered, how far the sources overlapped. The shape of the data, not the pages.

Developing  The dataset is published once the last race is called; this page is a snapshot of it.

01

The night as threads

One row per race, one mark per turn. Dots: an engine's first page naming the winner; hollow dots: a different name. Black tick: AP's call. Hover to name a row, click to open it, press play to run the night.

Colour Office Engine

02

The six engines, through the night

How often each engine answered, in 15-minute windows; the ticks beneath a sparkline are calls sent straight to the engine's own API. Nearest: the engine it overlaps most, by the domains both cite. Serper and TinyFish overlap most; the others barely overlap.

03

Six engines reach twice the sources

Distinct domains cited per search: the best single engine in that call, and all six together. Medians and .

best single engine in the callall six together

04

Who brought what

Merged results each engine contributed, and how many no other engine found.

found only by this enginealso found by others

05

Where the winning page sat in the list

For each called race, the benchmark records the first page that named the winner. This is where that page sat in Jevlin's merged list at that turn. The ranker orders by corroboration, so pages several engines agree on come first. The first page to name a winner usually came from one engine alone, and it sat low until the others caught up.

Merged rank at the moment of the first find

every engine's first find ()the earliest find per race ()

Who was first, and from where

Races where each engine was first (ties count for each), and the sites those first pages came from.

    06

    One race, turn by turn

    One session as stored: six engine rows across the turns. Each cell is one call, deeper colour for more pages, hollow for a failure, dark edge for the engine's own API. Dots: first page naming the winner. The line above is the consensus score. Click a cell for the whole turn.

    07

    Agreement fades after the call

    Median consensus score by minutes relative to AP's call, over called races. Before the call the engines overlap more; after it the coverage spreads across many outlets and they overlap less.

    08

    What one search contains

    The record stored for every turn, with illustrative values from a median turn. Page text, URLs and titles are in the stored document and left out here.