Somebody asked ChatGPT who they should hire for the thing you do, and it gave them three names, and you weren't one of them.
That's the whole problem, and it's a different problem than the one SEO solves. AEO, which stands for Answer Engine Optimization, is the work of being one of those three names. It's the practice of getting your business quoted inside the answers that ChatGPT, Perplexity, Gemini, and Google's AI Overviews hand people directly. Ranking fourth on a page of ten blue links used to be worth something. Now the answer arrives before anybody clicks, and if you're not in it, you may as well not be there.
The acronyms are a mess right now. AEO, GEO, LLMO, AI SEO. People use them interchangeably and there's no settled winner yet. I use AEO because "answer engine" describes what these things are. Don't get precious about the label.
Why this is a real shift and not a rebrand
Two numbers explain the urgency.
The first is that clicks are drying up at the top. When Google shows an AI Overview, the top-ranking page loses roughly 58% of its clicks, and across all queries that trigger one, click-through rate is down around 15%. So you can still be #1 and watch your traffic fall, because the answer sits above you and most people never scroll past it.
The second number is stranger, and I think it's the one that matters. Only about 12% of the URLs that AI tools cite overlap with Google's top-10 organic results. Twelve percent. Being great at classic SEO buys you a lot of things, but it doesn't reliably buy you a seat in the answer. These are two different games running on the same board.
You can be the best-ranked page for a question and still not be the answer to it.
The good news is that the traffic that does come through converts better. People arriving from an AI answer have already had the category explained to them and already had you vouched for, so they show up further down the funnel than a cold search click ever did. Fewer visitors, warmer visitors.
Two people worth listening to, who don't agree
This field is maybe eighteen months old and the people who know it best are openly arguing with each other. Here are two worth tracking, picked because they pull in opposite directions.
Rand Fishkin founded Moz, then SparkToro, and has spent fifteen years publishing real click data instead of vibes. His read is the deflationary one. Yes, 68% of US Google searches now end without a click, but Google's AI Mode is still a rounding error at roughly a third of one percent of searches, and the AI-search gold rush is running well ahead of the evidence. His conclusion is that traffic as a KPI is structurally broken, and that the answer isn't a new acronym, it's building brand influence in the places your audience already spends time, whether or not those places send you a visit you can attribute.
Mike King runs iPullRank, was Search Engine Land's Search Marketer of the Year in 2025, and comes at this from information retrieval. His read is the constructive one. These systems chunk your content, embed it, fan a query out into sub-questions, and retrieve passages, and every one of those steps is something you can build for. He calls it relevance engineering. The point is that getting cited by AI isn't magic, it's a pipeline with observable inputs.
I think they're both right and the disagreement is smaller than it looks. Fishkin is telling you not to torch your budget on a channel that's still small. King is telling you the mechanics are real and learnable. Put those together and you land somewhere sane: don't reorganize your business around AEO, and don't be sloppy about it either. Do the structural work, because it's cheap and it compounds and it makes the site better anyway. Spend the rest of your energy on being worth citing.
What actually works
Most guides open with tactics that do nothing, so let me start with the one to skip.
Skip llms.txt. It's the file everyone recommends and right now it's theater. About 10% of domains ship one, 97% of those files have never been fetched by anything, and Google has said outright that it doesn't use them for AI Overviews or AI Mode. It costs you twenty minutes and it makes you feel like you did something. Do the rest of this instead.
Everything below does work, roughly in order of return per hour spent:
| # | What to do | Why it works |
|---|---|---|
| 1 | Answer the question in the first 50 words | Models extract passages, not pages. Put the direct answer directly under the H2, then elaborate. |
| 2 | Write in question-shaped headings | H2s phrased the way a person asks them. The model is matching a question to a chunk of text, so make the match obvious. |
| 3 | Ship JSON-LD schema | FAQ, HowTo, Organization, Article. Unlike llms.txt, structured data really does help machines parse what your page is and who's saying it. |
| 4 | Build comparison tables | "X vs. Y," "alternatives to Z," "best tools for [job]." Models love a clean table with named competitors. It's the single most citable format there is. |
| 5 | Get mentioned where the models read | Reddit is the #1 cited domain across ChatGPT, Perplexity, Gemini, and AI Overviews. Wikipedia and Reddit together account for over a quarter of ChatGPT's US citations. |
| 6 | Be specific and be attributable | Numbers, dates, named things, a real person's name attached to a real claim. Vague marketing copy is unquotable, so it doesn't get quoted. |
Number 5 is the one people resist and the one with the biggest multiplier. Domains with heavy Reddit presence average around 3.9x the ChatGPT citations of domains with barely any. You can't fix that on your own website. It's a go-be-a-real-participant-where-your-customers-already-talk problem, which is slower and less comfortable and works better.
Finding those places doesn't take a tool. Ask yourself where you would look if you were trying to get a read on something you'd never bought before. You'd search it, you'd skim a comparison page or two, you'd check reviews, and you'd probably go read what unaffiliated strangers said about it somewhere. That's the list. For most software it means Reddit, Capterra, G2, and Product Hunt. If you have an app, App Store and Google Play reviews count too. Whatever your version of that list turns out to be, that's where the models are reading about you, and right now somebody else is writing that copy.
What a citable page actually looks like
Every page that gets cited has roughly the same skeleton, and it's the same skeleton whether a person or a model is reading it.
- A question as the heading. The literal question somebody would type, not a clever title.
- The answer immediately underneath, in two or three sentences, before any setup or story or context-building.
- The step-by-step below that. How to actually do the thing.
- An FAQ at the bottom, catching the adjacent questions your first answer raises.
Clay does this about as well as anyone right now, and it's worth pulling up one of their pages just to look at the shape of it. The reason the structure works on models is the same reason it works on people. Somebody showed up with a question and you answered it before making them work for it. If a page is only good for machines, you've built slop, and slop gets cited once and trusted never.
How to know if it's working
This is where it falls apart for most people, because the old dashboard doesn't measure the new thing. Three things to watch:
Ask the engines directly. Once a month, run your ten most important buying questions through ChatGPT, Perplexity, Gemini, and Google AI Mode, and write down whether you appear and what they say about you.
Do it in an incognito window, or better, on somebody else's computer. These systems personalize hard against your account and your own history, so checking while logged in as yourself shows you a flattering version of reality that no prospect will ever see. That personalization is also why this is genuinely difficult to measure, and why any single check is a data point rather than a verdict. It's manual and it's tedious and it's still the closest thing to ground truth you have. There are tools that automate it now. They're fine, they're just not necessary at the start.
Watch referrals by source. AI referral traffic shows up in analytics under real hostnames, so segment it out. Small numbers are normal. The direction is what you're reading.
Watch what people say when they arrive. The tell that AEO is working shows up on a discovery call, when somebody already knows what you do and you spend the first ten minutes on their problem instead of on your explanation.
What this looked like when I did it
I built this machine for CrewLAB earlier this year. They're a sports team management platform, and the people who actually buy are club treasurers and administrators, the ones who own the money and the registrations. The goal was never traffic. It was booked calls with those people, from a channel that would keep working after we stopped pushing it.
The build was six deep pages, one per dimension of the problem their customers actually have, plus four comparison tables naming the tools those customers were already using. Every page answered a real question in the first paragraph and carried FAQ schema, and they all pointed at one free financial-health audit tool sitting on a crawlable subfolder URL. Real HTML, not a black-box iframe, because a tool the models can't read is a tool that can't be cited. Then we instrumented the whole funnel, so we could answer the question that actually pays the bill: did anybody book a call off this.
Eleven articles, one tool, one measurable path. The part I'd underline is that it was quality over volume. Six pages that go deep beat sixty that skim, because the model is hunting for the page that most completely answers the question, and depth is how you win that.
Where to start this week
Start on your own website, always. It's the only surface where you control both the words and the code, and it's the one crawlers read most reliably. Social platforms are a worse bet than they look, and Instagram in particular is close to invisible here, partly because it's hard to crawl and partly because Meta and Google are in a slow existential fight over who owns the future of the internet. Neither is inclined to do the other any favors.
If you do nothing else, do these three, in this order:
- Take your five highest-intent buying questions, the ones a prospect types right before they'd contact you, and write one deep page for each. Not a blog post. A page you'd be proud to have quoted.
- Add FAQ and Organization schema to those pages and to your homepage. It's an afternoon of work.
- Run those five questions through ChatGPT and Perplexity today and screenshot what comes back, so you have a before.
The third one takes fifteen minutes and it's the one everybody skips, and then six months later there's no way to prove anything changed.
AEO doesn't replace SEO, and anybody telling you it does is selling something. The two overlap and they feed each other, and a crawlable, well-structured, useful site tends to win at both. I wrote the companion piece on the fundamentals here: SEO: A Five-Minute Primer.
If you'd rather someone just built the machine, that's what I do. The CrewLAB version is written up here.
Figures cited: Conductor's 2026 AEO/GEO benchmarks, SparkToro's 2026 zero-click study, SE Ranking's llms.txt analysis, 5W Research's state of AI citations, and Search Engine Land's citation-source reporting. Positions attributed to Rand Fishkin (SparkToro) and Mike King (iPullRank) are my summary of their published work, not quotes. Numbers in this space move fast, so treat them as direction rather than gospel.
