Sometime in the last month, an agency probably emailed you about answer engine optimization. Maybe it was a newsletter subject line, maybe a slide in a pitch. The acronym is AEO, and the premise is simple: people are starting to ask AI assistants the questions they used to type into Google, and for a fee, someone will get those assistants - ChatGPT, Claude, Perplexity, Google’s own AI answers - to cite your site when they respond. The same way they used to sell SEO, which is just helping you show up when someone does type a question directly into Google.
The shift underneath the pitch is real. How someone finds a treatment center, a law firm, or a financial advisor is changing, and AI is now part of it. The problem is the tactics being sold to capture it. The bill for those tactics arrives whether or not they do anything, so it is worth knowing which ones work.
We run an AEO setup on this very site. So we can tell you what the data actually shows, including about the things we built ourselves.
Why this matters: someone checks you with an AI before they call
Start with why anyone should care, because the stakes here are not abstract.
In the work we do, the buyer does a lot of quiet homework and then decides on trust, usually with a referral somewhere in the mix. That has not changed. What has changed is one step in the middle: the referral who used to walk straight to your website now asks an AI about you on the way.
It is the same pattern in every field we work in. The 2026 Legal CX Report from Case Status and Researchscape found that how a client is treated now rivals the outcome itself in winning loyalty, and that referrals bring in roughly 47 percent of new clients while just 5 percent say advertising most influenced their choice. Trade coverage in financial planning described high-net-worth clients using AI tools to vet advisors, and made a point worth repeating: the same web presence that makes you findable on Google is what makes you findable to an AI. There is no separate switch to flip. In behavioral health, the trade press spent the month describing buyers who want proof in hand before they will take a meeting at all.
The person who was going to call you anyway now runs your name past a machine first. You cannot buy your way into what that machine says about you. So the only useful question is what earns a place in it.
What the AEO data actually says
Here is where the AEO pitch and the evidence part ways. Two things landed in the same week in June, and both matter.
The first is from Ahrefs, which looked at more than 137,000 sites. The finding most owners need: of the sites that had published an llms.txt file - a tidy text file you are told the assistants will read - 97 percent got zero requests for it in a month. The assistants are not going looking for it. A larger Ahrefs study found that schema markup, another staple of the AEO checklist, had close to no effect on whether AI tools cited a page.
The second is from Google itself. In its 2026 guidance on appearing in AI answers it says, in plain words, that you do not need to create special machine-readable files or markdown to show up in Google Search, and that Google ignores them. That is the company whose answers you are trying to appear in, telling you the file does nothing for it.
Two more findings from the same research, because they change the picture. AI citations have come loose from rankings: only about 38 percent of the pages cited in Google’s AI answers now come from the top ten results, down from 76 percent a year earlier, and assistants regularly cite pages that do not rank on Google at all. And the one move that measurably helped was the unglamorous one. Adding original research and real numbers to a page lifted its visibility in AI answers by 41 percent, more than any tactic tested. For what it is worth, separate research this year put the backdrop in one number: roughly two-thirds of US Google searches (68 percent) now end without a click to any website.
Put those together and the checklist falls apart. The thing that gets a page cited is having something in it worth citing, which has frankly always been the end goal of search across the internet.
The boldest experiment, and the line it crosses
The most ambitious public test of this came from Ramp, the spend-management company, this spring. They did what the pitches only hint at: a Cloudflare Worker served AI bots a different version of their pages than people saw, with an offer addressed directly to the agents, to see whether assistants would relay it to the humans asking about their category. It is the closest thing to a controlled experiment the field has, and three findings are worth your time.
Format first. Ramp served bots three versions of the same content - plain markdown, stripped HTML, and schema markup - and only markdown reliably surfaced in AI answers. Schema, the format built for machines, underperformed, which matches the Ahrefs result. Ramp is honest that a targeting flaw muddies this, so read it as a strong hint, not a law.
This is not actually at odds with the Ahrefs and Google findings above; they answer different questions. Ahrefs and Google are about discovery - the assistants will not go fetch a special file you publish, so your llms.txt sits unread. Ramp is about digestion - when an assistant does read a page, clean markdown is easier for it to parse and surface than schema-stuffed HTML. Worth noting how Ramp got that markdown in front of the bots, though: a Cloudflare Worker detected the bot and served it a markdown version while people got the normal rendered page. That is content served conditionally to machines, which is a different thing from what we do - we serve the same rendered page to every visitor, human or bot, and separately publish an open .md copy of each post at a parallel URL that anyone or any model can fetch. Same content either way; nothing is detected, swapped, or hidden. Hold that distinction; it is the line the next section is about.
Then the finding that matters most: the pages that surfaced new content were the ones the assistants already cited. Ramp calls it “agent trust,” and it behaves like domain authority - an assistant pulls from the sources it already trusts. Pages with no existing citations surfaced nothing, in any format. You do not format your way in. You earn your way in, and then the page can carry something new.
And it was wildly model-dependent. Claude relayed the offer consistently, down to the exact link. Perplexity mentioned it vaguely. ChatGPT, across thirty-two days, never surfaced it once, while plainly reading the content. What works on one assistant can do nothing on another, which is why testing against a single model tells you almost nothing.
Now the line. Ramp’s test was, in plain terms, an attempt to get agents to recommend them - an above-board offer, but aimed at the machine instead of the person. We will not do that, and we would tell you not to. The moment your marketing exists to change what an AI says about you rather than to be worth saying, you have rebuilt the hype machine with new wiring, and in healthcare, law, or finance that is a trust problem waiting to surface, not a growth tactic. Tellingly, even Ramp’s data shows the payoff half has not landed: agents relayed the offer, but no human acted on it yet. The bold version is still unproven. The boring version, being the source worth citing, is the one their own numbers keep pointing back to.
The part you cannot fake
That should sound familiar, because it is the same thing that has always earned trust. It bears repeating only because AEO is being sold as a shortcut, and there is no shortcut.
There is a second trap worth naming. Getting cited is not the same as being believed. A study this June drew exactly that line: a brand can show up in AI answers and still not persuade the reader it knows anything. In the work we do, that distinction is the whole game. A family choosing a treatment center, a person hiring a lawyer after the worst week of their life, someone handing a stranger their retirement - none of them are won by topping a list. They go to whoever reads like they actually know the subject, because they do. The ones who win go a step further: they read like they understand the reader’s particular situation, not just the subject - a distinction worth its own piece.
This is what we mean by AI in the right hands. The slop approach, where you point a model at a keyword and publish whatever falls out, fails on both counts: nothing original for an assistant to surface, and nothing a reader can feel a person stood behind. The tool is fine. What matters is whose hands it is in and what they aimed it at.
So what do you actually do
A short and honest answer.
Publish things only you could publish. Your real cases, your own numbers, the judgment of the people who do the work. That is the highest-return move in all the data, and it is the one no agency can fake on your behalf. It is also why we tell clients their marketing should sound like their front-line, not a brand committee.
Make sure a machine can actually read the page. This is the dull half, and plenty of sites get it wrong: they block the very crawlers that would have cited them. There is a real difference between the bots that train models (GPTBot, ClaudeBot) and the bots that fetch a page to answer a live question (OAI-SearchBot, Claude-SearchBot, PerplexityBot), and you want to allow that second kind. The exact names drift, so confirm them against each provider’s own docs before changing a robots.txt. It is a short sitting with whoever runs your site.
Keep the page clean and structured, so a person and a machine can both follow it. Then measure the right thing. Google now shows you how often you appear in its AI answers but not how often anyone clicked, and an impression is not a patient or a client. That is the same trap a good agency helps you avoid in the first place: knowing what to measure, not just what looks busy.
Then there is the file we built. We serve a clean text version of every page on this site, and we keep an llms.txt. The honest read on it: it helps coding tools and keeps our own house tidy, and it is not the reason anyone cites us. We left it up because it is good hygiene. We are not going to sell it to you as a strategy.
The honest version
There is no AEO trick. There is the same job there has always been: be genuinely worth citing, to a person, and now to the machine standing in front of the person. That kind of work compounds. The tricks decay the week the next model ships.
If someone is selling you an AEO package, you are allowed to ask what is in it beyond a file and a checklist. If the answer is thin, you learned something for free.
When you want a second read, book a 20-minute teardown. We will look at your site and your content, tell you what is actually earning trust and what is just noise, and hand you the short list of what we would fix first. No deck, no pitch. It is the kind of thing a content engine, not a content mill, is built to see.
Common questions
Do I need an llms.txt file?
Not for the reason you have been told. As of mid-2026, the major AI assistants do not go looking for an llms.txt file, and Google says it ignores it. It is harmless and tidy, and it is genuinely useful to coding and developer tools, so there is no reason to take one down. Just do not mistake it for a reason anyone will cite you.
Does AEO replace SEO?
No. The fundamentals overlap almost entirely. Clean, fast, well-structured pages with something worth reading on them are what earn both a Google ranking and an AI citation. AEO is mostly SEO with the honesty that the prize is now sometimes an answer instead of a click.
Will schema markup get me cited by AI?
The June 2026 research found schema had close to no effect on AI citations. Keep it for the traditional rich results it still earns in regular search, but do not add it expecting an assistant to cite you because of it.
How do I know if AI assistants can read my site?
Two things to check. First, that your site is not blocking the AI crawlers that fetch pages to answer questions (the ones that send a real visitor, as opposed to the ones that only train models). Second, that your important pages are fast and cleanly structured, since an assistant reads the same page a person does. Whoever runs your site can confirm both in a short sitting.
Is AEO worth paying for?
It depends entirely on what you are paying for. If the package is a file, some markup, and a promise, the data says you are buying very little. If it is the production of genuinely good, original, expert content that happens to also be legible to machines, then you are paying for the thing that works, and AEO is just a newer name for it.