AI Website UX Audit
Fetches a page, reduces it to the outline a visitor meets, and has a model name up to six things they would struggle with. No score, no design opinions.
This tool needs JavaScript: the page is fetched here and read by a model as an outline.
One page. It is fetched here and reduced to its title, headings, buttons and opening text before anything is sent to a model.
This is one of the few tools here that asks the server rather than your browser, because a browser cannot do it. Nothing you put in is stored; the request is rate limited by address and the answer is not cached.
Fetches a page, reduces it to the outline a visitor actually meets — the title, the headings, the buttons, the opening words — and has a model name up to six things somebody would struggle with.
The reduction matters as much as the model. Sending raw HTML would cost ten times as much and tell it less; what a reader meets is the structure, and that is what gets sent.
How to use
Put in one page. The home page is a fine start and an unrepresentative one to stop at: the page that needs the work is usually a product or a signup.
Example
Page https://example.com/pricing
Read as 23 outline lines: title, headings, buttons and the opening text
- Three plans are named but the headings never say who each is for, so the visitor has to infer it
- The primary button says "Get started" in four places with no indication of what happens next
- The FAQ answers objections the page has not raised yet, which suggests the order is wrong
Strongest: the opening sentence states what the product does in plain words
What this cannot see
the design no layout, no colour, no spacing
anything after JavaScript the HTML as delivered
what your visitors do the only real UX evidence is watching five of them use itPitfalls
This reads structure, not design. Spacing, hierarchy, contrast, the thing your eye goes to first — none of it is visible here. A page can read well as an outline and be a mess on screen.
No score. A number out of a hundred for user experience is a marketing device. Six specific findings are worth arguing with; 72/100 is not.
Nothing runs. The HTML is read as delivered. A page that builds its content in the client gives this very little to read, which is itself worth knowing: it is what a crawler sees too.
Five real users beat any audit. Watching five people try to do the thing your page is for finds more, and finds it in the order that matters. This is a cheap first pass, not a substitute.
The model is opinionated and sometimes wrong. It is told to say “I cannot tell from the outline” rather than guess about anything visual, and it will still occasionally assert something confidently. You are the reviewer.
It costs money. Each audit is a page fetch and a model call, capped daily and monthly, with the cost printed at the bottom.
Compatibility
Two steps: the page is fetched by the server, through the same guard as every other server tool here — http and https only, private and reserved addresses refused, redirects followed by hand with the checks repeated — and then reduced to an outline: title, meta description, every h1 to h3, the first twenty-five button or call-to-action labels, and the opening six hundred characters of text.
That outline is what goes to the model, with instructions to give at most six findings, each naming what a visitor would struggle with and why, and one line naming what the page does best. No score, no preamble.
Cost is counted from the tokens the provider reports, against a monthly ceiling the site owner sets, and the key never leaves the server. Nothing about the page or the findings is stored.