Why AI-built landing pages all look the same
Convergence is not a coincidence. Half the sites we scanned resolve to one of six typefaces, and the reason is in how the tools were trained.
Open five newly launched developer-tool sites in five tabs and the tabs blur together. Same dark hero with a soft gradient behind it, same tight sans-serif headline, same row of three cards, same rounded button in the top right. The sameness is measurable, and it has a mechanism behind it.
The measurement
Across the sites SlopCheckr has scanned, half resolve to one of six typefaces for their dominant text run.
Inter, Poppins, Roboto, Montserrat, Geist or DM Sans.
47 of 93 sites scanned · measured 31 Aug 2026
Inter alone accounts for most of that share.
The computed font-family of the page’s dominant text run resolves to Inter.
40 of 93 sites scanned · measured 31 Aug 2026
Layout converges the same way. The three-card feature row appears on close to half of them, and the fully rounded primary button on more than half.
A feature row of three sibling cards at equal width.
41 of 93 sites scanned · measured 31 Aug 2026
The primary button resolves to a border-radius at or above half its height.
55 of 93 sites scanned · measured 31 Aug 2026
Six typefaces out of the thousands available, on half the web we can see. That is a much narrower distribution than taste alone would produce.
Four mechanisms, none of them mysterious
Defaults are sticky. Every framework ships with a font stack, a spacing scale and a button radius. Changing them requires a decision, and shipping requires only silence. The default wins every time nobody argues with it, which on a small team building fast is most of the time.
Training data is the web as it was. A model that generates a landing page has read a great many landing pages, and the ones it has read most are the ones that were most copied. The output regresses toward the mean of its input, and the mean of the input was already narrow because of the first mechanism.
The feedback loop closes. Generated pages ship, get indexed, and become training data for the next model. Each pass tightens the distribution. This is the part that makes the trend accelerate rather than plateau: the process is now partly feeding on its own output.
Nobody is punished for it. A generic page still converts. It loads, it explains the product, the button works. The cost shows up only in comparison, and comparison happens in the visitor's head across four tabs, where no analytics tool can see it.
Why it costs something anyway
The damage shows up in the position a visitor occupies at the moment of choosing, rather than in any single page.
Someone evaluating four vendors in one session is holding four near-identical impressions. When nothing distinguishes the pages, the visitor falls back on whatever else is available: price, whichever name they saw first, whichever site a colleague mentioned. A page that looks like every other page has quietly handed the decision to something outside the page.
For a product whose whole claim is judgement, taste or craft, the cost is sharper. The site is the first sample of the work. A generic site makes a claim about the product that the copy then has to argue against.
What actually helps
Changing the typeface is the smallest useful move and the one most often mistaken for the whole job. Swapping Inter for another sans-serif that fifty thousand other sites also use relocates the problem.
The changes that hold up are structural. Break the three-card row when the content does not come in threes — two cards, or five, or a list. Give the page one asymmetric layout that the reader remembers. Cut the gradient that carries no information. Write the headline about the specific thing the product does, in the words a customer would use, rather than in the register of a category.
None of that requires abandoning generation tools. It requires reading what they produce as a first draft rather than a finished page.
The rubric behind the figures above, including what each detector actually looks at, is published on the methodology page. The pattern-by-pattern checklist is in how to tell if a website was AI-generated.
Sources
- One of six default faces as the primary face — 51%, 47 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
- Symmetrical three-column card grid — 44%, 41 of 93 sites scanned, last updated 31 Aug 2026
- Fully rounded pill call-to-action — 59%, 55 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.