The Cost of Being Wrong in Public

01A mistake travels on its own

A wrong number does not need help to spread. It sits on a page, gets read, gets repeated, and every copy of it is another place it can be found. The correction has to do all of that work again from zero, starting later, with less attention, and usually without the people who saw the first version ever coming back to see the second. A correction reaches a fraction of the people the mistake did, and none of the machines.

That last part sounds harsh, so here is what I mean by it. A machine that read your page last month is not watching for edits. It has whatever version it took, and the new one only exists for it once it comes back, reads again, and the new reading outweighs the old one. For a search index that can take days. For the memory of a trained model it can take until the next model.

The one that reached the public

One of my sites had an article about search updates whose history was simply made up: an update that never happened, and dates that did not match the official record. Nobody had checked it against the source. I rewrote it in July against the official status page, every date, every duration. That same article later became the page one assistant read most on that site, 378 times in fourteen days. I cannot know what it kept from the version it read before the fix, and that is exactly the point. Once a machine has read your mistake, you no longer control which version it remembers.

02Fixing it is not the same as correcting it

Silent fixes

The usual instinct is to change the number and move on, because a visible correction feels like an admission. But a silent fix leaves every reader who saw the old version holding it, and it gives a machine nothing that says the page changed. A dated note does. It tells a person what was wrong, and it gives any crawler that comes back a clear reason to replace what it stored. If the wrong figure also lived in other places, the rules from two sources disagreeing apply: change every copy, because the version in more places wins.

Floors survive, peaks do not

The other kind of wrongness is slower. When the portfolio grew from 21 products to 22, some pages on this blog still said 21, and I left them. They were written as a floor, at least 21, which is still true and can be checked by counting. That is the habit from publishing claims with numbers: a claim built to be true later does not need correcting when the world moves. The cheapest correction is the one you made unnecessary when you chose how to phrase the claim.

03Catch it the night before

Two that never went out

The ledger above has two green rows, and they matter more to me than the red ones. The night before the last batch went live, a preview pass caught a component card claiming a rule had deleted zero real words, a figure nobody had ever recorded. It caught a phrase calling some pages the least read on a site when the audit only showed they were rarely read. Both were small. Both would have been quoted by somebody exactly as written. A number in a chart needs the same source as a number in a sentence, because a machine reads both the same way.

The rule I write against

So the test I use now is not whether a sentence is true today. It is whether I would be comfortable with a stranger repeating it in a year, in a different context, attached to my name, with nobody checking. Anything that fails gets a source, a date, a floor, or deleted. It is the same patience that an old cached reputation taught me from the other side: the past version of you keeps circulating long after you have moved on. Being wrong in public costs less than being wrong quietly, because only the public version can be corrected.

None of this makes me careful by temperament. It makes me careful by process, which is more reliable. Every entry on this blog now goes through the same gates before it ships, and the gates exist because of the rows in that table.

Tomorrow: two findings that look like they contradict each other, about listicles and explainers, and why they do not.