A Scroll, Not a Read
Every counter you inherited was built to go up, because arrival is the easiest thing to count and the most flattering. On building measurement that can refuse, proving the refusal, and why the honest number on this site is the one I declined to collect.
By Jason Waldrip 16:20 7 min read
I pulled up one of my own essays on my phone, put my thumb near the bottom of the screen and flicked. Two thousand words went past in about a second. A slip came up in the corner and told me no. Derailed. That was a scroll, not a read. Back to the top to re-rail.
I wrote that refusal. Everything else about the feature was straightforward, and that one sentence took most of the thinking.
Here is the part worth sitting with. On that same page load, before my thumb moved, a beacon had already gone out to Google. This site runs GA4 pageviews and nothing else, a few lines of inline script that fire on load. By the standard of that number, I read the essay. By the standard of the other counter on the same page, I never opened it.
One of those two numbers reports back to me. It is not the one that knows anything.
arrival is the cheap thing to count
Every analytics default you have ever inherited counts arrival, because arrival is a single event at a known moment and attention is a judgment call spread over minutes. So the industry standardized on the easy one and then argued about the dashboards.
Chartbeat went and looked. Across two billion visits in a month, Tony Haile reported that “a stunning 55% spent fewer than 15 seconds actively on a page,” and that filtering down to article pages only improved it to roughly one visitor in three under fifteen seconds. His other finding is the one I think about more: there is no relationship whatsoever between how much a piece gets shared and how much attention an average reader gives it.
That was 2014. The counters did not change. The industry just got better at filling them.
what this counter refuses
The read tracker on this site measures position in words, not pixels, and it only counts ground the reader has actually covered. A piece marks read when two things are true at once: the end is on screen, and enough visible time has accumulated to have plausibly read the thing at 650 words a minute. There is a four second floor underneath that, because even a one-line take needs a few seconds on screen.
650 is deliberately generous. Marc Brysbaert’s meta-analysis of reading rate puts adult silent reading of English non-fiction at about 238 words a minute, and his headline point is that the real numbers come in well under what gets quoted in both the research and the popular writing about it. So the threshold sits at nearly three times the measured average. I would rather credit a fast reader than accuse a real one.
Derailing is stricter about one thing only. A fling is half the piece or more coming on screen faster than ten thousand words a minute, across at least two screens of scrolling. Skimming is fine. Skipping a section is fine. Scrolling back up to check something is fine, because progress is measured against a high-water mark rather than your current position, so re-reading covers no new ground and can never look like a flick. The train comes off the line for a fling and for nothing else.
And when the browser refuses the write, because storage is full or disabled, the tracker reports a refusal instead of a read. The comment above that function says why: so the caller never claims a read it lost.
the negative case is the test
A measurement that has never been shown rejecting something is decoration. So the proof script for this feature does the obvious half, drives a real article through the real tracker at a reading pace and checks that it lands, and then does the half that actually matters. It finds a piece long enough to flick through, flings it, waits out the full dwell the piece would have needed, and asserts two things: the read was not stored, and the count did not move.
That assertion is the feature. The rest is plumbing.
I ran into the same discipline building the film pipeline, where a machine judge only gets a vote after it has been proven on a known-bad example, a blue sailboat standing in for a shot that was supposed to be a red delivery van. Different domain, same question. What have you watched this thing correctly reject?
the question I should have been asking in design reviews
For twenty years I have read metrics the way everybody reads them. Is the number going up, is it going up fast enough, what moves it. I have built dashboards for boards and for standups and I have been right about plenty of them.
The question underneath, the one I never put on a review template, is this: what does this number refuse to count, and can my own output move it?
At Brandfolder I walked into a product crashing on a memory leak. Features shipped that quarter would have looked respectable. Story points, commits, releases, all of it defensible on a slide. Uptime was the one number my team’s effort could not flatter, because it was not measuring us at all. It was measuring whether the thing worked for the people using it. Fix the leak and it moves. Ship four features on top of the leak and it does not.
That used to be a nice distinction. Now it is the whole ballgame, because volume is free. Posts published, pull requests opened, tests added, lines changed, tickets closed: I can move every one of those by lunch with a fleet, and so can you, and none of us should get credit for it. The numbers that survive that are the ones measuring somebody else’s behavior. Did it work. Did it hold. Did anyone read it.
the number I decided not to collect
Here is the part I expect to get argued with about. The read log lives in local storage on the reader’s own device and goes nowhere else. No account, no server, nothing leaves the browser. The page says so out loud, under the counter: your progress stays in this browser.
Which means I built the honest counter on this site and then declined to collect it. I can tell you precisely how the mechanism works because I wrote it. I cannot tell you how many people have finished this essay.
That was not an oversight and it was not only privacy. Charles Goodhart made the underlying point about monetary statistics in 1975: an observed regularity collapses once you put pressure on it for control purposes. The version everyone quotes, that a measure which becomes a target stops being a good measure, is Marilyn Strathern’s, from a 1997 paper about auditing British universities. Fifty years of people nodding at it, and for most of my career it was still a slow-motion risk, something a team drifts into over quarters. The drift is not slow anymore. Point a fleet at any target you can see and it will be hit by Friday, honestly, without anyone deciding to cheat. If “reads” showed up on a chart I check, I would start writing for the chart, and the counter would turn into one more number my own production can move.
So it reports to the one person it should report to, which is the reader. The queue on this site already taught me what happens when I widen the stage that was never narrow. I did not need another gauge pointed at my own output.
If you got this far by reading it, your browser knows and I do not. That seems like the right way around.
Jason Waldrip is a builder, entrepreneur, and technologist. He takes on fractional CTO and CAIO work through The Bushido Collective, for founders drowning in AI-generated code and teams scaling past the leadership that got them here. If your dashboard is full of numbers your own output can move, that is the engagement. Work with me.
Written with AI assistance. The frame, the calls about what counts, and the decision not to collect it are mine.