How does AEO actually work?
AEO works by influencing what an AI engine retrieves before it writes anything. Modern engines fan a question out into multiple searches, gather candidate sources, select individual passages, and compose an answer from them. Optimizing for that pipeline means making a site accessible to the crawlers that feed each engine, then supplying clear answers to the exact questions buyers ask, backed by corroboration on sources the engine already trusts.
Retrieval comes before generation
Google's documentation describes AI Overviews and AI Mode as able to "fan out" into multiple searches across subtopics and sources before composing a response (Google's AI features documentation). The practical consequence is that when a buyer asks about your category, the model is assembling sources in real time rather than reciting something it memorized during training. Your page has to clear retrieval first, then get selected as a passage worth quoting, and it has to carry enough source credibility for the model to name you ahead of a competitor it already has reason to trust.
Kevin Barber, our Head of AI Growth, is direct about the prerequisite: "It's super important that we provide the website in a structured, machine-readable way that gives AI everything it needs to ingest your copy and understand your brand." In practice that means putting the direct answer in the first sentence under each heading, using headings phrased the way buyers actually type their questions, and then checking whether any fact you want cited still makes sense once it's lifted away from the page around it.
Each engine reaches your site differently
Crawl permissions decide whether any of the content work matters. OpenAI uses OAI-SearchBot for ChatGPT's search discovery, so a blanket bot block quietly removes you from that surface no matter how good the writing is (OpenAI's publisher FAQ). Perplexity separates its indexing crawler from user-requested fetches, which means the rules you set for one don't govern the other (Perplexity's crawler documentation).
Before we touch content, we check robots.txt, WAF and bot-mitigation rules, indexation status, and whether the answer text is present in the served HTML or gets rendered later by JavaScript. That last one is the issue we run into most often on otherwise healthy sites, and it's usually a fast fix once somebody looks.
How the loop runs
We start with a baseline: a fixed panel of buyer prompts run across ChatGPT, Claude, Gemini, and Perplexity, recording which brands get named, which sources get cited, and how your company gets described on the occasions it does appear. The prompts where you're absent or described inaccurately become the gap list, and the content plan is built directly off that list rather than off a keyword export.
From there we structure the on-site answers and brand facts, then go earn corroboration on the sources the engines are already pulling from, since a claim you make about yourself carries less weight with a model than the same claim appearing on a publication it already trusts. Then we rerun the identical prompt panel and close whatever is still missing. That rerun matters more than any single asset we publish, because the engines keep shifting their source preferences and without a second measurement you have no reliable way to tell whether a change helped. We put together a demo site that walks through the mechanics if you want to see the structure applied end to end.
