Random article IDs put the share of dev.to posts with no reactions at 92%

dev.to's API will list articles by tag, and paging that listing looks like the supported way to enumerate a corpus, except that the listing is not ordered. Walking twelve pages of the `webdev` tag, page 10 ran from 2026-08-31 down to 2026-08-29 while page 11 ran from 2026-09-13 down to 2026-08-27 ^[8]. Page 11 therefore held posts newer than anything on page 10, the pages overlapped, and posts went missing between them ^[9]. The post being hunted for, published 31 August with 60 reactions, was on none of the twelve pages ^[7]. Forty minutes of paging had already gone into the corpus by then ^[21].

The gaps skew toward the quiet posts. Posts the listing returned averaged 0.40 reactions and the ones it skipped averaged 0.15, a factor of 2.7, and 19.0% of the included posts had at least one reaction against 5.1% of the excluded, a factor of 3.7 ^[10][1][2]. So the listing selects on the variable under study.

Those two subgroup rates have a second use. The 460-post random sample had 8% of posts with at least one reaction ^[3]. If that 8% is the weighted average of 19.0% among listed posts and 5.1% among unlisted ones, the listing covered about a fifth of them: 19.0p + 5.1(1 - p) = 8 solves to p = 0.209 ^[4]. The stated interval on the 92% is plus or minus about two and a half points, which stretches the same calculation to anywhere between roughly 3% and 39% ^[3][8]. So it is an order-of-magnitude figure. Read at random, the zero-reaction pile included "Delta E is a distance, not a score" and "How Farm.js compiles React components into direct DOM updates" ^[4].

The title-and-tag comparisons in the post come off the big corpus, 45,400 August posts, because the random sample is too small to slice ^[14]. Cover images go with 23.7% of posts getting at least one reaction against 11.2% without ^[15]. Four tags beat one, 16.0% against 3.5% ^[18]. First-person titles, ones with "I" or "my" in them, go with 24.7% against 13.9% ^[16]. For any of those gaps to transfer to the site as published, cover-image use would have to be independent of whether the listing returns a post, and inclusion in the listing already tracks reactions ^[10]. The post does not test that, and it says the corpus leans optimistic in absolute terms ^[14].

Some of the standard recommendations came out the other way. Titles beginning with "How" did worse than titles that do not, 12.2% against 15.1% ^[19]. "Sampling through the listing quietly hands you a rosier site than the real one," the author wrote ^[11].

The instrument error is documented in the post itself. An earlier draft written off the tag corpus put the zero-reaction share at 85%, and the random-ID figure is 92%, seven percentage points of optimism from the sampling method alone ^[12][7]. The author also checked whether a fortnight is long enough for a post to finish collecting reactions: the share with any reaction was 13.6% at 14 to 20 days old and 15.0% at 42 to 48 days, with no trend in between ^[13].

The evidence supports the level but not the effects. The median August post drawn by ID got zero reactions ^[9], and any percentage published off a tag-paged corpus is too high. Whether the cover-image, first-person and four-tag gaps hold on an unbiased sample is untested, because the post says its 460-post draw is too small to slice ^[14].