For twenty-five years you could check where you stood. Type your own company name into Google, run the searches your customers run, and the answer came back ranked and visible. You could see your position, watch it move, and tell whether last quarter's work had paid off. The whole discipline of SEO rested on that one underrated privilege: you could audit yourself.

Open ChatGPT and ask it to recommend a supplier in your field, and that privilege is gone. The model returns three names. You've no idea whether yours was the fourth, the fortieth, or never in contention. There's no results page to scroll and no ranking to inspect. You're either in the answer or you aren't, and most of the time you simply can't tell which.

That's the actual shift, and it's bigger than the one everyone's talking about. The conversation about answer engine optimisation tends to fixate on tactics — structure your content this way, add that schema, earn these citations. All sensible. But it skips past the genuinely uncomfortable part. The surface where people now discover you has stopped showing you your own reflection. We spent two decades learning to manage what we could measure, and the thing we relied on to measure it has been replaced by something we can't see into.

The glass box became a black box

HubSpot's recent research puts a number on how fast this has moved: 32% of B2B buyers now discover new vendors through generative AI chatbots. A third of your potential pipeline is meeting your category through a conversation you're not party to and can't replay. And the buyers who use AI this way aren't browsing idly — HubSpot found they're more likely to convert than traditional organic visitors, which means the discovery you can't see is also the discovery that's worth the most.

Search was a glass box. You could watch the mechanism work, more or less, and the feedback loop was honest: do the work, check the ranking, adjust. The answer engine is a black box. The work goes in, an answer comes out, and the relationship between the two is opaque even to the people who built the models. We used to get an early warning when something slipped — traffic dipped, a ranking fell, and we'd go and look. Now the first sign of trouble is a recommendation we never appear in, by which point the shortlist has already formed without us.

This is one half of what I've called marketing's two-front war — the front where the audience itself changes, fought on a surface you can't see into and that, in most companies, nobody actually owns.

I find that destabilising, and I think most business owners haven't felt it yet because they haven't gone looking. Try it: ask the major assistants the questions your customers would ask, and read the answers as a stranger would. It's a sobering ten minutes.

The half I left out last time

I argued a couple of months ago, in AI didn't break marketing, that the models reward substance over volume — that the businesses producing structured, authored, genuinely useful content were pulling ahead while the keyword-stuffers disappeared. I stand by all of it. But I was only describing half the job.

Being useful is what gets you eligible to be cited. It doesn't tell you whether you actually are. You can publish the most authoritative guide in your sector and have no mechanism for knowing whether ChatGPT mentions it when someone asks. The old world collapsed those two things into one — if your content was good and your SEO was sound, you ranked, and you could see that you ranked. The new world has prised them apart. Producing the work and verifying it landed are now separate problems, and almost nobody is treating the second one as a problem at all.

Three names, not ten blue links

There's a reason absence hurts more now than it used to. A search results page had ten blue links, then a second page, then a third. Even a mediocre ranking kept you somewhere in view, and a determined buyer could find you three scrolls down. The answer engine doesn't work in tens. HubSpot's figures suggest buyers now begin with an average of 7.6 potential vendors and whittle that to 3.5 before they'll talk to anyone. Increasingly that whittling happens inside the model, before a human has seen a single website.

Three or four names is the whole shortlist. Being left off it isn't the equivalent of ranking eighth; it's the equivalent of not existing. The customer can't choose to ignore you, because you were never offered as an option. This is the part that changes the stakes: the penalty for invisibility used to be lost traffic, and now it's being left out of the decision altogether.

It's the natural endpoint of something I wrote about back in 2024, in the future of search: the move from a list of links you browse to a single answer you're handed. Back then it read as a forecast. It's now just the weather.

Ads are arriving inside the answer

If you were hoping the new surface would at least settle down long enough to learn it, it won't. It's being commercialised in real time. Microsoft launched AI Max for Search in April, an open pilot that pushes ads into Copilot and Bing's AI surfaces and includes formats that drop selling points like free shipping directly into the conversation. Paid placement is arriving inside the answer, which means the organic-versus-paid line you knew from search is being redrawn in a space you can't yet inspect.

It gets stranger at the buying end. The Business of Fashion has reported that as shopping agents start acting on customers' behalf, they distort the metrics underneath — triggering ad impressions and clicks no human ever saw, and cutting off the behavioural data trail that retailers have always relied on to understand demand. So the surface is harder to see into, and the instruments you used to read it are giving false readings. That's not a stable platform to optimise against. It's a moving one, in the dark.

A practical audit

So what do you actually do? If you've spent years learning to rank on Google, some of that instinct still serves you and some of it now works against you, and telling the two apart is most of the job. Don't panic, and don't rush out to buy whatever AEO tool is loudest this week. Three things are enough to start.

Split your content by who it's really for. Go through what you publish and sort each piece into one of two buckets: built for a human to find and read, or built for a model to cite. Most businesses have a pile of the first and almost none of the second. The second is the stuff written to answer a question cleanly, attributed to a named person with real expertise, structured so a model can lift it without distortion. You need both, but you've probably only been making one.

Start measuring share of model now. This is the part people skip, because it's new and a little awkward. A category of tools has appeared specifically to answer the question search used to answer for free — which brands the models recommend, and which they don't. Mentio is one of the clearer examples: you give it your brand and competitors, it runs the questions a real buyer would ask across ChatGPT, Claude, Gemini and the rest, and it hands you a score for how often you show up against your rivals. It's imperfect and the methodology is young. It's also a great deal better than the nothing most businesses currently have. Don't get attached to the specific tool, though. What matters is that you can no longer assume your visibility — you've got to check it, the way you once checked a ranking.

Decide how much of each you need. This isn't all-or-nothing, and the right mix turns on three things: how your buyers actually find you today, how considered the purchase is, and how crowded your category is. A local firm whose customers still arrive by word of mouth, buying something simple in a market with few rivals, can afford to move slowly. A B2B software company selling a high-consideration product into a crowded field — where a third of buyers already open with a chatbot — has no slack at all, and sits at the far end of the same scale. Forget "are we doing AEO". The question that matters is what proportion of your discovery now happens somewhere you can't see, and what you'll actually do about it this quarter rather than next year.

What I'm not sure about

I don't know whether the black box stays black. It's entirely possible the platforms hand us dashboards within a year — Google and Microsoft both have every commercial reason to, since advertisers won't spend into a void forever. It's possible the measurement tools standardise into something as trusted as Search Console once was. If that happens, today's anxiety will look quaint, and the businesses that scrambled will have over-rotated.

But I'd rather act early here than wait. Measurement is cheap right now and most competitors aren't bothering, which is exactly the window where a small advantage compounds. Ten minutes and a modest subscription buys you the answer; the alternative is finding out, eighteen months from now, that you slipped off the shortlist and never knew.

The search bar used to hand you that answer for free, whether you liked it or not. It doesn't any more. The job now is to build your own way of seeing — and to be honest enough to use it when it tells you something you'd rather not hear.

Quick answers

What is answer engine optimisation, in plain terms? It's the practice of making your business discoverable inside AI assistants like ChatGPT, Claude and Gemini, rather than inside a traditional search results page. Same goal as SEO — being found — but the surface, and the way you get cited, are different.

How is it different from SEO? The mechanics overlap, but the crucial difference is visibility into your own performance. SEO gave you rankings and traffic you could measure directly. With answer engines you're recommended, or not, inside a model whose reasoning you can't inspect — so measurement is now a deliberate, separate effort rather than something the platform hands you.

Do I need to start now, or is this a next-year problem? It depends entirely on where your buyers are. If a meaningful share of your customers already research through AI — and in B2B that's now around a third — it's a today problem. The measurement is cheap and most competitors aren't doing it, which is the best possible time to start.