← Writing

How to tell if a website was AI-generated

Seven patterns that show up when a site is built from a model default, each with the rate we measured across the sites SlopCheckr has scanned.

31 August 2026

A site assembled from a generation tool's defaults tends to carry the same handful of visual habits. They are easy to see once you know the list, and easy to measure once you can render the page and read its computed styles.

Seven of them are below, each with the rate it appears at across the sites SlopCheckr has scanned. Every one of these is a defensible choice in isolation. A rounded button is fine. What makes the list worth walking is how often the seven arrive together, unchosen, because the tool that built the page reached for the same defaults that every other tool reached for.

The seven tells

1. A fully rounded call to action

The primary button resolves to a border radius at or above half its own height, so the ends are semicircles.

This one is the most common single tell in the corpus, and the cheapest to check: open the page, inspect the main button, and read its computed border-radius. A value like 9999px is the giveaway, because it is the idiom a framework reaches for when the intent is "as round as possible" rather than a considered radius.

2. The three-card feature row

A row of three sibling cards at equal width, usually icon over heading over two lines of body copy.

Three is what fits a twelve-column grid without arithmetic. The pattern is worth noticing when the content does not actually come in threes: pages that have two strong points and one filler card, or five that were cut down to fit the row, are showing you the layout choosing the argument.

3. Inter as the body face

The computed font family of the page's dominant text run resolves to Inter.

Inter is a good typeface, which is exactly why it is everywhere. It is the default in several design tools and the first suggestion of several generation tools, so choosing it deliberately and arriving at it by inertia produce identical CSS. The typeface alone proves nothing. It is informative alongside the rest of the list.

4. A radial glow behind the hero

A radial gradient layer sitting on or behind the first viewport, usually a soft purple or blue bloom under the headline.

The effect had a specific origin in a handful of well-known developer-tool sites around 2021, and it propagated from there into templates, then into training data, then back out into new pages that have no relationship to the sites it came from.

5. A glassmorphism panel

A translucent surface with a backdrop-filter blur behind it.

Worth checking for contrast as well as originality: text on a blurred translucent panel changes legibility depending on what scrolls underneath it, and the resulting ratio frequently fails WCAG AA even when the design was checked once against a static background.

6. The sparkle icon

The sparkle glyph used as the visual shorthand for "AI".

This is the youngest convention on the list and the fastest spreading. It arrived with a wave of assistant features and became the default affordance almost immediately.

7. The whole set arriving together

Any single pattern above is weak evidence. The score exists because the combination is stronger evidence than the sum of its parts: a page with a pill button is unremarkable, while a page with a pill button, a radial glow, a glassmorphism panel and a three-card row has almost certainly been assembled rather than designed.

The median is low, which is the honest result. Most of the web is ordinary rather than generic, and a detector that flagged everything would be useless.

What a high score does not mean

It does not mean a model wrote the page. These are patterns, and a human designer working quickly from the same references produces the same output. The scan measures what the page looks like, and makes no claim about who or what produced it.

It also does not mean the page is bad. A conversion-optimised template with a pill button and three feature cards can work perfectly well. The cost of the patterns is specific: a visitor comparing four vendors in one session cannot tell the four apart, and the page does no work to be remembered.

Checking a page yourself

Without any tooling, open the page and answer four questions. What is the primary button's border-radius? What is the computed font-family on the body text? Is there a gradient layer behind the hero that carries no information? Does the first feature section contain exactly three equal cards?

Four yeses is worth a closer look. The full rubric, including the five scored dimensions and the detectors behind each one, is published on the methodology page, and the write-up of what happened when this rubric was pointed at our own marketing site is in the self-audit.

Sources

• Fully rounded pill call-to-action — 59%, 55 of 93 sites scanned, last updated 31 Aug 2026
• Symmetrical three-column card grid — 44%, 41 of 93 sites scanned, last updated 31 Aug 2026
• Inter as the primary face — 43%, 40 of 93 sites scanned, last updated 31 Aug 2026
• Radial gradient glow behind the hero — 40%, 37 of 93 sites scanned, last updated 31 Aug 2026
• Glassmorphism panel — 32%, 30 of 93 sites scanned, last updated 31 Aug 2026
• Sparkle icon in the interface — 23%, 21 of 93 sites scanned, last updated 31 Aug 2026
• Median slop score across the corpus — 15/100, 93 of 93 sites scanned, last updated 31 Aug 2026

Every figure above is computed from the 93 sites SlopCheckr has scanned, by scripts/compute-corpus-stats.mjs. Figures the corpus cannot yet support are listed, with the reason, on the methodology page.