Somebody looking for a plumber in Manchester used to type the query, look at a list of ten links, and pick one. That user increasingly asks ChatGPT instead, reads a written answer, and follows one of the three sources it names. If you are not one of those three, your position in the classic rankings is irrelevant to that person. They never saw the list.
What answer engine optimization actually is
Answer engine optimization is the discipline of making your site the source an AI system reaches for when it answers a question in your category. The mechanics are different from ranking, because the thing being selected is different. A search engine picks pages. An answer engine picks sentences.
That distinction drives almost everything else. A page can rank first and still never be quoted, because its useful content is buried in paragraph nine behind an introduction about how the industry is changing. A page can sit at position six and be quoted constantly, because its second sentence answers the question cleanly enough to lift.
AEO, GEO and SEO are not the same job
The terms get used interchangeably and it causes real confusion when you are buying the work, so here is how I separate them.
- SEO earns a position in a list of links. The unit of success is a ranking.
- AEO earns a citation inside a generated answer to a specific question. The unit of success is being named.
- GEO, generative engine optimization, is the broader job of shaping what models say about your brand across every prompt, including prompts that never mention you. The unit of success is share of voice.
In practice the work overlaps heavily and any agency selling them as three separate retainers is selling you the same work three times. What matters is that all three depend on the same foundation: clean, credible, well structured information that a machine can read without guessing.
How answer engines pick their sources
None of these systems publish their selection logic, so anyone claiming certainty is guessing. What can be observed, by running hundreds of queries and recording what comes back, is a consistent pattern.
The model retrieves a set of candidate documents, usually via a search index. It then reads them and decides which passages it can restate confidently. Confidence is the operative word. A passage that is unambiguous, self contained, specific and corroborated elsewhere is safe to quote. A passage that is vague, promotional, or dependent on three paragraphs of preceding context is not, so it gets skipped in favor of a competitor who wrote more plainly.
This is why the winners in AI citations are so often not the biggest brands in a category. They are the sites that happened to write clearly.
Structure: the first 60 words decide most of it
The single highest leverage change you can make is to answer the question at the top of the page, before any framing, in plain declarative sentences. Analysis of citation patterns consistently finds that a large share of quoted passages come from the opening portion of a document rather than its middle.
Concretely, on a page targeting a question:
- State the answer in one or two sentences, in the first 40 to 60 words, using the same words a person would use to ask it.
- Then expand. Context, nuance, exceptions and your opinion all belong after the answer, not before it.
- Give every H2 a specific, descriptive name. "How answer engines pick their sources" is extractable. "Key considerations" is not.
- Write each section so it survives being read alone. If a paragraph only makes sense after the three before it, a model cannot lift it.
- Use lists and tables where the content is genuinely list shaped. Structured formats are quoted at noticeably higher rates than prose, because they are easier to lift without distortion.
There is a version of this advice that turns into writing for machines and produces unreadable pages. Resist it. The reason answer first structure works is that it is also better for humans, who have been skimming for the answer since long before any of this existed.
Schema: removing the guesswork
Structured data does not persuade a model. It removes ambiguity, which is a different and more useful thing. Without it, a model has to infer from your page text whether Charliez Digital is a person, a company, a product or a place, and inference is where you get misrepresented.
The types that carry weight for most businesses:
- Organization or ProfessionalService with name, url, logo, areaServed, and sameAs links to your real profiles. This is your identity anchor.
- FAQPage on any page with genuine questions and answers. This is the single most extractable format there is, because it maps exactly onto how an answer engine works.
- Article or BlogPosting with a named author, datePublished and dateModified. Freshness and attribution both feed confidence.
- Service with areaServed, so a model knows you can actually serve the person asking.
- BreadcrumbList, which helps the system understand where a page sits in your structure.
Validate everything before you ship it. Malformed schema is worse than none, because it teaches the model that your markup cannot be trusted.
The part nobody wants to hear
On page work gets you into the candidate set. Off site corroboration is what gets you quoted repeatedly instead of occasionally.
When a model finds a claim on one domain, it hedges. When it finds the same claim on five independent domains, it states it as fact and attributes it to whichever source it trusts most. That is why review platforms, industry directories, association listings, genuine guest articles and real participation in the communities your buyers read matter more here than they do for classic rankings.
Reddit deserves a specific mention because it makes up a strikingly large share of citations in some systems, Perplexity especially. That is not an invitation to spam it, which will get you banned and does not work anyway. It is a reason to actually answer questions in the subreddits where your customers already are.
This is slow work and it is the reason most AEO engagements underdeliver. The on page part takes a fortnight. The corroboration part takes months, and it is the part that lasts.
How to measure it without a tool
There is no Search Console for AI citations. Anyone selling you a dashboard is inferring, not reporting. What you can do is build a manual baseline, which is more work and more honest.
- Write down the fifteen to twenty questions your buyers actually ask before they buy. Not keywords. Questions.
- Run every one through ChatGPT, Perplexity, Claude and Google AI Overviews. Record whether you are named, which competitors are, and which page the model cited.
- Repeat on a fixed schedule, quarterly at minimum. Compare citation share over time, not absolute counts, because model behavior drifts.
Two warnings from doing this repeatedly. Results vary between sessions and users, so a single run tells you very little and you need the same question set every time. And a citation is not a visit. Treat this as brand presence measurement, and expect referral traffic from it to be real but modest.
Five mistakes I see constantly
- Burying the answer. Four hundred words of industry context before the page says anything useful. The model reads the top and moves on.
- Writing for the model instead of the reader. Keyword stuffed, robotic, question shaped headings with no substance. It reads as low quality to a person and as low confidence to a model.
- Schema with nothing behind it. FAQPage markup on a page with no visible questions and answers. It is a guidelines violation and it is transparent.
- No named author. An anonymous page is a page with no accountability attached. Put a real person on it, with a real role and a link to a real profile.
- Skipping the baseline. Starting the work without recording where you were, which makes it impossible to know afterwards whether any of it worked.
None of this is complicated. It is mostly the discipline of saying the useful thing first and then proving it, which is also just good writing. The advantage available right now is that very few businesses in most categories are doing it at all.
