Research note 01 · Second edition · September 2026 · Detector v1.1.1
The Same Page
We opened 4,725 Show HN launch pages as the Internet Archive saved them around their launch, every year from 2016 to 2026, and ran the Slop Patterns detector on every one. We only counted rules that held up when we checked them by hand.
By Pavithra Lamahewa, Principal UX Architect, Precious StudioData and method below
of 2026 launch pages carry at least one of six AI-era design tells. In 2016 it was 2 pages in 310. In 2026 it is 199 in 425.
have at least one basic defect, down from 72% in 2016. The pages got more correct.
Faint bars show 95% intervals. Hover a point for the count.
Launch pages didn't get worse. They picked up the same few looks.
The pages got more correct. Heading outlines are in order more often, the pages have landmarks, and a shared link shows a preview. The basic defects we can measure fell from 72% of pages in 2016 to 37% in 2026, most of it after 2020.
At the same time, a narrow set of looks spread: three matching feature cards, a tiny uppercase label over section headings, a pulsing dot, a gradient headline. In 2016 almost no launch page had any of them. By 2026 nearly half had at least one, and one in five had two or more. Until 2020, no page in our sample had two.
The rise started well before ChatGPT: under 1% in 2016, 4.5% in 2019, 12% in 2022. Our best guess is that template kits and shared component libraries were already pushing pages together. Then it sped up, to 22% in 2024, 27.5% in 2025 and 47% in 2026.
None of this shows who, or what, made any single page. Only about 2% of the 2026 pages carry visible markup from an AI site builder, so the look has spread well past the builders' own hosting. The timing lines up with AI site builders and coding assistants going mainstream; the data can't say more than that.
The six tells, year by year
Each one is a documented Slop Patterns entry with its sources. Two of them arrived almost overnight. The small uppercase label over section headings and the pulsing dot sat at or below 1% of pages every year until 2025, then showed up on about one page in six in 2026. The gradient headline went the other way: it peaked in 2024 and has faded a little since.
What got better
These defects are about structure, not style, and all three fell, mostly after 2019. Our best guess at why: current frameworks and starter templates include landmarks, preview metadata and a sensible heading order by default. Our library lists these as common in generated code, and they still are; on launch pages they became rarer.
Inter caught up with the system font
Until 2018, not one page in our sample set its body in Inter. In 2026, 23% did (87 of 372), and another 7% used Geist. That puts Inter level with the system font stack (25%), which had been the most common choice since 2019.
This is not a collapse. Every year still shows between 65 and 97 distinct body typefaces. What changed is which face is most common.
Does it cost upvotes?
Not that we can see. In 2026, launch pages with at least one tell had a median of points on Hacker News and pages with none had . The rank correlation between the number of tells and points was . Readers are not visibly rewarding or punishing these looks. We will measure it again next year.
We checked our own detector first
Before counting anything, we pulled random hits for 29 rules and judged each one by hand against the saved page and its screenshot. On whole live pages, most of our rules were wrong more often than we would accept. They were written for snippets of generated code. On real sites they tripped on framework CSS the page never used, on blog quotes, and on text hidden in scripts.
We fixed what we could and checked again. Nine rules reached 75% or better, and only those are counted in this note. We tuned them on the pages from this note's first edition (861 pages from 2019, 2022, 2024 and 2026), then checked the nine once more on a fresh set: 2026 launch pages that had never been used to tune anything. This edition runs the same nine rules, unchanged, on a sample five times larger. The rest of the rules are marked “worth a look” in the tool until they pass.
| Rule | First check | After fixes | Fresh pages | Status | Most common miss |
|---|
How this was done
Sample
700 random Show HN posts per year, January to August of every year from 2016 to 2026, drawn from about 31,000 posts that had 3+ points and linked to a website, one page per hostname. Links to GitHub, app stores, video, social and blog platforms, documents and package registries were excluded. Posts were collected in three-day windows, because the search API returns at most 1,000 results per query.
Capture
Each page was loaded from the Internet Archive at the snapshot nearest two weeks after launch, kept if it fell between 7 days before and 90 days after (median: 6 days after). It was rendered in headless Chromium and stripped of the archive's toolbar. 4,725 of 7,700 were usable (310 to 492 per year). The rest had no snapshot in that window, failed to replay, showed under 200 characters of text, or loaded no stylesheet.
Detection
Slop Patterns detector v1.1.1, from the rendered HTML and every stylesheet the page loaded. CSS rules whose selectors don't appear in the page's HTML are set aside (a static check, so it can miss some).
Counting
Only the nine rules that passed the hand audit. AI-era tells: A3, A10, A15, A22, A23, A36. Basic defects: A47, A55, B115. Intervals are 95% Wilson intervals; the 2016–2019 vs 2026 difference in tells per page was tested by permutation.
Limits
Show HN leans toward developers, so this is not the whole web. Archive replays can drop late-loading content. A static reading can't see hover states or anything drawn only by script. The rules were tuned on the first edition's pages, which overlap this sample, and the audit checked false hits, not missed ones. 2026 covers January to August only. None of this says a person used AI on any given page.
Not included
We don't name any site. The rules, their evidence and every entry's sources are public at sloppatterns.com.
Is your page on the same page?
Run the same detector on your site in about ninety seconds, or have Pavithra review it by hand.