When someone asks ChatGPT or Perplexity "what is the best tool for X", the answer names two to five brands. Answer engine optimization is the work of becoming one of them. Here is how it differs from SEO, how to measure where you stand, and the six changes that reliably help.
Written September 2026. The category is young and the assistants change monthly; the mechanics below are the ones that have held up across re-measurement.
Answer engine optimization (AEO) — the same idea also travels as generative engine optimization (GEO) and LLM SEO — is the practice of making a brand more likely to be named, recommended and cited when a person asks an AI assistant a question in that brand's category. The "answer engines" are ChatGPT, Perplexity, Gemini, Claude, Microsoft Copilot and Google's AI Overviews and AI Mode.
The important difference from search is the shape of the output. A search results page has ten slots and a hundred million pages compete for them. An assistant's answer has room for a handful of names, chosen from a few pages it retrieved a moment ago, and it presents them as a recommendation rather than a list. If you are not in the answer you are not "on page two" — you do not exist for that question.
Assistants with web access do something close to this for a buyer question: rewrite it into a few search queries, fetch a small number of pages, and write an answer grounded in whatever those pages say. Three consequences follow.
For a "best X" or "X alternatives" question the retrieved pages are mostly comparison articles, review sites, directories and forum threads — not vendor homepages. When we audited Ghost, the assistants named it in 91.7% of answers but cited ghost.org in only 12.5%; they were describing Ghost using pages on Substack, Wix and WordPress. How other sites describe you is most of your visibility.
The same question asked on two days can return different names. A single measurement is a coin flip; a dozen questions across two assistants, repeated, is a measurement. Any tool that reports one number from one run is showing you noise.
A surprising share of sites block the AI crawlers in robots.txt or behind bot protection, or render their key facts only through JavaScript, which most retrieval fetchers do not execute. Those sites can be excellent and still be invisible, because the fetcher saw an empty page.
You need a question set, not a keyword list. Write the ten to twenty things a buyer actually types — "best X for small teams", "X alternatives", "X vs Y", "is X worth it", "what do people complain about with X". Ask each assistant every question and record, per answer:
| Number | What it tells you | Typical fix if it is low |
|---|---|---|
| Mention rate | How often your brand name appears in the answer | Get onto the third-party pages the assistant is reading |
| Citation rate | How often your own site is linked as a source | Crawlability, plain-HTML facts, question-shaped pages |
| Named instead | Which competitors fill the slots you are missing | Find which of their pages get cited and match the format |
| Sources cited | The pages the answers were assembled from | This is your outreach and listing to-do list |
Keep mention rate and citation rate separate. They have different causes and different fixes, and a blended "visibility score" hides which one you have.
Roughly in order of effect for a typical small or mid-sized company:
1. Let the crawlers in. Check robots.txt for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and the rest, and check that your CDN's bot protection is not serving them a challenge page. This is a five-minute fix and it is the most common reason a good site is never cited.
2. Be on the pages that get cited. Your measurement told you which review sites, directories and comparison articles the assistants are reading. Make sure you are listed on them, accurately, with the same category words the buyer uses. This is the single biggest lever for mention rate and it is mostly listings work, not writing.
3. Publish question-shaped pages. One page per real buyer question, with the answer in the first paragraph and the comparison in a table. Assistants retrieve pages that look like answers to the query they just generated. A page titled "X vs Y for small teams" gets pulled for that question; a homepage does not.
4. Put the facts in plain HTML. Pricing, who it is for, what it integrates with, what it does not do — as text in the initial HTML, not rendered later by a script. If a fetcher without JavaScript sees nothing, you get cited for nothing.
5. Add structured data. Organization, Product or SoftwareApplication and FAQPage JSON-LD make the entity unambiguous: this name, this site, this category. It does not make you rank; it stops you being confused with something else.
6. Keep an llms.txt. A short plain-text summary of what you are and where the important pages live. Not every assistant reads it yet; it costs nothing and some do.
Then re-measure after four to eight weeks with the same question set. Expect movement on citation rate first — it is the part you control directly — and mention rate later, as third-party pages update.
For the measurement, yes — doing a dozen questions on two assistants by hand, repeatedly, is tedious and you will stop. There is a comparison of the paid dashboards and their real prices; they range from $29 to several hundred a month and almost all of them stop at the chart. For a one-off answer, a pay-per-check run is a few dollars.
For the work — the listings, the question-shaped pages, the fixes — you need either a person or a service that includes it. That is the gap most of the category leaves open.