The AI concierge is coming to retail. What it means for your store.

Two shifts are happening at once, and stores keep confusing them. The first is outside your walls. Shoppers now ask an assistant where to buy something, and the assistant names a store. The second is inside your walls. Your own site can answer a question at eleven at night, in your voice, on your policies. Same technology underneath. Different work entirely. Run them as one project and you'll do both badly.

Start outside. A shopper opens an assistant and asks for a wool coat under three hundred, in Hamilton, this week. Back comes a short answer with a few stores in it. You're in that answer or you're not. There's no page two to fight your way onto.

That's the part owners underrate. Search gave you gradations. Fourth place still got clicks, and a sharp snippet could pull attention away from the top. An answer behaves more like a referral from a neighbour. It's short and confident, and it doesn't name everyone in town.

None of this creates new shoppers. The same person who would have opened a search box opens an assistant instead. Same demand, new doorway.

Machines can only recommend what they can read.

You can run the diagnosis yourself tonight. Start with your own store name. Search it and read every set of hours that comes back at you. The Business Profile, the directory a former manager filled in, the listing from two addresses ago. Any of them can be the one a machine believes.

Then open one of your own product pages and read it cold. Count the facts a machine could lift straight off it. Price. What's actually in stock. The size range you carry. Which store has it on the floor today.

If the count comes back thin, the page is selling a mood. A machine can't quote a mood. It can quote a line that admits the boots run narrow through the instep. It can quote a delivery radius and a return window. Structured data carries facts like those in a format machines already expect, which is why the dull markup job keeps turning up first. And accuracy is the part that bites, because you can't correct a machine after it has already quoted you.

Your site should answer when your staff can't.

Now the half you fully control. On a Saturday afternoon your staff are on the floor and the chat window sits there blinking. At eleven at night nobody's even in the building. Both of those hours are busy on your website. A concierge is the counter that stays open through them.

The job description is short. It works from your own catalogue and your written policies, never from general knowledge about coats. Fit is the question that stops the sale. It gets asked in the fitting room, and it gets asked at midnight by someone who would rather not stand in one. When the question outgrows the concierge, a person takes the thread over.

Logging matters more than it sounds. Every conversation is a record of what shoppers wanted and couldn't get. Three people asked about wide widths on Tuesday. You don't carry them. That's a buying signal your point of sale will never show you, because the sale never happened.

Stock, policy and price come out of your systems, nowhere else. Invention is the expensive failure. A shopper takes the afternoon off to collect a rug the bot said was waiting at the back counter. It isn't. That story reaches the neighbourhood before your apology does.

Which is why most of the build is plumbing rather than personality. The concierge has to read the same inventory the till reads. It has to know the delivery van covers Guelph on Thursdays. It has to say I don't know out loud and pass the question along, without dressing the gap up as an answer. Feeds break on a Tuesday and get noticed on a Friday. If the concierge is reading that feed, it spent three days confidently telling people the wrong thing.

The work that ranks you is the work that gets you recommended.

The desk takes a position here, and plenty of people will argue with it. Assistants are reading the same web everyone else reads. Your pages, your profile, your reviews, the forum thread where somebody asked which store in town carries wide sizes. A store that was already clear and well structured tends to be the store that gets named.

So the sequencing is dull and it matters. Structure and copy come before measurement. Until a machine can read your store, there's nothing to read back. That work lives on your product pages and in your feed. If it's been sitting half-done, it was sitting there before assistants entered the conversation. Findability isn't a separate product bolted on beside that job. If someone quotes it on its own, ask what the work actually is.

One more thing, said plainly. Nobody can promise you a place inside an assistant's answer. There's no line item for it and no button that buys it. The systems keep changing, and so do the sources they lean on. What reads well in spring may read differently by fall. If someone guarantees you a spot, they are guessing with a straight face.

Keep your own file on it. Type your shoppers' questions into an assistant and save the answers with the date on them. Fix whatever the answer got wrong at the source. Try again in a few weeks. That file belongs to you whether or not you ever hire anybody, and it settles arguments opinions can't.

Three questions is enough to start with. Bring what comes back and we'll work from where the answers went wrong.