Chapter 9

The new chooser

In October 2025, I asked an AI assistant a simple question, the kind a buyer would ask: "Who are the top digital agencies in Brisbane, Australia?"

It came back with about ten names. I'd heard of most of them. Evolut wasn't one. Not in the list, not in a footnote, nowhere.

I laughed. Then it stopped being quite so funny. I felt we did better work than most of the agencies on that list. But they were clearly doing a better job of being found.

It wasn't for lack of content. Our website is far denser than most agencies' sites, and it's full of case studies. The problem was that almost all of it was us talking about us. There was little or nothing outside our own site to back any of it up, and the machine had no one else's word to go on.

Around then I decided I was going to master this new game of getting brands into AI answers.

For the first time, something other than a person is reading your market. It reads far more of it than any buyer could, and it will tell you what it thinks. It isn't a careful referee, though.

The newest adviser

Buyers have always leaned on advisers: the retailer who decides what goes on the shelf, the broker who recommends a lender, the friend who knows cars. An AI assistant is the newest. It isn't yet the buyer, but it's moving that way. Today people use it to compare options, then decide for themselves. It already writes much of the shortlist, and that's most of the choice. Next it will place the order, starting with the repeat purchase and the routine reorder.

In a 2026 Ipsos study across 15 markets, 27% of consumers who knew of AI tools used them for product research. Only 9% would let them buy without asking (Ipsos). That's an attitude, not observed behaviour.

Use is growing fast. BCG found nearly a third of consumers now use AI somewhere in a purchase, about triple the rate of 18 months earlier (BCG).

You can ask it

No marketer could ever interview the retailer, the broker and the friend about every buying moment in the market. The machine will answer all day.

Ask it the questions your buyers ask, many times over, because one answer is one draw. Where it prefers a rival it will often say why, and the why is a list of work to do.

How much it matters so far

The effect is strongest where the purchase is complex. The best evidence is peer-reviewed. Kaiser and Schulze analysed 973 online stores, with more than US$20 billion in combined revenue, from August 2024 to July 2025 (Marketing Science).

Sites in complex categories drew 4.6 times the share of ChatGPT referrals that simple categories did. On those sites, visitors sent by ChatGPT converted better than five traditional channels, organic search and email among them. Across all 973 sites, AI referrals were under 0.2% of visits, and overall they converted below every traditional channel except paid social.

It is growing and improving quickly. Adobe's retail data, reported by Decrypt, shows AI-sent visitors converting 38% worse than others in March 2025. A year later they converted 42% better (Decrypt). By May 2026 the gap was 54% (Digital Commerce 360).

Similarweb counted 770.7 million visits a month sent by AI assistants worldwide in the year to May 2026, up 117% on the year before (Similarweb). It's still a small share of all visits, and that's the point: most of the choosing now happens inside the answer, before anyone clicks.

The shortlist moves to the machine

The list is written early, and AI has started to write it. The machine leans on its own memory. In one preprint experiment, AI models picked the famous brand on a tie, but dropped it when a rival offered better specifications.

So AI doesn't level the field for challengers. Where information is thin, the known brand wins the tie. A challenger wins by giving the machine better evidence, not more content. Tally, a ten-person maker of online forms, says AI assistants now bring it a large share of its new users.

There is no page one

When volunteers asked AI assistants the same request for brand recommendations, they almost never got the same list twice. So there's no position to hold, and a tool that reports your rank in an AI answer is mostly reporting noise.

What did hold steady was how often each brand appeared across many runs. Selection by a machine is a rate, not a rank. Chapter 14 shows how to measure it.

Named is not recommended

An AI can name five brands and recommend one. Counting how often you're named is visibility. Counting how often you're the one recommended, and why, is Selectability.

The gap is wide: in one adviser's sample of 31,200 answers, the brands it tracked were the first choice in under a third. And it changes what people do. In one 2026 preprint, a recommendation moved people to search for a brand two to three times more than a neutral mention (Iannelli and Ai).

The seven stages apply to the machine as they do to a person: does it know you, find you, believe you and prefer you?

The buyer checks the machine

The machine advises and the buyer audits, harder when more is at stake. In a 2026 TrustRadius survey of 1,862 technology buyers, 63% used AI in their research. And 94% fact-checked what it told them at least some of the time (TrustRadius).

So in a considered purchase, the machine's answer starts the checking, often on your own website. What the buyer finds there has to match what the machine said.

Two memories

Marketing works by leaving a memory. There are now two memories to leave: one in the buyer, one in the machine. The buyer's is built by experience and advertising. The machine's is built from what has been written about you, much of it by other people.

Studies of what AI cites disagree widely on how much. My reading is that your own site supplies the facts, and others supply the verdict.

The machine can be fooled

In March 2026 China's state broadcaster showed how easily. An insider used software bought openly online to invent a smartwatch, with fake reviews and experts who didn't exist. Soon a popular chatbot was recommending it (China Media Project).

Nobody yet knows for how long that will work. Two things still favour real evidence. Buyers check the machine when the stakes are high, and the platforms now treat manufactured evidence as spam.

The machine reads what you have published, and what's said about you, in one sitting. So what should it find?

Whose evidence this is. Ipsos sells market research. BCG sells consulting. Adobe sells analytics and commerce software. TrustRadius runs a software review site. Similarweb sells web analytics. Peer-reviewed source: Kaiser and Schulze, in Marketing Science. Iannelli and Ai is a preprint.