What AI says about you when you're not in the room

I noticed it in myself before I noticed it in any analytics dashboard. Somewhere over the last couple of years, "I'll search for it" quietly became "I'll ask". Not for everything. But for the questions that used to mean opening six tabs and triangulating, the ones that start with which, or should I, or who's good at, I now ask an assistant and argue with the answer until it's useful.

I'm not unusual. ChatGPT passed 900 million weekly users in February, according to OpenAI's figures as reported by Reuters, and by June, independent app analysts were calling it the fastest app in history to reach a billion monthly users. Google's Gemini is climbing the same curve. Back in early 2024, Gartner predicted that traditional search engine volume would fall by a quarter by 2026 because of AI assistants. Here we are in the year they were pointing at, and whether that exact figure lands is Gartner's business. Google, for its part, says people are searching more than ever, and both things can be true. The pool of questions is growing, and a widening slice of it never touches a results page.

Which raises a question most business websites were never built to answer. When somebody asks one of these tools about your field, your area, your kind of service, what does it say? Does it mention you? Does it describe you correctly? You're not in the room for that conversation. Your website is, in a sense. That's the whole subject of this post.

Being the answer instead of a result

Search optimisation, whatever else it turned into, was always a fight for position. Ten links on a page, and the argument was about who stood where.

Answer engines don't hand you a page of links. They read widely, make up their mind, and give the person a paragraph, usually with a name or two in it and sometimes a short row of sources underneath. Being cited is the new page one. So the fight has changed shape. You're no longer competing to be third on a list; you're competing to be mentioned, or cited, at all. The failure mode is new too. You don't get outranked. You simply don't come up. No notice, no penalty, no consolation prize of page two. Silence, which looks exactly like not existing.

That's what answer engine optimisation, AEO, actually is once you strip the acronym off. It's the work of making sure that when one of these tools is asked a question in your field, it can find you, understand what you are, and feel safe saying your name. Not a bag of tricks. Mostly it's clarity, applied with unusual discipline.

The machines were always reading. Now they're deciding.

Here's a number that reframed things for me. Cloudflare, who sit in front of an enormous share of the web's traffic, have started publishing what they call crawl-to-refer ratios: how many pages an AI platform fetches for every one human visitor it sends back. For some platforms the figure runs from a few hundred to one all the way into the tens of thousands to one.

Sit with that. The busiest reader your website has ever had is software. It fetches everything, over and over, and it clicks on nothing. The ratios are rough measures, and they wobble for technical reasons, but the direction isn't in doubt. Then a person asks a question, and some version of what those machines took away decides whether you're worth mentioning.

So if software is doing most of the reading, it's worth asking what software actually understands when it reads you. The honest answer, for most websites, is: less than you'd hope.

Schema, or the label on the tin

Every website is really two websites. There's the one humans see: the design, the photography, the words you agonised over. Underneath it sits a second one, written for machines, which most owners have never once looked at. The shared vocabulary for that second website is called Schema.org, set up by Google, Microsoft and Yahoo back in 2011, with Yandex joining the same year, and it has been quietly available for fifteen years.

Structured data is not clever. That's its entire virtue. It's a set of plain declarations sitting invisibly in the page: this page describes an organisation, here is its name, here is the person who runs it, and that person holds these qualifications. This is a course. It costs this. It starts then. A human reading your site infers all of that from context, tone and layout. A machine would much rather be told. Inference is expensive and machines get it wrong; declarations are cheap, and they're only ever as wrong as the person who wrote them.

For years the reward for maintaining that second website was rich results in Google: the review stars, the event dates, the FAQ dropdowns. Useful, but a garnish, which explains the state it's in almost everywhere. The human-facing site gets redesigned, rewritten and fussed over every time the business evolves. The machine-facing one is whatever a plugin generated by default years ago, still describing a service that's been discontinued, a colleague who left, an address from two moves back. Or it says nearly nothing at all. Nobody looks at it, so nobody notices. Humans read this year's business. Machines read a stale one, or a blank.

What's changed is who's doing the reading. Fabrice Canel, a principal product manager on Microsoft's Bing team, confirmed on stage last year that schema markup helps Microsoft's large language models understand content. A structured data engineer at Google said around the same time that many of Google's systems simply run better with it. The other assistants are quieter about their internals, but they're all drawing on search indexes that have rewarded machine-readable clarity for a decade.

The stakes are different now, though. The machine used to decide where to rank you. Increasingly, it decides how to describe you. I'd want the description working from a label I wrote, not a guess it made.

One honest caveat, because this corner of the industry attracts snake oil: markup alone gets nobody recommended, and anyone who promises you a block of code will get your business named by AI should worry you. Google's own guidance says its AI features need no special markup at all, just the same clear, crawlable, corroborated site the rest of this post is about. Structured data is one layer of being legible. It's simply the layer you control completely.

E-E-A-T, which matters more than its name deserves

Picture two pages giving the same advice on the same subject. One is anonymous: no author, no history, no trace of who stands behind it. The other is written by a named person whose qualifications turn up where you'd expect qualifications to turn up, on a site run by an organisation whose details match everywhere you look. Now imagine you're an answer engine, about to repeat one of them to a stranger with your own name on the reply. It isn't much of a decision.

Google has a name for what separates those two pages. Its Search Quality Rater Guidelines, a long public document anyone can read, call it E-E-A-T: Experience, Expertise, Authoritativeness and Trust. It's how Google describes the qualities it wants its results to have, especially anywhere the answer could affect someone's money or health. The thing to understand about it is that it isn't a score, a setting or a tag. There's nothing to install. It's the accumulation of checkable facts: a named human behind the advice, who verifiably exists beyond that one page, with credentials that hold up somewhere that isn't their own website, at an organisation that is the same organisation, with the same details, everywhere it appears. Boring questions. Answerable questions. That's the point.

The part I find genuinely interesting: an answer engine has a sharper appetite for this than a search engine ever did, because it isn't ranking you, it's repeating you. Its own credibility rides on the sources it chooses, so it behaves the way a careful researcher on a deadline behaves: it corroborates, imperfectly and in a hurry, and it favours the sources where corroborating is easy. These systems still get things wrong, and confidently. That's an argument for leaving them less room to guess about you, not a reason to ignore them. A page of confident claims with no named author, no verifiable organisation and no footprint anywhere else isn't offensive to a machine. It's just noise, and noise doesn't get quoted.

This is where the two halves of this post meet, because that second website is how you hand the evidence over. E-E-A-T is the case you're making; structured data is the filing system it arrives in. Done properly, the author of a page isn't a name in italics, it's a declared person, connected to their qualifications, their organisation, and the places their identity can be verified independently. The machine doesn't have to play detective across your site and half the internet. It follows the connections you laid down and finds the story holds. Corroboration stops being work, and the sources that are easy to corroborate are the sources that get repeated.

Most small business websites fail these checks in completely mundane ways. Not through dishonesty. Through vagueness. The expertise is real; it's just written down nowhere a machine could check, and connected to nothing.

Fewer visitors, better ones

Now the uncomfortable bit, because there's no point pretending it away. Answer engines absorb clicks. A person who gets a full answer in the chat often has no reason to visit anyone, and plenty of businesses have watched their traffic graphs sag over the past two years and felt something close to dread.

But watch what happens to the clicks that survive. Adobe's analytics arm has been tracking visitors who arrive at retail sites from AI assistants, across a very large volume of visits, and the turnaround in that data is remarkable. In March 2025, those visitors converted noticeably worse than everyone else. Twelve months later they converted 42% better, and were spending more per visit and staying longer once they arrived. Shopify, working from its own commerce data, reported this spring that AI-referred sessions convert at roughly 50% higher rates than organic search on product pages, across almost every category they measured.

The mechanism isn't mysterious. Think about what that visitor did before landing on you. They asked a specific question. They compared options inside the conversation. They were told about you, by name, with context, by a tool they trust. By the time they click, the decision is half made. The old search click was a coin toss walking in the door; this one arrives with a reason.

I've seen the same thing up close. On one ecommerce project I work on, visitors arriving from AI assistants are converting at a significantly higher rate than the site's other channels. It's still a trickle next to Google, and a good chunk of it hides in analytics filed under "direct" because of how the assistant apps pass people along. But the trickle behaves differently, noticeably enough that we watch it now. Fewer visitors, better visitors is not the trade anyone asked for. It's still a trade worth being on the right side of.

Where this goes next

The direction of travel seems fairly clear to me. Assistants are becoming things that act rather than just answer: shortlist three options, check the details, fill in the form, book it. Every step of that is a machine reading websites on someone's behalf and deciding which ones it can work with and vouch for.

Which sharpens the two-website point considerably. The sites that state plainly what they are, who's behind them, and why that person should be believed, in ways both people and machines can verify, will get named, quoted and used. The vague ones won't be penalised. They'll just be skipped, quietly, and nobody sends you a letter about it.

None of this calls for tricks, and the tricks don't work anyway; you can't charm a corroboration engine. What it calls for is making true things checkable. Who you are. What you actually do. Who wrote this, and why they're qualified to. Structure it so a machine can read it without guessing, back it up somewhere beyond your own pages, and keep it consistent. Unglamorous work. Most of the durable kind is.

There's a simple way to find out where you stand, and it costs nothing. Open whichever AI assistant you use, but do it in a private window, or signed out, so it isn't being polite because it knows you. Then don't ask about your business. Ask the way a customer would: who should I use for this service, in this area, what are the good options, who do people rate. See whether you come up at all. That's the real test, because being described wrongly is annoying, but not being mentioned is invisible, and invisible is the expensive one. If you do appear, then ask about your business by name and see how well the description matches reality. One conversation is a glance, not a measurement; the answers shift with wording, location and the day you ask. But as a first look at where you stand, it's bracing.

It's a strange feeling the first time, reading a stranger's summary of your life's work. Stranger still when the summary has a gap where you should be.

If you don't like the answer, that's fixable.