Bugs your users can see, with no error to log.

The GenClass catalog of visual bugs collects what breaks on screen in AI-built sites, from studies, builder forums, code searches and audits, and ranks each bug by how strong the evidence is that it is common. These are its families, in its order. Under each: the measured checks for that family, by the names their reports use. A family's other bugs have no check yet.

Unreadable text

Text the same colour as its background, too faint, too small, or set over an image.

Checks: text-invisible

Missing or broken media

Images, icons and charts that fail to load, never existed, or render stretched.

Checks: image-broken, image-css-broken, image-distorted, media-failed, icon-font-failed

Missing styles or fonts

CSS the page asks for and never gets: Tailwind classes that generated nothing, theme tokens that resolve to nothing, fonts and stylesheets that fail to load.

Checks: stylesheet-failed, font-failed, css-var-undefined, csp-blocked

Layout overflow and clipping

Content wider than its box: the page scrolls sideways on a phone, cards blow out, labels are cut off.

Checks: page-overflow-x

Overlap and occlusion

Something paints over what the visitor needs: a sticky header over a heading, a dialog behind the page.

Checks: focus-obscured, control-unreachable, dialog-covered

Hidden or unreachable controls

Controls that exist but cannot be used: an invisible layer takes the click, the page stays frozen after a modal closes, actions appear only on hover.

Checks: click-blocked, pointer-events-leak, scroll-lock-stuck

Template and data leaks

What should never reach a visitor: undefined, NaN, [object Object], Lorem ipsum, garbled characters.

Checks: text-template-leak, text-placeholder, text-encoding

Blank page and missing content

The page, or a part of it, renders nothing.

Checks: page-blank

Motion and shift

Content jumps while the page loads, or waits at opacity 0 for an animation that never runs.

Checks: layout-shift

Dark mode

A dark theme applied halfway: dark text left on a dark surface, light cards left behind.

Checks: text-invisible

Responsive

Bugs that appear only at some widths: a phone shown the desktop layout, rows that never wrap.

Checks: viewport-meta-missing, page-overflow-x

Alignment, spacing and consistency

Nothing is broken, but edges, gaps and buttons do not line up.

No check: design quality, not breakage

Other

Fields labelled only by placeholder text, emoji used as icons, and pages that render differently in Safari, Firefox and Chrome.

No measured check yet

Design and thresholds: the visual checks design. The checks never feed the network model, and they read no pixels.

Every number below is for GenClass visual checks v1 (branch wave2/visual), measured on GitHub runners in Chromium 153 on October 13, 2026, and copied with its label from the runtime results, sections 8 and 9. These are measurements of fixed checks with documented thresholds, not of a model. The page check runs the same checks, except click-blocked; these numbers were measured on the runtime's bench, not through the page check.

• 35 of 35 planted faults found on fixture pages, each on its element, 23 kinds, 24 tests passed. These are fixture pages written by the same author as the checks, so this shows the checks work as designed.
• 0 false reports and 0 fixes applied in 128 page views of unmodified pages (Table 2).
• 8 findings in 96 loads of public home pages, 5 confirmed real by a person from the element's crop, 3 unconfirmed, 0 check errors (Table 3, run 4).

Kind	Planted	Found	Recall
text-invisible	70	70	1.00
image-broken	36	36	1.00
image-css-broken	32	32	1.00
media-failed	36	36	1.00
stylesheet-failed	32	32	1.00
font-failed	28	28	1.00
icon-font-failed	36	36	1.00
csp-blocked	36	36	1.00
text-template-leak	54	54	1.00
text-placeholder	28	28	1.00
text-encoding	28	28	1.00
image-distorted	36	36	1.00
page-overflow-x	22	22	1.00
viewport-meta-missing	11	11	1.00
layout-shift	18	18	1.00
page-blank	18	18	1.00
click-blocked	18	12	0.67
pointer-events-leak	18	17	0.94
scroll-lock-stuck	36	24	0.67
control-unreachable	32	32	1.00
focus-obscured	18	18	1.00
dialog-covered	18	18	1.00
css-var-undefined	28	27	0.96

Every click-blocked and scroll-lock-stuck miss is on one of the six demo pages, which update their DOM all the time; both checks wait for a moment with no DOM change, so they stay silent on such pages. css-var-undefined missed one phone load. The pointer-events-leak miss is a compat app whose requests were in flight during the reads: since the review a lock is not judged while a request runs.

Pages	Page views	Findings	Confirmed real	False reports	Fixes applied
Demos, 7 pages	28	2	2	0	0
Templates, 4	16	0	0	0	0
Compat apps, 7 apps x 3 scenarios	84	0	0	0	0

The two findings are one real bug in two colour schemes: on the demos index at desktop width, the content under the hero headline moves 59 px 0.2 s after navigation start (a layout-shift window of 0.34). It is labelled real and the demo is left alone. The same zero false reports held in all five runs.

Run	Sites measured	Loads	Findings	Per load	Confirmed real	Check errors	Unconfirmed
1	32	100	49	0.49	5	44	0
2	33	100	12	0.12	5	5	2
3	34	102	7	0.069	5	1	1
4	31	96	8	0.083	5	0	3
5	33	97	6	0.062	not looked at	not looked at	6

The 5 confirmed real findings in run 4 are 3 bugs on 2 of the 40 sites, each found in two colour schemes where it applies: on a phone, a statistic 16 px wider than the 390 px viewport, so the page scrolls sideways; on a phone, a layout shift of 0.27 that moves the content below by 782 px; on a phone, a keyboard-focused link hidden under a fixed announcement banner. Bot challenges, HTTP errors and rate-limited loads are not measured.

Loads	Count	Pass time	Pass within budget	Longest slice
Local pages, 1x CPU	128	6.8 / 32.1 / 53.4	100%	6.2 / 10.5 / 12.5
Local pages, 4x CPU throttling	32	32.8 / 112.4 / 113.6	78.1%	14.3 / 18.3 / 20.1
Public sites, 1x CPU	117	40.3 / 95.2 / 158.4	95.7%	8.9 / 18.9 / 51.1

Three of the proposed cost gates are not met yet: p95 pass time on public sites is 95.2 ms against at most 50 ms, the longest slice at p95 is 18.9 ms on public sites against at most 12 ms, and the first pass finishes inside 100 ms at 4x throttling on 78.1% of local pages against at least 90%.

• The learned model has no numbers yet. Nothing on this page claims a detection rate for it until one is measured and published.
• Measured on GenClass's own apps and 31 to 34 of 40 public home pages per run. Recall is on faults planted in the author's own forms, and the public sites were also used to tune the checks between runs, so they are not held out.
• No pixels. Text over photos, video and gradients is skipped, and canvas and WebGL content is not checked.
• Only what the visitor's browser shows. Once they ship in the runtime, the checks run in your app in the colour scheme, width and hover state each visitor is using. The page check adds a phone view and dark mode.
• Two checks need Chromium. Layout shifts and failed CSS background images have no browser signal elsewhere.
• Not an accessibility audit. Readable text below WCAG AA contrast is not reported, and nothing here makes a site accessible or WCAG compliant.
• Design quality is not reported as a bug. Uneven alignment and spacing are in the catalog; no check reports them.
• It does not repair CSS in general. The three fixes above are the only changes it can make, and only when you turn them on.

Paste the address of a public page. A cloud browser loads it, runs the measured checks (all but click-blocked, which needs a visitor's clicks), and shows each problem on a screenshot with what was measured and a suggested fix.

Page address

Scan

A cloud browser loads that one page as a desktop and as a phone and reports each problem with its element and what was measured. No form or POST request reaches your server. How it works →