Your next customer might never see your homepage. They ask ChatGPT "who builds SaaS MVPs for non-technical founders," read the three names it hands back, and email one of them. No click. Just a mention, or no mention.
That shift has a name. Answer Engine Optimization (AEO) is the practice of getting your SaaS MVP cited inside AI-generated answers — ChatGPT, Perplexity, Google's AI Overviews, Claude — instead of fighting for a blue link nobody clicks anymore. It overlaps with traditional SEO, but it optimizes for a different reader: a model that quotes one sentence, not a human who skims a page. I rebuilt my own portfolio around this over a weekend, and the changes were smaller and stranger than I expected.
What is answer engine optimization, really?
AEO is making your content easy for an AI to extract, trust, and quote with attribution.
Traditional SEO gets you ranked. AEO gets you cited. That distinction drives everything else, because the mechanics differ. A model doesn't index your page the way Googlebot does. It pulls a passage, checks whether it answers the question cleanly, and decides whether to name you as the source. You can rank on page two and still get cited, if your page-two answer is the most quotable one available. You can rank first and get skipped, if your answer is buried under three paragraphs of throat-clearing.
So the unit of optimization changes. It isn't the page anymore; it's the passage. Every claim that matters should stand on its own, out of context, in 40 to 60 words, because that is the shape a model lifts into an answer.
Your site now has two readers
Most founders miss this part. A growing slice of your traffic isn't human.
Webflow has reported that AI agents already make up around 20% of traffic on some sites, growing roughly 350% year over year. Those aren't people browsing your pricing page. They're assistants fetching your content to answer someone else's question, then leaving. You will never watch them bounce. They don't bounce. They read, summarize, move on, and the human decision happens somewhere you can't instrument.
What does an agent want that a human doesn't? Clean text. No carousel, no cookie wall standing between it and your value proposition. The sites winning AEO serve structured, dense content a model can parse in one fetch. Which points straight at four small files most sites skip.
The four files that make a SaaS legible to AI
You don't need a replatform. You need four boring files.
llms.txt is a plain-Markdown index of your important pages — a sitemap written for a model instead of a crawler. Who you are, what you do, links to what matters.
llms-full.txt is the one with teeth. It inlines the full text of your key pages, every case study and post, into a single document, so an assistant answers a question about you in one request instead of crawling ten URLs and giving up halfway. I generate mine from the same data that builds the site, so it never drifts out of date. If you do one thing on this list, do this one.
AGENTS.md tells an agent how to act: how to evaluate your service, where to verify live work, how to request a quote. It's the gap between a model guessing and a model knowing.
pricing.md gives structured pricing a model can quote without inventing a number. For a B2B SaaS where "how much does this cost" is the opening question, a machine-readable answer beats a "contact us" wall.
All four run live on this site. llms-full.txt is the one worth a view-source.
The robots.txt mistake that hides you from ChatGPT
Allowing GPTBot does not put you in ChatGPT search.
This cost me an afternoon of confusion. GPTBot is OpenAI's training crawler. The bot that fetches pages to cite in ChatGPT's search answers is a different user agent: OAI-SearchBot. Allow one and block the other and you've let a model train on you while staying invisible in the product. Per OpenAI's own crawler docs, they are distinct, with distinct jobs.
The same split runs across providers. Anthropic separates ClaudeBot from Claude-User. Perplexity runs PerplexityBot and Perplexity-User. Google documents Google-Extended as its AI control. A robots.txt written in 2023 almost certainly allows the training crawlers and silently ignores the newer search and user-triggered fetchers — the exact bots that produce citations. Open the file. Add the citation bots by name. Ten lines, and it's the cheapest AEO win you have.
GEO: write to be quoted, not to be read
Generative Engine Optimization is the content half. Two rules carry most of the weight.
Write dense, not long. Assistants strip filler before they quote, so a 900-word post where every sentence earns its place beats a 2,000-word post padded to a target. Lead each section with the answer; put the qualifier after.
Name things accurately. Models are semantic engines. Call your feature a "smart workspace hub" instead of a "shared inbox" and a model fielding "who makes a shared inbox for clinics" can't match you. Vague naming reads as brand personality to a human and as noise to a machine.
Then test it. Open ChatGPT, Claude, and Perplexity and ask the five questions your buyers actually ask. "Best developer for a Flutter MVP in Europe." "Alternatives to a dev agency for a SaaS build." Are you cited? Are the facts right? That query log is your AEO rank tracker. Actually — run it before you change anything, so you have a baseline to compare against.
Where this goes wrong
The tempting mistake is serving one thing to humans and another to crawlers.
Don't. Showing Googlebot different content than your users is cloaking, and Google has penalized it for two decades. The safe form of "Markdown for agents, HTML for humans" is additive: .md versions, llms.txt, structured data — extra surfaces anyone can fetch, never a swap on your canonical page. The human page and the machine file tell the same story.
And don't overbuild. AEO sits on top of being genuinely useful and genuinely findable. If your traditional SEO is broken and your content is thin, no llms.txt rescues you. These files amplify good content; they don't replace it. A model still has to find a real answer worth quoting. Your job is to make sure that answer is yours, cleanly stated, on a page it's allowed to read.
Next steps
- View the live files: llms-full.txt, and the related SaaS authentication guide.
- See how I build production SaaS: BookBed and Callidus.
- Planning a build and want a number? Estimate your project.
What do ChatGPT and Perplexity say about your product right now? Go ask them the question your best customer would. Whatever comes back, or doesn't, that's your starting line.