The Shop Assistant Is Now Something You Can Rent

On 9 September Instacart put a shopping assistant called Clementine in front of customers across the United States and Canada, and on the same morning began selling a white labelled version of the same machinery to grocers who want it running under their own name. Shipt announced one that day too. For a shop with one warehouse and one person answering email, none of that is the story.

What actually shipped, and who is it for?

Four separate things landed inside about thirty hours, from three companies, and they are not the same product wearing different names. Instacart launched Clementine, a consumer assistant that turns a request like a week of budget friendly kid lunches into a ready cart. Shipt launched Ask Shipt, which does something similar from a text prompt, an uploaded recipe or a photograph of a meal. Instacart also expanded Cart Assistant, the white label product a grocer puts on its own site. And a mid sized home goods retailer quietly reported six months of results from an assistant it had already been running since June.

Two of those four are marketplace assistants. They live on somebody else's property, they read somebody else's catalogue and they end with somebody else's checkout. If you sell through Instacart, Clementine is a new front door into a store you already stock. It is not a tool you operate. The other two are the ones worth reading closely, because they sit on a retailer's own domain and answer questions about a retailer's own products.

Why does a rug retailer matter more here than Instacart?

Because Instacart answered a question you do not have. It has 1.6 billion lifetime orders and a catalogue of two billion items to draw on, which is exactly the asset a small shop will never assemble. Rugs Direct is closer to your size and it published the part everyone leaves out: the setup.

Digital Commerce 360 reported that Elevated Brands Group spent four to six weeks training the assistant on its own material, meaning articles, blog posts and product specifications it had already written. Chief merchandising officer Sharon Gautschi described the approach as data integrity work, and said add to cart rates for shoppers who used the assistant ran eight to nine times higher than for shoppers who did not, across both of the group's retail brands. In the same breath she said the majority of customers are still using classic search.

Both halves of that sentence are load bearing. One month to six weeks of content preparation is a real cost for a small business, and it is not engineering work. It is somebody sitting down with the catalogue and answering, in writing, the questions customers keep asking. That is the same job we described in turning a support inbox into a help centre, and if you have already done it, most of the setup is behind you.

What does the eight times number actually mean?

It means less than it looks, and understanding why will save you from a bad forecast. A shopper who opens a conversational search box and types a full sentence about a hallway runner has already decided to buy a rug. A shopper who lands from a search result and bounces in nine seconds never touches the assistant at all. So the group being measured is self selected toward high intent, and the comparison group contains everyone who was only browsing.

The honest way to read it: the assistant is doing well by the people who choose to use it, and the retailer has not claimed a site wide conversion lift. Nobody in the reporting claimed one either. If you switch an assistant on and your overall conversion rate does not move in month one, that is the expected outcome, not a failure. The thing to watch instead is whether the shoppers who engage it convert better than shoppers who used your old search box on comparable queries, which is a harder measurement and the only one that answers the question.

What does an assistant on your own site need from you?

Four things, and only one of them is technical.

The first is product attributes that are actually filled in. Ask Cleo answers questions about material, thickness and dimensions on Rugs Direct, and about dimmability, ceiling height and colour temperature on the lighting brand. It can only do that because somebody put those fields in the product records. An assistant sitting on a catalogue where half the attributes are blank will either say it does not know, which annoys people, or infer, which is worse. Attribute completeness is the unglamorous prerequisite, and it is the same work that decides whether a product page reads well to an AI shopper in the first place.

The second is written answers to the questions you already get. Not marketing copy. The reply you send when someone emails asking whether the rug sheds, whether it works with underfloor heating, whether the colour in the photo is accurate under warm light. Those replies are the training material. Elevated Brands Group used its own blog posts and articles for exactly this.

The third is a place where every unanswered question lands. The value of a conversational search box is not only that it sells things. It is that it records, in your customers' own words, every gap in your catalogue and every objection you never knew about. A shop that reads that log weekly gets more out of the tool than a shop that watches the revenue line.

The fourth is a named person who fixes wrong facts. Product data drifts. A supplier changes a fabric, a size chart gets updated, a discontinued line stays live. When the assistant repeats an old fact confidently, someone has to correct the source, and that person needs to exist before launch rather than after the first complaint.

Where does the money go?

Vendors price this per session, per resolved conversation or as a platform fee bundled into search. None of the four launches published a price, so treat any figure you are quoted as the start of a negotiation rather than a market rate. What you can budget with confidence is the part nobody invoices you for: the four to six weeks of writing and data cleanup that Elevated Brands Group described. For a one person shop that is not four to six weeks of full time work, but it is not an afternoon either.

There is also a question of where the assistant sits on the page, which sounds cosmetic and is not. Ask Cleo runs in two places at Rugs Direct: as a site level search tool for someone who arrives without a specific product in mind, and as an agent on the product page itself, answering questions about the item already on screen. Those are different jobs with different economics. The site level one competes with your navigation. The product page one competes with the customer closing the tab to go and ask elsewhere, which is the more expensive loss and the easier win. If a vendor offers only one of the two, ask which, and price it accordingly.

There is a second cost arriving on a timer. Modern Retail reported that Instacart plans to bring advertising into both assistants by the end of the year, with advertising general manager Ali Miller saying the company is purposefully going slowly. Read that as a preview of the category. Once an assistant becomes a placement, the ranking inside it becomes something brands pay for, and the free organic period every new channel enjoys will close here as it closed on marketplace search. If you sell through a marketplace assistant rather than your own, plan for that.

Is this actually mainstream yet?

It is early and it is real. The same Modern Retail piece cites Gartner data putting generative AI chatbot adoption among multi brand retailers at 21% in 2025, up from 5%. Gartner analyst Greg Carlucci added the caution worth pinning above your desk: shoppers are open to using AI while they shop, and they still want to keep control of the experience.

The traffic side is smaller than the noise suggests. Rugs Direct and its sister brand get under 10% of their traffic from AI platforms including ChatGPT, Claude and Perplexity. That matches what we found when agentic shopping turned out to be under 1% of Etsy's traffic. An assistant on your own site is therefore not a bet on AI referral traffic. It is a bet on the visitors you already have, which is a much safer bet and a much smaller prize.

What should you measure in the first ninety days?

Three things, in this order, and none of them is revenue.

1. Engagement rate. What share of sessions open the assistant at all. If it sits under a few percent, the entry point is wrong, not the model. Rugs Direct's own executive said most shoppers still use classic search, and that was after a full quarter live.
2. Unanswered rate. What share of conversations end without the assistant giving a usable answer. This is your catalogue quality score, expressed in customer language, and it should fall every week that somebody works the log.
3. Correction rate. How often you have to fix a fact the assistant stated. A number that will not fall is a signal that your product data has no owner, which is a business problem the vendor cannot solve.

Only after those three settle does a conversion comparison mean anything, and it should be run against comparable queries in your old site search rather than against your whole traffic. Otherwise you are measuring who chose to talk to a robot, not what talking to one did.

Does an assistant help if you sell in more than one language?

It can, and the week produced a concrete example. Heritage Grocers Group worked with Instacart to build Spanish language capability into its version of Cart Assistant, which is the kind of customisation a white label deal is for. The catch is the same as everywhere else in translation work: an assistant that answers fluently in a second language while reading product attributes that only exist in the first will produce fluent nonsense. We went through why that failure is hard to spot in the piece on what actually blocks a sale in a multilingual store. Fluency is not accuracy, and an assistant makes the gap between them harder to see, because it never sounds unsure.

Who is liable when the assistant is wrong?

You are, in most of the places you sell. An assistant on your domain, wearing your brand, speaking to your customer, is your statement. That is not a theoretical risk. A Canadian tribunal has already held a company to a refund policy its own chatbot invented, which we covered in the piece on what happens when your chatbot makes a promise. The white label arrangement does not move that exposure. Instacart chief commercial officer Ryan Hamburger framed the deal as keeping data a competitive advantage for Instacart and its retail partners, which is a statement about data, not about who answers for a wrong answer.

Two practical guards. Keep the assistant out of anything that states policy: returns windows, delivery guarantees, price matching, warranty terms. Let it answer product questions and hand policy questions to a page you control. And log every conversation in a form you can search later, because the first time a customer says the assistant told me, you will want the transcript.

Should a small shop switch one on this quarter?

If your catalogue is small and simple, no. When a visitor can see everything you sell in two scrolls, a conversational search box solves a problem that does not exist, and you would be paying to answer questions your product page already answers.

If your catalogue is large, technical or full of attributes that decide fit, the case is stronger. Rugs, lighting, tools, parts, ingredients, sizes, anything where the customer question is which of these is right for me rather than do you have this. That is where the eight to nine times figure is coming from, even discounted for selection.

The order of work matters more than the vendor choice. Fill in the attributes first. Write the answers second. Only then shop for the tool, because every vendor demo will look excellent on a clean catalogue and none of them will fix a dirty one for you. If you want to see what a shop looks like when the product data is structured from the start, our AI store builder generates the catalogue schema and the product pages together, which removes the first of those two jobs.

One more thing worth saying plainly. Instacart's own assistant reports an average order value above $115 for the customers who use it, according to the Digital Commerce 360 account of the launch, and that figure is drawn from grocery baskets built out of a two billion item catalogue with ten million daily inventory signals behind it. It is a fact about Instacart. It is not a target, a benchmark or a promise, and any vendor who quotes it at you as one is selling rather than explaining.

The pattern underneath all four launches is the same and it is older than any of them. The assistant is a better front end onto whatever you already know about your products. Where a business has written that knowledge down, the tool has something to work with. Where it has not, the tool has a catalogue of blank fields and a confident voice, which is the worst combination on a commerce site. Instacart spent nearly fifteen years accumulating the knowledge it is now renting out. The equivalent asset in your shop is the email you have been answering for six years, and it is already written.

Start there. The vendor can wait a month.