Answer Engine Optimization (AEO): how to get cited by ChatGPT, Claude and Perplexity

Answer engine optimization is search optimization for a world where the result is an answer, not a list of links. When someone asks ChatGPT, Claude, or Perplexity a question, the engine reads a handful of sources, writes one synthesized reply, and cites the pages it leaned on. AEO is the discipline of making your pages the ones it leans on — findable by the engine's crawler, legible to its model, and quotable in its answer.

The shift from classic SEO is bigger than it sounds. In SEO you compete for a position on a results page, and the user still clicks through and judges your site themselves. In an AI answer there is no page of ten results — there is one reply and a short list of citations. If you are cited, you are the authority the user hears. If you are not, a competitor's framing becomes the answer, and the user may never see that you exist.

You will also hear this called generative engine optimization (GEO). The two terms describe the same practice; GEO tends to emphasize how you appear inside the model's generated prose, AEO the mechanics of being retrieved and cited. This guide treats them as one. One thing AEO is not: a replacement for SEO. Answer engines retrieve from conventional search indexes, so fast pages, clean HTML, and crawlable architecture still underpin everything below.

Your content reaches an AI answer through two doors. The first is training data: what the model absorbed about your brand before it shipped. That moves slowly — on model release cycles — and rewards being consistently described across the web over years. The second is live retrieval, and it is where a small site can win this quarter. Perplexity searches the web for nearly every query; ChatGPT and Claude search when a question calls for fresh or specific information. In retrieval, the engine is choosing sources right now, at answer time.

The retrieval pipeline looks like this: the engine turns the user's question into search queries, pulls a shortlist of pages from an index, fetches and reads them, then synthesizes a reply and cites the passages it actually used. Two consequences follow. First, you have to be fetchable in real time — a blocked crawler or a JavaScript-walled page never makes the shortlist. Second, you win at the passage level, not the page level: a modest site with one clean paragraph that fully answers the question routinely beats a stronger domain that buries the answer under eight hundred words of preamble.

When the model reads its shortlist, it favors passages that answer the question completely without needing the rest of the page, plain declarative statements over marketing prose, and sources whose identity is unambiguous. An honest caveat: no engine publishes its selection recipe, so everything in this guide comes from provider documentation and observable behavior. The five levers below are the reliable, low-regret moves — the things that clearly help and cannot hurt.

Lever one is embarrassingly basic and the most common failure: your robots.txt may be blocking AI crawlers without you knowing. Some CDNs and firewalls offer one-click AI-bot blocking, and blanket disallow rules copied from templates catch these bots by accident. Note the distinction between training crawlers and answer crawlers: blocking OpenAI's GPTBot keeps you out of future training data, but OAI-SearchBot is what feeds ChatGPT's live search features — you can allow one without the other. Blocking PerplexityBot removes you from Perplexity's citations outright.

Verifying takes ten minutes: fetch your own /robots.txt and read it, then check your server or CDN logs for the user agents below. While you are there, confirm your core pages render their content without JavaScript execution — crawlers vary in how much JS they run — and load fast enough that a real-time fetch does not time out.

Lever two is structured data. JSON-LD does not make weak content strong, but it removes ambiguity: Article schema establishes dates and authorship, FAQPage marks exactly what is a question and what is its answer, HowTo labels steps, and Organization says who is publishing. Ambiguity is expensive when a machine decides in milliseconds whether your page answers the question. One rule keeps you safe: schema must mirror what is visibly on the page — never mark up content that is not there.

Lever three is the content pattern this very page uses: phrase headings as the questions people actually ask, and put a direct, self-contained answer in the first two to four sentences underneath. Not a teaser, not context-setting — the answer, complete enough to be quoted verbatim without the rest of the page. Elaborate after. This mirrors how retrieval systems chunk pages into passages: the heading tells the engine which question the chunk addresses, and the answer block gives it something clean to lift.

Beyond structure, feed the engine specifics. Comparison tables, numbered steps, and definition-style openings are all formats a model can extract without interpretation. Concrete, verifiable statements get quoted; vague ones give the model nothing to work with — "pricing starts at €39 a month" is citable, "affordable plans" is not. Counterintuitively, acknowledging limitations helps too: a passage that states what something does not do reads as complete and balanced, which is exactly what a synthesized answer needs.

Lever four is entity signals: helping engines resolve who you are before deciding whether to trust you. Add Organization (or Person) schema with sameAs links pointing to your real profiles — LinkedIn, GitHub, industry directories. Use one consistent name and description everywhere, give authors real bios, and accumulate third-party pages that describe you the same way you describe yourself. Engines cross-reference; an entity that checks out consistently is safer to cite than one that appears from nowhere.

llms.txt is an emerging convention, proposed in 2024: a markdown file at /llms.txt that gives language models a curated map of your site — what it is, and which pages best explain it. Here is the honest assessment: so far, no major engine has publicly confirmed it uses llms.txt, and there is no proven ranking or citation effect. It also takes about an hour to write and breaks nothing. That makes it cheap insurance against the convention being adopted — worth doing fifth, never first, and never instead of the four levers above.

Measurement is where AEO differs most from SEO: there is no built-in console showing your AI-answer impressions. The workable substitute is a disciplined manual loop, and it tells you more than any rank tracker — because you see not just whether you are cited but how you are characterized when you are.

Set expectations by mechanism. Retrieval-based citations can appear within days or weeks of your pages being indexed and improved; presence in training data shifts only when new models ship, which is a matter of months. Treat AEO as a monitoring discipline with a monthly cadence, not a campaign with a launch date.

The five levers are mostly a one-time project. The measurement loop is not — it is recurring, repetitive, and the first thing that slips when client work gets busy. That shape of work is what AI agents are for. Brohns includes a built-in GEO auditor: an agent that scans how findable your site is in AI answers, benchmarks you against competitors on the questions that matter to your business, and generates concrete fixes — crawler rules, schema, answer blocks — instead of a score with no next step.

The auditor's authority ends at the recommendation. Every fix arrives as a draft with the agent's reasoning attached to a visible timeline, and you decide what actually changes on your site — the agent can analyze all day, but publishing anything is a decision that stays yours. If AI findability is one piece of a larger goal, like more inbound leads for an agency, you describe that goal in plain language and Bro assembles the auditor into a team alongside agents that find and qualify prospects. The 14-day free trial includes 500 credits and needs no credit card, which is enough to run your first scan and see what the engines currently say about you.