What an AI shopping assistant does and how to choose one for your store

It's 10pm and someone is on your product page looking at a rain jacket. They like the color. They want to know if it's waterproof or just water-resistant, whether it runs small, and whether it'll fit over a fleece. Your team logged off hours ago, the product description doesn't say, and the reviews disagree with each other, so they open another tab and there's a decent chance they don't come back.

An AI shopping assistant is built for that moment. It answers the questions shoppers have while they're deciding, recommends products that match what they've told it, and gets out of the way when they're ready to check out. For a small ecommerce team, it's the closest thing to having your best salesperson on every product page at the same time.

This guide covers how AI shopping assistants work, where they help most, what they need from you before launch, how to evaluate one, and how to tell whether yours is paying off. It's written for brands where the CX team might be one to five people who answer chat, work the inbox, and process returns in the same afternoon.

What an AI shopping assistant is

An AI shopping assistant is software on your site, usually living in the chat window, that uses AI to help shoppers find and choose products. It reads your catalog, understands questions written the way people talk, and answers in your brand's voice. The better ones also know who they're talking to.

What it does on your site

Most of the work falls into four jobs:

• Product questions. "Is this waterproof?" "Does the 32 inseam come in a tall?" "Can I put this in the dryer?" These are the small questions that stall a purchase.

• Guided discovery. A shopper describes a need and the assistant narrows your catalog down to a few good options.

• Recommendations. The assistant suggests a matching item or an upgrade when it fits the conversation.

• Cart and order help. Shipping thresholds, return windows, promo codes, and "where's my order" for people who already bought.

How AI shopping assistants work

Under the hood, it comes down to four parts.

Your product data. The assistant reads product pages, specs, size charts, and availability from your site or catalog. If that data is wrong or stale, the answers will be too.

Customer history. An assistant connected to your customer service platform and your ecommerce platform can see past orders and earlier conversations. That's the difference between a generic answer and "You ordered the 8.5 last time, and this style runs about a half size small."

The language model. This is what lets the assistant understand a messy, half-typed question and write a natural reply. A language model on its own will fill gaps with confident guesses, so the assistant has to be grounded in your catalog and your policies to stay accurate.

Guardrails and handoff. These are the rules about what the assistant can say and do, plus a path to a person when a question falls outside those rules or the shopper asks for one.

AI shopping assistant vs. chatbot vs. site search

These three get lumped together, and they solve different problems.

A traditional chatbot follows a script. It's good at routing and at a fixed list of FAQs, and it stalls when someone asks something the script didn't plan for. Site search matches keywords to product listings, so it works well when the shopper already knows what they want and poorly when they're describing a problem, like "something for a wet spring in Portland."

An AI shopping assistant understands the question the way a person on your team would, searches your catalog on the shopper's behalf, and answers in a conversation. Because it's a conversation, it can ask a follow-up like "Are you commuting or hiking?" before it recommends anything.

Gladly Agentic Commerce works as an AI shopping assistant in this sense. It answers product questions from the latest information on your site, suggests matching items in the same conversation, and can reach out first when a shopper looks hesitant.

Where assistants help most in the buying journey

Discovery

"I need a jacket for a wet spring in Portland" is a hard query for site search and an easy one for anyone who's worked a sales floor. An assistant can ask one or two questions about how the shopper will use it, check what's available, and come back with a short list at a couple of price points. The shopper describes what they need in their own words and only sees the part of your catalog that fits.

Product questions and fit

In apparel and footwear, fit and performance questions come up constantly. "Is it waterproof?" has a real answer, and a shopper who gets it in five seconds is in a very different place than one who has to dig through 40 reviews to find it.

Fit is the big one. A good assistant can pull from your size chart, the product's fit notes, and what you've written about how a style runs, then give a direct answer. When it can't answer confidently, it should say so and offer a person, which beats a wrong size and a return label.

Cost questions belong here too. Baymard Institute puts the average cart abandonment rate at 70.22% across 50 studies, and extra costs like shipping and tax are the top reason shoppers give for leaving. An assistant that can answer "how much is shipping to Canada?" or "how far am I from free shipping?" before checkout removes a surprise that would otherwise show up on the last page.

Complete-the-look and add-ons

Recommendations work best after the shopper has decided on the main item. Someone who just chose a pair of hiking pants is open to hearing about a shirt that pairs well or a belt that fits the loops. Someone still comparing two pants styles isn't.

KÜHL, the Salt Lake City outdoor brand, saw this on their own team. Nancy Orgill, Customer Support Manager, described how the conversations changed once AI took over the repetitive questions: "Now I hear them confidently making recommendations. When someone calls about pants, they naturally transition to discussing our Renegade shirts and other complementary items."

An AI shopping assistant can make the same kind of suggestion on chat, as long as it's grounded in what goes together in your catalog. If you're building out your approach, our guides on product bundling strategy and how to increase average order value go deeper on what to pair and when.

Returning customers

Many AI shopping assistants treat every visitor like a stranger. That's fine for a first-time shopper and a missed opportunity for the person who's bought from you four times.

An assistant that knows a shopper's order history can skip questions it already has answers to. If someone bought your trail runners in a 9 and wrote in last spring that the toe box felt tight, the assistant can suggest the wide version of the new model before they ask. If they returned a jacket because the sleeves were short, it can point them to the tall fit. This is the part of personalization shoppers notice, because it sounds like a store that remembers them.

It also matters after the sale. The same conversation history that makes recommendations better helps with exchanges, order questions, and returns. We cover that side in our guide to conversational AI for ecommerce.

What results look like

A useful number here is the resolution rate on product help and recommendation conversations, meaning the share of those conversations the AI answers fully without handing off to a person.

Smith Optics, the eyewear and goggle brand, reached a 67% resolution rate on product help and recommendation conversations with Gladly Agentic Commerce.

Tecovas, the Western boot and apparel brand, resolves 55% of product help and recommendation questions with Gladly.

The revenue side often shows up on your team as well as in the AI. At KÜHL, AI resolves routine questions across chat, email, SMS, and voice, which gave the team time to help customers find the right fit and suggest the right accessories. KÜHL saw a 120% increase in revenue per call. Rothy's describes the same pattern on chat.

The AI answers what it can, and your team spends more of the day on the conversations where a person makes the sale.

What an assistant needs from you

An AI shopping assistant is only as good as what you give it. Most of the setup work is content, and a small team can do all of it.

Clean product data and a current catalog

Start with the pages the assistant will read. Product descriptions should state materials, care, dimensions, and fit in plain words. Size charts should be accurate and easy to find. Discontinued items should be marked as discontinued. If a spec only lives in a PDF or a product image, the assistant probably can't see it.

The same cleanup helps with AI agents shopping on behalf of consumers, which we cover in our guide to preparing your product feed for AI agents.

Policies and brand voice written down

Your assistant needs your return window, exchange rules, shipping thresholds, warranty terms, and promo rules in writing. It also needs to know how you talk. If your brand says "y'all" and never says "valued customer," write that down.

With Gladly, your CX team writes these instructions as Guides in plain language, then updates them after launch without a developer. KÜHL's AI, for example, is trained to match the company's tone and generates answers from KÜHL's own knowledge base.

A handoff path to a person

Decide what happens when the assistant can't answer or the shopper asks for a human. Who picks up during business hours, and what does the shopper see at 10pm? The handoff should carry the full conversation, so nobody has to ask the shopper to repeat themselves.

Rothy's Head of CX, Lauren Inman-Semerau, put it this way: "Because we're on Gladly and can read everything that has happened, my agents are caught up when they enter the chat."

How to evaluate an AI shopping assistant

Every AI-powered shopping assistant looks good in a vendor demo. The questions below help you see how an assistant will do on your store, with your products and your team.

Accuracy and grounding in your catalog

Bring your 20 hardest real questions from the last month of chat and email. Include fit questions, compatibility questions, and at least one question about a product you've discontinued. Ask the vendor to run them against your catalog, and watch for answers that sound confident but aren't on your site.

Integrations with your ecommerce platform, reviews, and loyalty tools

The assistant should connect to your ecommerce platform for orders and products, and ideally to your reviews and loyalty tools, so it can reference what other customers said or what points a shopper has. Gladly connects with Shopify and BigCommerce, plus tools like Klaviyo, Yotpo, and Okendo through App Platform. With Gladly, these are pre-built connectors your CX team configures without engineers. When you talk to other vendors, ask which integrations are ready to use today and which need a developer for the initial install, since the answer varies a lot.

Who controls it day to day

Once it's live, someone has to update the assistant when you launch a product, change a policy, or notice a bad answer. If every change needs an engineering ticket, updates will lag. Look for an assistant your CX team can edit directly.

If you're a small team

When the same two people handle chat, email, and returns, keep the requirements short.

Require:

• Answers grounded in your catalog and policies

• Plain-language editing your team can do without a developer

• A handoff to a person that keeps the full conversation

• Basic reporting on conversations, resolutions, and handoffs

Skip for now:

• Launching on every channel at once. Start with chat on your highest-traffic product pages.

• Complex proactive campaigns. One well-timed prompt on a product page is plenty to start.

• Custom model work. You want a tool that learns from your content, and your content is the part you control.

Questions to ask on a demo

1. Where does the assistant get product information, and how quickly does it pick up a change on my site?

2. Can it see a returning customer's past orders and conversations?

3. What does it do when it doesn't know an answer?

4. What happens during a handoff, and what does my team see?

5. Who on my team can change what it says, and how?

6. Which of my current tools does it connect to, and what does setup require?

7. How do I measure its effect on conversion and order value?

8. How is it priced, and what counts as a billable conversation or resolution?

How to measure whether it's working

Set a baseline before launch

Pull at least 30 days of numbers before you turn the assistant on: site conversion rate, conversion rate for shoppers who used chat, average order value, chat volume, and your most common pre-purchase questions. Without a baseline, you're guessing.

If you can, hold part of your traffic or a set of product pages out of the launch for the first few weeks. Comparing shoppers who saw the assistant against similar shoppers who didn't gives you a much cleaner read than a before-and-after chart, which picks up seasonality and promotions along with the assistant.

The numbers to track

• Conversion rate for shoppers who chat, compared against your baseline and your holdout

• Average order value on assisted orders, meaning orders that followed a conversation

• Resolution rate on product questions, the share the assistant answers without a handoff

• Handoff rate, and what those handoffs have in common

A rising handoff rate on one product usually points to missing content on that page. That's useful, because it tells you what to write next.

Attribution with GA or Adobe events

Gladly can send chat events to Google Analytics or Adobe Analytics, including when a conversation starts and ends and when a proactive campaign fires. That lets you connect conversations to purchases in the analytics tool you already use and attribute revenue to AI.

To turn those numbers into a business case, try the CX ROI calculator.

Where assistants go wrong

Content gaps get mistaken for AI mistakes

When an assistant gives a bad answer, the first reaction is usually to blame the AI. Look at the source first. Often the product page never said whether the jacket was waterproof, the return policy on the site was out of date, or two pages said different things. Fixing the content fixes the answer, and it helps the shoppers who never open chat too.

A simple weekly habit works well for small teams. Pull the conversations that ended in a handoff or a thumbs-down, sort them by product, and fix the page behind the most common one.

Recommending too much, too early

An assistant that suggests three add-ons before the shopper has picked the main item feels like a pushy salesperson. Set it up to recommend after the shopper has decided and to back off when they ignore a suggestion. One suggestion at a time is usually enough. The KÜHL example works because the shirt suggestion comes after the pants are settled.

On-site assistants vs. consumer shopping agents

There's a second kind of AI shopping help, and it doesn't live on your site. Tools like ChatGPT research and compare products for the shopper. Adobe found that traffic from AI sources to US retail sites grew 393% year over year in the first quarter of 2026, and that AI-referred traffic converted 42% better than non-AI channels in March 2026.

Those AI shopping agents work for the shopper. Your on-site AI shopping assistant works for you. A shopper might discover you through an agent, land on your site with a follow-up question, and ask your assistant. Clean product data serves both. For the bigger picture, see our take on agentic commerce.

Getting started

Pick the 20 product pages that drive the most revenue, clean up their content, write down your policies and your voice, and decide who takes the handoffs. Set your baseline. Then launch on chat for those pages and review the conversations every week for the first month.

If you want to see how Gladly Agentic Commerce handles product questions, recommendations, and returning customers in one conversation, take a look at our AI shopping assistant.