A growing share of people no longer search, they ask. They type a question into ChatGPT, Claude or Perplexity, or they read Google's AI overview instead of clicking a result. If your site is invisible to those systems, you are missing an audience that never sees a traditional results page at all. The discipline of getting quoted by AI answers has picked up two ugly acronyms, GEO and AEO, for generative and answer engine optimization. Strip the jargon and it is a concrete, mostly technical checklist. Here is what actually moves the needle, and what is snake oil.
What GEO and AEO really are
Classic SEO tries to rank a page so a human clicks it. GEO and AEO try to get your content read, trusted and quoted by a language model composing an answer. The model is not admiring your hover animations. It is reading text, structure and metadata, deciding whether your page is a clean, citable source, and sometimes fetching it live. So the work splits into two questions: can these systems reach your content, and once they have it, is it easy to extract and trust? Almost everything useful follows from those two.
Let the AI crawlers in
The first mistake is blocking the very bots you want to quote you, usually by accident through an overzealous robots.txt or a security layer. Every site I ship explicitly welcomes the AI crawlers: GPTBot for ChatGPT, ClaudeBot for Claude, PerplexityBot for Perplexity, and GoogleOther for Google's AI systems. It is a few lines in robots.txt, and getting it wrong makes everything else pointless. You cannot be cited by a system that was turned away at the door.
Give them a file written for them: llms.txt
Beyond letting crawlers in, I ship an llms.txt on every site. It is a plain-text file, a bit like robots.txt, that describes the business in clean prose an AI can lift directly: who this is, what they do, the services, the key products, how to get in touch. On sites with a blog I extend it with a list of posts, each a title, link and one-line summary, because those summaries get quoted almost verbatim. It is a small file with an outsized effect, since you are handing the model a pre-digested, accurate description instead of hoping it infers one from your markup.
Structure the page so it can be extracted
Language models reward the same clean structure that has always been good practice, which is convenient. Real semantic HTML, one clear heading per page, a logical heading hierarchy, actual paragraphs and lists rather than text baked into images. On top of that, structured data in JSON-LD: an Organization or Person block so the model knows who is behind the site, Article or BlogPosting on posts, and FAQPage where you answer real questions, because a well-marked question and answer is the easiest thing in the world for an answer engine to quote. My CMS, Orbit, injects this automatically. Dated posts get article metadata and BlogPosting structured data with no manual work, and pages can carry hand-authored FAQ and Organization blocks. The point is that the page should tell a machine what it is, not make the machine guess.
An answer engine cannot quote what it cannot reach, and will not quote what it cannot cleanly extract. Almost all of GEO is those two sentences.
Be fast, and be a real source
Speed matters more than it looks, because some of these systems fetch pages live when they answer, and a slow or flaky page gets skipped. The static, edge-served sites I build load in milliseconds, which removes that whole failure mode. Just as important, and far harder to fake: be an actual source. Answer engines favour specific, factual, first-hand content over generic filler, because filler is exactly what they can already generate themselves. Concrete numbers, real examples, clear claims. The same qualities that make a page worth a human's time make it worth quoting, which is a pleasant kind of justice.
You can also watch this working instead of guessing. The analytics I put on every build, TrackTrendy, my own platform, shows referral sources, and visits arriving from AI assistants are a distinct, growing slice. Those visitors show up with unusual intent, because something already answered their question and pointed at you as the source. They skipped the comparison shopping stage entirely.
What does not work
Let me save you some money. Keyword stuffing does nothing for AI answers and reads as spam to the models trained to spot it. There is no secret tag that forces a citation. Paying for "AI SEO" that is just old keyword tricks in a new suit is money set on fire. And thin, machine-spun content aimed at gaming AI systems is self-defeating, since it is indistinguishable from what the model produces for free and gives it no reason to prefer you. The honest version of this work is unglamorous: accessible, well-structured, genuinely useful, fast pages that tell machines clearly what they are.
The short version: welcome the crawlers in robots.txt. Ship an llms.txt that describes you plainly. Use semantic HTML and JSON-LD so pages are extractable and trustworthy. Be fast, ideally static. Write real, specific content worth quoting. That is the whole discipline, and I build every one of those into a site by default rather than selling it back to you as a mysterious extra.
Want your site quoted by AI answers instead of skipped? Get in touch, or see how I build for search.