There is no single shopping agent algorithm to optimize for. Researchers who ran frontier models through randomized shopping experiments found that different models pick different winners, and a model update can reverse what the previous version preferred.
On episode 407 of the podcast, Tanner Larsson, CEO of Built to Scale, described his wife redecorating a fireplace mantel. She uploaded a photo to ChatGPT, described the look she wanted, rejected a few suggestions, asked where to buy the rest, and ordered all of it. “But a lot of our shopping in our household comes from it,” he said.
In that story, an assistant did the finding and comparing, and a person approved the result. That is what an AI shopping agent is, and it is already how some of your customers shop. The question for a Shopify brand is narrow: when an agent builds that shortlist, what makes it pick your product?
This page answers that from the shopper’s side. The merchant setup, the checkout decision and the protocols sit with their own guides, linked where they come up.
An AI shopping agent does three jobs for a shopper: it researches and shortlists products against a stated need, it watches for a condition such as a price drop, and on some platforms it completes the purchase. Discovery is the job every agent does today and the one that decides whether your product is ever considered.
The five agents most of your customers can reach read different sources and finish the sale in different places. That difference matters more than any feature list, because it tells you where an agent learns about your product. Here is the state of each as of November 17, 2026.
A few dated facts sit behind that grid. OpenAI refocused ChatGPT shopping on discovery on March 24, 2026, letting merchants use their own checkout. On November 13, 2026 it added ChatGPT Wallet, which lets the always-on agents it calls dots pay on a merchant’s site with a stand-in card number, starting with US Pro subscribers. Shopify’s agentic storefronts are on by default for eligible stores and feed ChatGPT, Google AI Mode and Gemini, Copilot and Meta. Google says its Shopping Graph held more than 50 billion listings in November 2025. Amazon replaced Rufus with Alexa for Shopping on May 13, 2026, and its Buy for Me feature can complete eligible purchases on other retailers’ sites. Perplexity added Instant Buy with PayPal for US users in November 2025, and in August a federal appeals court vacated the injunction against its Comet browser shopping on Amazon, while the underlying case continues.
Where the money changes hands is its own subject, covered in our Agentic Payments and Checkout hub. For what to switch on in your admin, see the guide to agentic commerce for Shopify.
AI shopping agents compare products by checking each option against the constraints in the shopper’s request, and controlled tests show price, ratings, reviews, endorsements and on-screen position all sway the choice, weighted differently by each model. The most useful controlled evidence I have found comes from a study that built a mock marketplace and watched frontier models shop in it.
In “What Is Your AI Agent Buying?”, revised in December 2025, Amine Allouah, Omar Besbes, Josué Figueroa, Yash Kanoria and Akshit Kumar showed models a results page of eight products in two rows and four columns, with no product detail pages, and asked them to pick one. They randomized positions, prices, ratings, reviews, sponsored tags and platform endorsements across six models from August and December 2025.
Position mattered, but not in one direction. The three August models favored the top row, while their December successor GPT-5.1 leaned the other way; the authors note that GPT-4.1’s preferred slot was GPT-5.1’s least preferred. Visible sponsored tags were penalized and platform endorsements such as an overall pick were rewarded. Sensitivity to price, ratings and reviews varied sharply between models. And a model update could drastically reshuffle which products won.
Keep the study’s limits in view. It tests the final choice among products already on screen, not how ChatGPT, Google or Amazon decide which products to retrieve in the first place, and its “market share” is simulated selection frequency, not retail sales. It is evidence about how agents behave, not an audit of any live service.
That is still enough to retire one idea. An earlier version of this page ranked five selection signals by weight. The research does not support a fixed ranking, which is why that list is gone. A checklist tuned to how one model behaved this spring can be wrong after the next update, and you will not get a changelog telling you.
Agents can recommend few products because, in repeated controlled choices, models tended to concentrate demand on a handful of “modal” options and ignore the rest. That cuts against the hopeful view that AI shopping levels the field for small brands, and I think the truth sits between the two.
Three different things get blurred here. How deep an agent searches, how many products it shows the shopper, and how often repeated choices land on the same few products are separate questions, and the study only measured the third. On episode 478 of the podcast, Morten Poulsen, founder and CEO of the product information platform Plytix, spoke to the first: “an AI doesn’t care about going to page seven. It just cares about giving you the best product.” That is his read of how agents search, and it matches what shoppers see when an assistant surfaces a brand they had never heard of.
Searching deep is not the same as recommending widely, though. I’d argue the two observations fit together: an agent may well find you on page seven, but on a broad request it has many well documented options to settle on. My working hypothesis, not a proven result, is that a smaller brand’s best odds are narrow, specific requests (“waterproof trail runner for wide feet under $140”) where its product is plainly the right fit, rather than broad ones (“best running shoes”).
That points the work at specificity. Pick the five or six use cases where your product is genuinely the best fit, and make sure every place an agent reads says so in plain words.
A brand gets picked by an AI shopping agent first by being eligible and retrievable, and then by giving the agent facts that answer the shopper’s constraints, say what the product is best for, and agree with each other. You cannot control a model’s quirks, but you can control those facts, which is why they are the place to start.
Eligibility comes first and is mostly settings. Shopify’s agentic storefronts only carry products that qualify for Shopify Catalog, from stores with their policies completed, and each channel has its own rules. Check that before rewriting anything.
Once a product is in the candidate set, the facts do the work. Answering constraints means the numbers a shopper filters on are written as text: size, material, compatibility, capacity, temperature range, what is in the box. Saying what it is best for means naming the situation and the person. Rob Langenback, President and COO of the search agency Eight Oh Two, told me on episode 480 that his team writes “this product is best for” lines into product copy, adding, “I think that’s really important because that’s what the tools are also looking for when they’re making the recommendation.”
Agreement is good practice for a simpler reason: an agent may read your product page, your feed or a retailer listing, and if they disagree on a weight or a size, you cannot predict which one it repeats. One source of truth, synced to the feed Shopify sends out, removes that risk.
The study adds a finding worth knowing. A seller-side agent that rewrote descriptions to suit the shopper’s query raised the product’s simulated selection share significantly in 33 percent of experiments after one round, and on average for five of the six buyer models, with Gemini 3 Pro Preview the exception. The tested intervention was an optimizing agent, not ordinary copywriting, so my practical reading is modest: copy that answers the specific question gives an agent more to work with, and sometimes that measurably helps. For the field level checklist, our guide to the nine facts most Shopify product pages leave out covers it, and being cited across the wider web belongs to generative engine optimization for Shopify.
Two cautions. The penalty the study found applied to a visible sponsored tag on a listing; it says nothing about whether buying distribution inside a live assistant pays, so treat any paid placement as its own test with its own measurement. And do not chase model behavior with monthly rewrites of your whole catalog. Get the facts right once and keep them right.
A product that meets every requirement in a shopper’s request can still miss an agent’s shortlist when the facts that prove it are scattered across images, PDFs and banners instead of the listing. This is a labelled composite, built from patterns that come up when I talk with Shopify merchants and agency partners, not one real store.
Picture a cookware brand doing about $45K a month with 60 products. Its best seller is a $72 ten inch ceramic nonstick skillet called “The Everyday Pan,” with 140 reviews averaging 4.6 stars. A shopper asks an agent for “a 10 inch nonstick pan that works on induction, PTFE free, oven safe to 500 degrees Fahrenheit, under $80.”
The listing confirms size, nonstick coating and price. The other three requirements are true but hard to find: induction compatibility appears only in an image inside an FAQ accordion, the oven rating lives in a care guide PDF, and “PTFE free” sits in a homepage banner. The description is two lines about Sunday breakfasts, and most reviews mention gas stoves. An agent working from the listing and feed may not be able to confirm three of the six requirements, while a competitor that states all six plainly gives it nothing to guess about.
What I would change first: rename the product so the title says what it is (“Everyday 10 inch ceramic nonstick skillet, induction ready”), move the missing facts into the first lines of the description and into product metafields, and add a “best for” line naming the cook it suits. Then add one honest limit, such as the hand wash care the coating needs, because a page that admits a trade off is easier to trust. Finally, add a post-purchase review prompt asking which stovetop the buyer uses, so future reviews carry the facts too.
What I would remove: the plan to buy an AI copywriting app to regenerate all 60 descriptions. The problem here is a handful of missing facts on the products that earn the money, not 60 thin paragraphs, and a bulk rewrite will not supply facts nobody has written down. I would also drop any idea of an agent-only landing page, since agents read the catalog and the listing. My estimate is a couple of afternoons on the top ten products with tools the store already has, though a store with a locked theme or an agency in the loop will take longer.
You can get a directional read on whether agents are choosing you by running the same specific shopping requests through two or three agents on a regular schedule and recording which products they name. Answers vary from run to run and shift with model updates, so the pattern across several runs tells you more than any single answer.
Write five requests the way a customer with a real constraint would type them, run each two or three times in a private window, and note the date, the agent, whether you appear, how you are described and who appears instead. Monthly is a reasonable starting cadence for a small team, not a threshold from any study. Our 30 minute AI search visibility check lays out the full routine. If an agent describes your product wrongly, check what it cited, whether your feed is current and whether the fact is published anywhere it can read before you assume a cause. If it names a competitor, read that listing and see which of your shopper’s constraints it states that yours does not.
Stage changes the scale, not the method, and these bands are rules of thumb. Under roughly $500K a year, fix your top ten products and check by hand. Between $500K and $2M, fix the template so every product carries the same facts, and give one person the routine. Above $2M, the usual gap is consistency between your product pages, feeds and marketplace listings, so reconcile those to one source before buying any tracking tool. At that size clean facts are table stakes rather than an edge: agents such as Alexa for Shopping also draw on a shopper’s own history, and a rival with equally good data can still win, so treat each agent channel as its own test with its own measurement.
Agents bring shoppers who arrive already decided and with no history with your brand, which makes the second order its own job; our ecommerce customer retention guide picks that up. For the technical layer connecting agents to stores, see the Agent Protocols hub.
Disclosure: I am in the Shopify Partner Program and earn affiliate commissions from Shopify. Names on this page may link through /recommends/ links that can earn a commission.
An AI shopping agent is an AI assistant that finds, compares and shortlists products for a shopper based on a request written in plain language, and on some platforms completes the purchase. Examples shoppers use today include ChatGPT shopping, Google AI Mode and Gemini, Microsoft Copilot, Amazon’s Alexa for Shopping and Perplexity. Each reads different sources: merchant feeds, Shopify Catalog, Google’s Shopping Graph, Amazon’s catalog or the open web. Some also watch for a condition, such as a price drop, and act when it is met. For a brand, the discovery job matters most, because it decides whether your product is considered at all.
AI shopping agents decide by matching the constraints in a shopper’s request against product facts they can read, and controlled tests show price, ratings, reviews, endorsements and on-screen position also sway the choice. A study revised in December 2025 found agents penalized visible sponsored tags, rewarded platform endorsements and differed sharply between models, including in which screen positions they favored. It also found model updates could reshuffle which products won. It tested the final choice among products already on screen, not how live assistants retrieve them. So clear, consistent product facts are the part you control, not a guarantee of selection.
A small Shopify brand can get recommended by ChatGPT over bigger brands, and its best odds are usually requests specific enough that its product is plainly the right fit. Controlled research found agents tend to concentrate repeated choices on a few products, which favors well documented options on broad requests. That makes narrow requests naming a use, size, material or constraint the more realistic target, though this is a practical judgment rather than a proven rule. Pick the five or six use cases where you are genuinely the best fit and make sure your title, description, metafields and reviews all say so in plain words.
Sponsored labels did not help in a published controlled test: in the “What Is Your AI Agent Buying?” study, agents penalized products carrying a visible sponsored tag and rewarded platform endorsements such as an overall pick badge. That result is about the label on a listing, not about whether buying distribution inside a live assistant pays, which the study did not test. Platforms are still building ad formats and few have published results. Treat any paid placement inside an assistant as its own experiment with its own measurement, and keep clear product facts as the foundation either way.
Check what AI shopping agents say about your products on a regular schedule, and monthly is a reasonable starting point for a small team. Use the same five or so specific shopping requests each time in two or three agents, run each two or three times because answers vary, and record the date, the agent, whether you are named, how you are described and who appears instead. Treat the results as directional rather than proof of why anything changed. When an answer shifts, compare the products that replaced you against your own listing to see which constraint they state that you do not.