Answer engine optimization. The haunting, named.
AEO · also sold as GEOAEO is the work of getting a business named inside the answers AI assistants write. Not ranked on a page of links. Named, in a sentence, at the moment someone asks what to book, buy or visit.
Generative engine optimization (GEO) is the same work under a different banner, and LLM SEO is a third name for it. Vendors pick whichever term their buyers search for. Whatever the label, the mechanics don’t change: the assistant searches, reads a handful of pages, and names the businesses those pages support. AEO makes sure some of those pages support you.
In the register: AEO is the whole monthly cycle. Log prompts, test and stamp, investigate who took your spot, ship the fix, re-run. Every VANISHED stamp is an AEO problem with a work item attached.
AEO vs SEO. Links vs names.
Same ground · different prizeSEO earns you a position in a list of links. The customer still has to click, compare and decide. AEO earns you something smaller and more final: your name in the answer itself, often with no list underneath and no clicking left to do.
The two aren’t rivals. AI assistants run on search indexes, so a site that ranks nowhere is hard for an assistant to find at all. SEO gets you into the room. AEO determines whether the assistant says your name once it’s read the room.
In the register: the “Ranks top-10 for its own prompt set in classic search” factor is exactly this bridge. Page three in classic search is rarely page one in an AI answer.
Answer share. The meter, in industry units.
AI share of voiceAnswer share is the percentage of relevant AI answers that name your business. Ask the engines the questions your customers ask, many times, across many prompts, and count who gets named. Fifty prompts, named in eleven answers: a 22% answer share.
It’s the number an AI visibility report is built around, usually broken out by engine, because a brand can be solid on Perplexity and a ghost on ChatGPT at the same time. The engines read different indexes and trust different sources.
In the register: the Apparition Meter is your answer share, drawn as opacity. A faint wisp at 15%. Solid and smiling past 60%. Same math, better ghost.
First-pick rate. Named first, booked first.
Position inside the answerBeing named isn’t the whole story. An AI answer usually offers three to six options, and the first one carries the recommendation’s weight. First-pick rate is the share of answers where you’re not just present but named first.
Two businesses can hold the same answer share and live in different worlds: one opens every list, the other closes them. When you read a visibility report, ask for both numbers.
In the register: every SIGHTED stamp records where in the answer you appeared. The Battleground shows who opens the list on your money prompts, and how often it’s you.
Your chief haunter. The rival in your place.
Strongest rivalEvery answer that doesn’t name you names someone else. Tally those names across a cycle and one rival usually dominates: the business the engines reach for when your prompt comes up. That’s your chief haunter, and it’s the most useful single fact in a visibility report.
It’s useful because it’s diagnosable. Look at which sources the AI cited when it chose them, and you’re looking at the exact pages you need parity with. A haunter isn’t beaten in general. It’s beaten source by source.
In the register: the Haunting Grounds keeps this tally per property, with every haunter’s count and the séance sources behind each appearance.
Like-for-like. Same prompt, next moon.
The recheck that countsAI answers vary run to run, so a visibility number only means something when it’s measured the same way twice. Like-for-like means re-running the identical prompt, on the same engine, in a fresh conversation, a cycle later, and comparing those two results and nothing else.
Change the prompt wording between cycles and you haven’t measured improvement. You’ve measured a different question. This is the most common way AI visibility numbers go wrong, in both directions.
In the register: prompts carry forward between cycles unchanged, and the ★ MATERIALIZED badge only fires on a like-for-like flip. Ghost in July, sighted in August. Proof, not vibes.
The crawlers. The machines that read you.
GPTBot, ClaudeBot & the restEach AI company sends more than one robot to your site, and they do different jobs. Some collect pages that may train future models. Some build the search index the assistant queries. Some fetch a page live because a person just asked about you.
The distinction matters because the common advice, “block the AI bots,” blocks all three at once. Block a training crawler and you’ve expressed a preference. Block a search crawler or a live fetcher and you’ve asked the assistant to forget you exist.
| Robot | Job | If blocked |
|---|
| GPTBot | Collects pages that may train OpenAI models | Your choice |
| OAI-SearchBot | Indexes your site for ChatGPT search | Keep open |
| ChatGPT-User | Fetches a page live when a user asks | Keep open |
| ClaudeBot | Collects pages that may train Claude | Your choice |
| Claude-SearchBot | Indexes your site for Claude search | Keep open |
| Claude-User | Fetches a page live when a user asks | Keep open |
| PerplexityBot | Indexes your site for Perplexity | Keep open |
| Perplexity-User | Fetches a page live when a user asks | Keep open |
| Google-Extended | Controls use of your pages for Gemini training | Your choice |
In the register: the “AI crawler access” factor checks your robots.txt for exactly this. A site that’s invisible to the search crawlers fails the whole exam before the first question.
Zero-click. The sale you never saw.
Why your analytics can’t see thisA zero-click search ends without a visit to any website, because the answer itself was enough. AI assistants push this further: the guest reads a recommendation, forms a shortlist, sometimes books, and your analytics records nothing at all.
This is why AI visibility can’t be inferred from traffic. By the time a drop shows up in GA4, you’ve been vanishing from answers for months. The only way to know is to ask the engines directly and count.
In the register: that counting is the whole product. Every test prompt carries a funnel stage, and a miss at the moment of purchase counts three times a miss at the daydream stage.