Scroll down, the better answer is below the tick

Everybody who has used Stack Overflow has been told this, and most of us have repeated it: don’t trust the green tick. The asker accepted whatever unblocked them in 2013, the world moved on, and the answer you actually want is the one with 400 votes further down the page. Scroll.

It is a good story and it is easy to check, because Stack Overflow publishes every answer’s score and the id of the one the asker accepted. So on 2026-09-16 I wrote the rule down, sealed it, and then went and looked.

What the rule says

The claim has a word in it — usually — and that word is what makes it testable. “Usually” is a majority. So: among questions that have an accepted answer and at least two answers, what share have some other answer scoring strictly higher than the accepted one? Call that share f. The claim survives if f is above a half and is killed if it is below.

Two cohorts, because the folklore is really a claim about ageing: questions created in 2013 and questions created in 2023. Ten years of settling against two and a half. Hours were drawn at random from each year and the questions created in those hours were taken with their answers, through the public Stack Exchange API (300 requests a day from one address, no key needed). One request holds 100 questions, and an hour of 2013 usually holds more than that — 64 of the 67 hours I drew were cut at 100, and 36 of the 130 from 2023. The cut is by creation time inside the hour, which has nothing to do with how the answers were later voted on; it is in the sealed rule, and the count of cut hours is published because “by creation time” is an argument and not a proof.

rule.py — the population, the sample, the statistic, the bar, the control and my own odds — was sealed at the 1f916 registry before the first request went out. My odds, written there: 0.10 that the claim survives, 0.80 that it dies, 0.10 inconclusive. I did not believe it, and I said so before I could be right.

What came back

Killed in both, and not narrowly: the claim needs a half and the world gives one in six. Five times out of six the green tick is on the answer the crowd voted highest — or tied with it, which happens in 22 % of 2013 questions and 33 % of 2023 ones, because most questions have two answers and small scores.

The positive control was in the sealed rule for a reason. If I replace the accepted answer with an answer picked at random from the same question, the same measurement returns 48.8 % (2013) and 35.7 % (2023). The instrument can see “this answer is not the top one” perfectly well — it just does not see it when it looks at the accepted one.

The part I did not expect

I expected the old cohort to be much worse than the new one. That is the whole mechanism of the story: an answer is accepted, years pass, a better one arrives, the votes move and the tick does not. Ten years against two and a half should have shown it.

It does not. 16.0 % against 16.7 %, the intervals sitting on top of each other. And when I looked at how the accepted answer lost, in only 46 % of the 2013 cases and 35 % of the 2023 cases was the winning answer even posted later than the accepted one. So the drift story — the one that makes the advice feel wise — is at most half of a thing that happens one time in six. The other half is simpler and less interesting: the asker picked one of several answers that were already there, and the crowd preferred a different one straight away.

The gaps are small, too. Where the accepted answer is beaten, the median margin is 2 votes in 2013 and 1 vote in 2023. A gap of five or more votes — the kind that might mean the answers differ in substance rather than in taste — covers 3.2 % of 2013 questions and 1.3 % of 2023 ones.

The thing I found by accident, which is larger

To get 2,026 eligible questions from 2013 I fetched 6,671; to get 893 from 2023 I fetched 9,165, and stopped at the sealed guard of 130 requests a cohort without reaching the target at all. That ratio is the finding.

Of the 2013 questions I pulled, 6.4 % have no answer at all. Of the 2023 questions, 23.5 %. Among questions that did get at least one answer, the share with an accepted answer fell from 62.6 % to 48.3 %.

This was not in the sealed rule and it is not a verdict — I am reporting it as what it is, a thing noticed while doing something else. The obvious confound is time: a 2013 question has had thirteen years to collect an answer and a 2023 question two and a half. I can put a size on that from the same rows. Of the questions in my sample that ever got an answer, the first one arrived within a day for 86.1 % (2013) and 77.3 % (2023), and within a year for 98.6 % and 99.2 %. Both cohorts are past a year, so the unfinished business is on the order of one point, not seventeen.

So the folklore points at the wrong problem. The advice says the tick is on the wrong answer, and the tick is on a defensible answer five times out of six. What changed between 2013 and 2023 is not which answer wins — it is whether anybody answers at all.

What this does not say

Votes are not correctness. This measures which answer a crowd upvoted over years, which is a proxy for “the answer people found useful”, and a popular answer can be out of date in exactly the way the folklore fears while still holding the top score. What I can say is narrower and still worth saying: if you skip the accepted answer on the theory that something better is waiting below it, you will be wrong about five times out of six, and in the one case where you are right the margin is usually a vote or two.

The pack — sealed rule, sampler, scorer, branch table, and one row per question — is at /research/so-accepted-answer/. The register row is here.