The accepted answer on Stack Overflow is usually not the best one

The advice everybody repeats about Stack Overflow: ignore the green tick, the accepted answer is usually stale and the highest-voted answer is below it. Among questions that have an accepted answer and at least two answers, what share have some other answer with a strictly higher score? Measured on two cohorts, questions created in 2013 and in 2023, because the claim is really about ageing.

rule.py, sealed at the 1f916 registry as rule.so-accepted-answer at 2026-09-16 16:11:26 Rome, before the first API request (the 2013 rows were on disk by 16:12:40 and the 2023 rows by 16:14:29). Hours drawn uniformly at random from each year (seed 20260916); one Stack Exchange API request per hour for the questions created in it, sorted by creation, 100 at a time, answers inline. Eligible: an accepted answer, at least two answers, and an inline answer list as long as the question's own answer_count. f = share whose accepted answer scores strictly below the maximum (a tie counts as top). Wilson 95 %: survives if the lower bound is above 0.50, killed if the upper bound is below it, otherwise inconclusive. Positive control in the same file: the same statistic on one answer drawn at random per question must exceed f by at least 10 points in each cohort, or the result is void.

computed after the data — conditional on the world as it came out, so it cannot show a branch was unreachable by design

Killed in both cohorts, with the control passing in both. 2013: 6,671 questions fetched over 67 requests (64 of those hours held more than the 100 a request returns and were cut at 100, by creation time inside the hour; 36 of the 130 hours in 2023), 2,026 eligible, 324 beaten, f = 16.0 % (14.5–17.7); control 48.8 %, margin +32.8. 2023: 9,165 fetched over 130 requests, 893 eligible, 149 beaten, f = 16.7 % (14.4–19.3); control 35.7 %, margin +19.0. The 2023 cohort stopped at 893 rather than the sealed target of 2,000 at the rule's own stopping condition — 130 requests per cohort, written into the sealed file because the API allows 300 a day for both — and not because the quota itself ran out (97 requests were left). The bar is still decidable at that size (see branches). Registered and not scored: exact ties at the top 21.9 % / 32.5 %; median margin where the tick loses, 2 votes / 1 vote; the winning answer was posted after the accepted one in only 46.0 % / 34.9 % of those cases; f among questions with five or more answers 23.0 % (n=183) / 18.2 % (n=22). After the data, not sealed: 6.4 % of the 2013 questions fetched have no answer at all against 23.5 % of the 2023 ones, and among questions with at least one answer the share with an accepted answer fell from 62.6 % to 48.3 %; the time confound is bounded by the same rows, where 98.6 % / 99.2 % of ever-answered questions had their first answer within a year.

I tested a version of the claim its speakers never asserted.

What is sealed is a claim about votes, and what people say is a claim about quality: "the accepted answer is worse" is not testable from the API, "the accepted answer is not the top-voted one" is. A popular answer can be out of date and still hold the top score, and that case is invisible here.