How the AI works
Software could count your words in 2005. Understanding that your headline promises nothing took until language models could read. Here is exactly what happens between your paste and your roast, with no mystique.
01
the page like a conversion expert
Your HTML is fetched (capped at 600 KB, public addresses only) or taken from your paste. Scripts, styles, and markup noise are stripped. What is left becomes a linear outline in visitor order: H1, buttons, links, inputs, images, and every paragraph with its length. Order is kept on purpose. "The first button appears after 900 words" is only a finding if you know where things sit.
IN: raw HTML / OUT: ordered outline + counts
02
against five jobs a landing page has
A language model judges the outline on promise, call to action, hierarchy, proof, and friction. It is instructed to use the whole 0 to 100 range, to never flatter, and to quote the page's real words when it criticizes them. It is not allowed to invent content that is not in the outline.
IN: outline / OUT: conversion_score, verdict
03
by cost and effort
The model returns four to six kills. Each has what is wrong, why it costs you, and the fix, plus a severity and an effort estimate. They come back ranked, so the first item is the best trade you can make today. No kill is allowed without its fix. That rule is in the prompt and in the schema.
OUT: kills[] { what, why_it_costs_you, fix, severity, effort }
04
headline, hero, and CTA
The model writes replacement copy for your product using facts from your page. No brackets, no placeholders, no made-up numbers. The current copy is returned beside it so you can see the before and after. Paid plans can request three alternates per slot, each taking a different angle for A/B tests.
OUT: rewrites { headline, hero, cta } each { before, after }
Stated plainly: Bluntpage is model-agnostic. Everything above runs through one thin routing layer. The product asks for "a validated roast of this outline" and the layer decides which model answers, with a fallback chain across models and providers.
Today that means current frontier models from major labs. When a better or cheaper model ships, we change one list and re-run our evaluation set of scored pages. Your roasts get better and you do nothing.
Your pages go to the model provider through their API for the single purpose of producing your roast. API traffic is not used to train their models, and it is never used to train ours, because we do not train any.
Guardrails
Every model response is validated with zod against a strict contract. Wrong shape, missing fix, score out of range: rejected and retried once with the errors spelled out. You never see half a roast.
The voice contract bans comments about the founder, the team, or anyone's intelligence. The subject of every criticism is a headline, a button, or a paragraph. A post-processing pass scrubs anything that slips.
If every model is unreachable, a rules-based pass still checks CTA position, copy length, form fields, and proof signals. It is labeled as such, and it does not use up a free roast.
Identical pages hit a 24 hour cache. The public demo is limited per IP. Your account limits are enforced on the server, not in the browser.
Honest limits
Two free roasts a month. No card. The page gets hurt, not you.