Your homepage, scored out of 100, including whether AI can read it at all.

There is no shortage of free website graders. Paste a URL, wait for a progress bar that is mostly theatre, receive a page of adjectives. Your hierarchy could be stronger. Consider improving your call to action. Nothing in that report tells you what to change on Tuesday morning, and nothing in it can be compared to anyone else's report, because there was never a rubric. Just a model describing a screenshot back to you in a confident voice.

The free roast is built the other way around. One homepage, desktop and mobile, scored out of 100 against a fixed, versioned rubric. Same rules for every site. That constraint is the whole product: a score only means something if the thing that produced it didn't move between your site and your competitor's.

Four lenses, 100 points

The score splits across four lenses, and every finding is tagged with the lens it came from, so you can see whether you have a persuasion problem, a craft problem, a speed problem, or a machine-readability problem. They need different fixes and usually different people.

• Conversion, 35 points. Does the page persuade? Value-proposition clarity, CTA presence and hierarchy, social proof, friction, and whether the copy says anything a competitor couldn't paste onto their own hero.
• Craft, 35 points. Is this well made, or does it read as template and AI slop? Typographic tells, layout clichés, stock gradients, copy smells. Findings cite the specific pattern rather than gesturing at vibes.
• Page Speed, 10 points. Scored live from Google PageSpeed Insights, mobile strategy. The Lighthouse performance score drives the points; LCP, CLS, TBT, FCP and Speed Index become the checklist.
• AI Visibility, 20 points. Can AI engines read, understand and cite you? Entirely deterministic presence checks, no model judgment involved.

Deterministic rules do as much of the work as rules can do. The model is only used where a rule genuinely can't be written, like judging whether a sentence is specific or generic. When it is used, it has to quote the evidence it found on your page. A finding you can't trace back to something on the screen isn't a finding, it's an opinion with a confidence score.

The lens nobody else runs

AI visibility is the differentiated part, and it's differentiated for a boring reason: it's deterministic, so it's cheap to be right about, and almost nobody bothers. Seven checks: llms.txt, structured data, AI-crawler access, semantic HTML, meta and OG completeness, sitemap, question-shaped content. Individually small. Together they decide whether a language model can form a usable representation of what you sell.

The lens is capped at 20 points on purpose. Presence checks can tell you whether the engines are able to read you. They cannot tell you whether the engines actually cite you when someone asks for a tool like yours. That requires running the queries and reading the answers, which is the Deep Roast's job. We'd rather cap the lens honestly than score you out of 40 on a proxy.

The lens, check by check: what each one looks for and how to fix it →

What you get back

A score with a band, one headline burn line, and five to eight findings. Each finding carries its lens, its severity, the evidence, either a quote from your copy or a crop of your page, and the fix. Desktop and mobile screenshots sit alongside. The AI-visibility checklist gets its own module, item by item, pass or fail.

No account, no card, no email gate on the score or the share. The email field appears once, late, and only unlocks the full report. We can afford that because the free roast isn't the business. It's an honest sample of how we work, sixty seconds long, and about your site rather than ours.