AEO is optimizing for the machines that answer questions instead of listing links. The work is mostly unglamorous: be findable, be quotable, be corroborated somewhere else, and be structured well enough that a model can lift a sentence out of you without guessing. Rankings become citations. Position becomes presence.
What an answer engine actually is
An answer engine is any system that responds to a question with a written answer rather than a list of results. ChatGPT is one. Google's AI Overviews are one. Perplexity, Claude, Copilot, Gemini and the voice assistant in a car dashboard are all one. They differ in plumbing and agree on behaviour: read a lot, decide what is relevant, write a short confident paragraph, sometimes name sources.
That last clause is the whole business. The engine collapses a page of options down to two or three names. If you are not one of the names, you are not in the consideration set at all — and unlike page two of Google, there is no page two to be on.
Ten blue links were a shortlist you curated yourself. An answer is a shortlist someone else already curated for you.
How a model decides who to name
There are two paths into an answer, and they behave differently.
Retrieval. The engine runs a search behind the scenes, pulls a handful of pages, and writes from what it read. This is what ChatGPT does when it browses, what Perplexity does always, and what AI Overviews do by design. Retrieval rewards pages that answer a specific question cleanly and load fast enough to be fetched.
Training memory. The model already absorbed something about you months ago. No fetch happens. This path rewards being described consistently in many places — directories, press, forums, industry lists — because the model learned a consensus rather than a page.
Most real answers mix both. Which is why AEO is not one tactic: you are trying to be the best available source right now and the name that already sits in the model's head.
Citations are the new rankings
The old metric was position: where do I sit for this keyword. The new metric is share of citation: across the questions my customers actually ask, how often does the answer contain my name, and what does it say about me when it does.
Three things are worth counting, and only three. Presence: are you named at all. Framing: are you the recommendation, a footnote, or the cautionary example. Accuracy: are the facts in the sentence about you true. A confident wrong answer costs more than an absent one, because customers act on it.
Share of citation — the percentage of a defined question set where an engine names you. Only meaningful against a fixed list of questions, re-run on a schedule. A number without its question set is decoration.
The eleven fixes, in the order they pay
This is the list we work through on every engagement. It is ordered by return, not by difficulty.
Half of invisible sites are invisible on purpose: a robots.txt line, a JS-only render, an aggressive bot firewall. Check for GPTBot, ClaudeBot, PerplexityBot and Google-Extended explicitly.
Put the extractable answer at the top of the page, before the story. Models lift the first coherent paragraph that matches.
A page covering nine topics gets cited for none. Split it.
Prices, hours, service area, credentials, capacity — in text, on your own site, in one canonical place. Not in an image, not only in a PDF.
Organization, LocalBusiness, FAQPage, Product, Person. Schema does not force a citation, but it removes ambiguity about what your sentences mean.
Same legal name, same address format, same category words everywhere. Models resolve entities by agreement; disagreement gets you merged with someone else or dropped.
Third-party mentions carry more weight than your own claims. Industry directories, local press, association listings, honest review platforms, relevant subreddits.
“Best X in Y” and “A vs B” are the highest-intent questions asked of assistants. If you never write those pages, someone else writes them about you.
Engines prefer sources that look current. A visible updated date and a real revision beat a rewrite of the intro.
When an engine states something false about you, find the source it likely read and correct that, rather than arguing with the model.
Answers drift weekly. A single audit is a snapshot; monthly re-runs are the actual product.
How to measure any of this
Write down thirty to fifty questions a real customer would type — including the ugly ones about price, trust and alternatives. Ask each of them to every engine that matters to you. Record whether you were named, how you were described, and which sources the engine leaned on. Repeat monthly and diff.
That is genuinely the method. It is tedious by hand, which is why Uwasa runs it for you and puts the diff in a report your client can read without a glossary.
Common questions
Does AEO replace SEO?
No. Retrieval still runs on search infrastructure, so a site that cannot rank usually cannot be retrieved either. AEO is a layer on top, not a replacement. The comparison guide pulls the three apart properly.
How long until anything changes?
Retrieval-path fixes — crawlability, question-shaped pages, schema — can show up within a few weeks. Anything that depends on the model's memory of you moves on a scale of months.
Can I pay to be in an AI answer?
Not the cited part, not today. Ads are appearing around AI answers, but the answer body is earned. Anyone selling guaranteed placement is selling something else.
What if the engine says something false about us?
Trace it. Wrong claims almost always come from a real source: a stale directory, an old press release, a competitor's comparison page. Correct the source, then re-ask.
The free check asks six engines about your business and shows you the answers.