Research

How it works

The check looks at a page the way a visitor’s browser draws it, not at a picture of it, and it does not use an AI model to judge it.

1. It opens the address in a real browser (Chromium), on a desktop screen of 1440 × 900 or a phone screen of 390 × 844.
2. It loads the page twice, once with the system set to light and once to dark, because many pages change with that setting.
3. It waits for fonts and images, scrolls to the bottom so content that appears on scroll is there, hovers over cards to see which react, and watches the headline for text that types itself out.
4. It reads the finished page: the text, the fonts in use, colours, gradients, shadows, corner radii, layout and animations, as the browser computed them.
5. Each of the 32 scored tells is a written rule over those facts, for example “the main heading font is one of the faces AI tools reach for by default”. A rule either matches or it does not, and records what it saw.
6. For 9 more tells, a rule can find the evidence but not decide, for example whether a testimonial is real. Those are shown under “For you to check” and never change the score.

The score

The score is out of 100. Each scored tell found adds its weight; the total is divided by the weight of all 32. A tell that people have confirmed more often weighs more.

The level counts tells, not weight: five or more found is High, three or four is Medium, one or two is Low, and none is No slop. Of the light and dark runs, the result opens on the one that scored higher.

A high score means the page shares many habits with other AI-made pages. It is not a verdict on the product, and well-made pages by people trip some of these too.

How reliable each tell is

Some tells show a figure such as “~80% reliable”. It is the share of past flags for that tell that a person agreed with when checking the pages by hand.

So far 12 tells have that figure, from 177 sites labelled by hand between April and June 2026 in the original tool’s dataset. The other tells were added in October 2026 and say “Not checked by hand yet” until enough pages have been labelled. Until then they get an average weight, so a new tell cannot outweigh a checked one.

The examples

The pictures on the Patterns page are crops of real pages that tripped each tell. To find them, the current check was run over 582 recent Show HN launches in October 2026. Each crop is cut around the part of the page the rule pointed at.

Every crop was looked at before it went on the page. 122 were kept. Error pages, cookie walls and crops where the tell was not visible were dropped. A tell with no good example says “Example to follow” rather than showing a wrong one. Motion tells are shown as stills, so they show where the movement is, not the movement.

The same research run is in the Rank, marked “Research run”.

What we keep

• Each result is saved with its screenshots, so opening it again, from the gallery, the rank or a shared link, shows the saved result and the date it was made instead of scanning again. “Scan again” on a result replaces it.
• Listed scans appear in the Gallery (the latest) and the Rank (all of them). Untick “List in the gallery” before scanning to keep a scan out of both.
• Addresses with a query string, a fragment or a password in them are never listed, since they may be private links.
• A site that is down, not found, answers with a server error, is paused, or blocks the scan is not scored, saved or listed. The same goes for an empty page or one that shows only a host’s error message.
• The check refuses private network addresses and only opens public web pages.

Limits

• It checks one page, as it looks when it loads. Pages behind a sign-in, or that block automated browsers, may score low because the check never saw them.
• A rule cannot judge taste. A page can pass every check and still look generic, or trip several and still be well made.
• Some rules are crude. The grid backdrop rule, for one, currently flags canvases and gradients that are not grids.
• The forty-second tell, text that reads as generated, is not checked.

Powered by fixmyland.ing

Lookalike Check is one of a family of checks at fixmyland.ing, each looking at one thing a page or app can get wrong.

• LiveDesignLookalike Check, this tool
• To followSecurity
• To followPerformance
• To followMore checks

Credits

Lookalike Check is built on Design Slop Cop by Adrian Krebs, used under the MIT licence. Its scanner, its first set of patterns and its hand-labelled dataset are the base this tool extends.