A growing share of the people evaluating your product never see your site. They ask an assistant, read three paragraphs of summary, and form an opinion about you from a description you didn't write. If the machine that wrote that description couldn't parse your site properly, it wrote something vaguer than the truth. Or it wrote about a competitor instead, because that competitor was easier to read.
This is the newest lens in the roast and the only one where the fixes are unambiguous. Twenty points, seven checks, all deterministic, with no model judgment and nothing to argue with. Either the file is there or it isn't. Here's what each check looks for and what to do when it fails.
1. AI-crawler access in robots.txt
We read your robots.txt and report GPTBot, ClaudeBot, PerplexityBot and Google-Extended as allowed, blocked, or unspecified. This is the check that most often produces a genuinely surprised founder. Blocking happens by accident constantly: a wildcard disallow copied from a template, a WordPress plugin's default, a security setting someone enabled in 2023 when the advice was different, or a CDN bot-protection rule nobody has revisited.
Deciding to block is legitimate, and publishers with licensing leverage make that trade deliberately. Discovering you blocked is not. The fix takes one line either way; what matters is that it becomes a decision instead of an accident.
2. llms.txt
A plain-text file at your root that tells a language model what your site is, what the important pages are, and how to talk about you. It's a young convention and support is uneven, which is exactly why it's cheap: an afternoon's work, no downside, and it's the one artifact where you get to write your own summary rather than hoping one gets inferred from your nav.
The roast checks presence and that it isn't empty boilerplate. Write it as if briefing a new salesperson who will never meet you: what you sell, who for, what you're not, and the five URLs that matter.
3. Structured data
JSON-LD, present and of the correct types. Organization and WebSite are the floor; Product, SoftwareApplication, FAQPage and Article earn their keep depending on what you are. Structured data is how a machine converts your page from prose into facts (name, category, price, parent company, support contact) without having to infer any of it.
4. Semantic HTML and heading structure
One h1, headings that descend in order, real landmarks (header, nav, main, article, footer), and content that survives being stripped of styling. Extractors follow structure. A page built from nested divs with visual-only hierarchy reads to a parser as one long undifferentiated block, and what gets quoted back is whatever fell nearest the top.
This check has a useful side effect: everything that makes a page legible to a model also makes it legible to a screen reader. The accessibility work and the AI-visibility work are largely the same work, which makes it an easy thing to get prioritised.
5. Meta and Open Graph completeness
Title, description, canonical, and the OG set that decides what your link looks like when someone pastes it into Slack, LinkedIn or a chat window. Truncated titles, a description auto-generated from your first paragraph, or no OG image at all: each one is a small tax paid every time your URL is shared, and shares are increasingly how a model discovers you exist.
6. Sitemap
Present, valid, and actually listing the pages you want read. The common failure isn't absence; it's a stale sitemap generated once, still advertising routes that 404 and missing everything shipped since. A crawler that hits three dead URLs in a row draws conclusions about how much of the rest is worth fetching.
7. Question-shaped content
Does anything on your site answer a question in the form a person would ask it? Assistants retrieve against questions. A page organised entirely around feature names and internal vocabulary matches almost nothing a real prospect types. A page that says "how much does it cost" and then answers it matches a great deal.
This is not a call for a content farm. One honest FAQ, written in customer words rather than product-marketing words, moves this check and reads better for humans too.
What the lens deliberately can't tell you
Every check above is a presence check. Together they establish that the engines are able to read you, understand you, and cite you accurately if they choose to. None of them establish that they do.
No schema, no llms.txt, and robots.txt blocks GPTBot. You've built a site AI physically cannot recommend.
That's why the lens is capped at 20 rather than scored out of 40 on a proxy. Finding out whether ChatGPT and Perplexity actually name you when someone asks for a tool like yours means running the queries, reading the answers, and recording who got mentioned instead. That's the Deep Roast's job, and one of the more uncomfortable slides in it.
Start with the free roast. The AI-visibility module comes back as a checklist, item by item, pass or fail, with the fix attached. Most sites can close three of the seven before the end of the week.
Seven checks, scored out of 20, in the same sixty-second roast.