The four checks that decide whether AI can find you.
Buyers increasingly ask ChatGPT, Claude, and Perplexity before they ask a search engine, and those assistants answer from whatever they can read about you. Answer-engine optimization (AEO) starts with a question search never had to ask: can the machine read you at all? Four deterministic checks answer it. None of them need a consultant to run.
Why this is a GTM problem, not a web problem
Search results were a list. The buyer read ten links and formed their own view. AI answers are a paragraph, and the assistant forms the view for them. When a VP of Sales asks "who helps mid-market SaaS companies fix pipeline velocity", the assistant composes an answer from the companies it can describe confidently. If it cannot describe you, it does not hedge. It leaves you out, or it guesses, and the buyer hears the guess.
That makes legibility a positioning decision before it is a technical one. The same ICP, category, and proof choices that shape your homepage decide what the machines should say about you. The four checks below are how you find out whether they can say it.
Check 1: is there an llms.txt at your root?
/llms.txt is a plain markdown file that hands AI systems a curated map of your site: what you are, which pages matter, what each one covers. Think of robots.txt, but for comprehension instead of permission. It is the single fastest fix in AEO because it is one file, and the crawlers behind the major assistants read it when it exists.
A good one has five properties: a top-level title, a one-paragraph summary as a blockquote, sections that group pages by purpose, links that stay on your own domain, and enough substance to be useful without bloating past a few kilobytes. If you already have one, grade it. If you do not, generate a starter and edit the descriptions until they sound like you.
Check 2: does robots.txt let the AI crawlers in?
This is the check that quietly ruins everything else. Many companies added blanket blocks for GPTBot, ClaudeBot, PerplexityBot, CCBot, and friends during the 2023 anxiety about training data, and never revisited the decision. Blocking a crawler stops training on your content. It also stops the assistant from reading your site when a buyer asks about you today. You cannot be cited by a system you have told to go away.
Open your robots.txt and look for any User-agent group naming an AI crawler with Disallow: /. Decide deliberately. For most B2B companies, being findable when a buyer asks is worth more than the training-data principle, and the two can be separated: OpenAI, for example, publishes distinct user agents for training and for search.
Check 3: is there structured data on the pages that matter?
JSON-LD is the part of a page written for machines: an Organization block that says who you are and where else you exist, a Service block per offering, FAQPage where you answer questions, Article on anything you publish. Models and the crawlers that feed them lean on this because it removes guesswork. A homepage with an Organization block and sameAs links to your LinkedIn and other profiles gives an assistant a verified identity to attach every other fact to.
Scan your homepage, your service pages, and your pricing or FAQ page. Each should carry at least one relevant block, and the Organization block should appear once, with a square logo, a plain description, and the same name you use everywhere else. Inconsistent names across profiles are how a company ends up described as three different entities.
Check 4: sitemap and per-page hygiene
The unglamorous layer. A sitemap.xml with accurate lastmod dates tells crawlers what changed and what to re-read. A meta description on every page is the sentence a model falls back to when it has to summarize a page in one line. An h1 that states what the page is about, and Open Graph tags so the page unfurls correctly when shared, close the loop.
None of this is new. What is new is that machines now read it to decide what to say about you out loud, to a buyer, in a sentence you will never see.
The test that matters more than the checks
Take the visible text of your homepage, paste it into an assistant with no other context, and ask: "Based only on this, what does this company do and who is it for?" Read the answer as if you were the buyer. If the model hedges on who you serve, that is not a model problem. That is the homepage failing to say it. We run this cold read as the last step of every audit, and it is routinely the most useful line in the report.
What to do this week
- Grade or generate your llms.txt with the free generator. Ten minutes, no signup.
- Read your robots.txt and decide, on purpose, which AI crawlers you block.
- Add an Organization block with sameAs links to every profile you control, and make the name identical across all of them.
- Run the cold read on your homepage and fix whichever sentence the model could not find.
One honest caveat
Nobody can guarantee placement inside an AI answer. The assistants weigh third-party corroboration heavily: mentions, profiles, and content elsewhere that agree with what your site says. Legibility is the prerequisite you control, and most of your competitors have not done it. The corroboration is the longer game, and it is the same game as building a reputation.
Where this fits
AEO lives inside our GTM Strategy work because it is a positioning problem first. The fixed-scope AEO Audit runs all four checks plus the cold read and hands you a prioritized fix list.