The year that was seven months long
Showing the misleading chart
A bike counter’s fourteen annual totals on a zero-based axis, nothing smoothed, nothing left out — and the last bar down 28.0%, the steepest fall outside 2020. The file behind it stops at 23:00 on 31 July, because the city refreshes it monthly. So the 2026 bar holds 212 days against every other bar’s 365 or 366, and 5,088 hourly readings where a full year has 8,760, drawn at a full year’s width. Compare January-to-July with January-to-July and the same 741,283 crossings are 18.4% up on 2025 and the busiest January-to-July the counter has ever recorded.
01The claim
Fourteen years of a bike counter, one bar per calendar year, and the axis behaving itself throughout: it starts at zero, runs to 1,200,000, has no break and no truncation. No year is left out from 2013 onward, nothing is smoothed, nothing is indexed, there is no second axis, and every bar is the same inductive loops in the same two sidewalks of the same bridge, summed the same way. Every value is printed above its own bar. The counter held above a million as recently as last year — 1,029,510 crossings in 2025, the fourth straight annual rise. This year it has logged 741,283. That is a fall of 28.0%, the second-steepest in fourteen years of counting and the steepest outside 2020, and it hands back four years of recovery in one: the bridge now sits 37.6% below its 2019 peak of 1,187,146, roughly where it stood in 2021. Every other fall in the record is smaller than 8%. Read-out for the board: hold the west-approach widening until demand recovers, and re-commission the 2019 promotional campaign, the last intervention on record that coincided with a rise. And note what we were careful about at the other end of the series. The loops were only switched on in October 2012, so that year holds a bare three months and would draw a misleadingly short bar — we start at 2013, the first complete year.
02The trick
The slide caught the partial year at the start of the series and plotted the partial year at the end of it. Seattle refreshes this feed monthly; its own metadata gives the refresh frequency as monthly and the last refresh as 5 August 2026, and the newest reading in the file is 23:00 on 31 July. So on the day the query ran — 31 August 2026 — the 2026 bar was not a year. It was 212 days of one, 5,088 hourly readings where a full year holds 8,760, which is 58.1% of a year drawn at 100% of a year’s width. The x-axis comes from the file and the height comes from whatever had been counted when somebody pressed download, so the missing five months arrive as a fall rather than as an absence, and they arrive in the position a reader trusts most, the end, where the story is supposed to be. Nothing else on the slide is wrong: the axis really does start at zero, every year really is plotted, the counts really are the counts. And the shortfall is about the right size to be believed — August to December has run at 40.2% of the calendar year across the thirteen complete years, so a bar cut on 31 July should come in roughly 40% light, and 28.0% is comfortably inside what a bad year could plausibly look like. Two things give it away without leaving the page. The 2026 bar is shorter while the traffic underneath it is heavier: 3,497 crossings a day so far this year against 2,821 a day across the whole of 2025, which is 24.0% more. And for the decline the chart shows to be real, August to December 2026 would have to come in under 288,227 — a figure only two August-to-Decembers in the record have ever undercut, 2020’s 281,246 and 2021’s 286,490, against a thirteen-year average of 379,295. One caveat belongs beside the record rather than in a footnote: Seattle publishes this feed as raw data and now titles it a combined bicycle and scooter counter, so “crossings” means detections by the inductive loops rather than bicycles specifically. That convention applies identically to every year on both drawings, and no step change appears anywhere in the 166 months of the series — the 2026 gains are spread across all seven months — so the comparison holds; it is the word “bicycle” that would be the overclaim. (Both drawings are ours; the transport desk on the first is invented, and every count on both is real.)
03The fix
Compare like with like, which here means the same slice of the calendar every time: 1 January to 31 July, every year, on one zero-based axis. That chart needs no caveat, because every bar on it is a complete measurement of the same thing — and it reverses the finding outright. The same 741,283 crossings are 18.4% above the 626,242 of January-to-July 2025 and 8.7% above the previous best, 2019’s 681,720, which makes 2026 the busiest January-to-July the counter has recorded in fourteen years of counting. The collapse and the record are the same number; only the comparison changed. The other honest option is to end the axis at the last complete period and say so in the caption — a chart that stops where the data stops is never wrong about where the data stops. If the unfinished bar has to appear, give it different ink and not merely a different footnote: hatch it, lighten it, label it with the span it actually covers, and put the cut-off date on the axis where the reader is already looking, because a caveat in the text and none in the picture is a caveat nobody reads. Resist annualising by multiplying up. Scaling 741,283 by twelve-sevenths gives 1,270,771, which assumes the missing months are average months, and on a bridge in Seattle they are emphatically not — August to December carries 40.2% of the year rather than the 41.7% the calendar would suggest, and the shape of the season decides the answer. Projecting on the shape the last thirteen years actually had gives about 1,239,000, still a record and still an estimate, which is why the honest drawing marks it as one and puts it in a dashed frame rather than a bar. Keep this apart from the cherry-picked window, which it resembles: a cherry-picked window is a decision about where to start and stop, and someone made it. Nobody chose this endpoint. The download did, and it moves on its own, which is why the same file drawn on 31 January says −93.6%, on 31 March −78.8%, on 31 May −56.4% and on 31 July −28.0% — four different findings from one unchanged dataset, all of them about the day the query ran. And keep the two flavours apart too, because they want different remedies: a period that has not finished on the calendar is fixed by looking at a calendar, while reporting lag bites periods that have already ended, and is fixed by plotting only the periods old enough to have settled and saying in the caption how old that is. On a dashboard, where the edge moves every morning, fix the window rather than the endpoint — a trailing twelve months, or this period against the same period last year — because those stay honest without anybody remembering to re-check them.