- 01Agentic commerce stalled at the button, not at the reader
- 02Declared intent: where AI shopping lives today
- 03Inferred intent: what personal agents change
- 04The gatekeeper split: wall, travel, open
- 05What content has to do, by persona and by task
- 06Visibility is now a moving target
- 07What to watch next
- 08Frequently asked questions
- 09References
A shopper opens ChatGPT and types, "What's a good grain-free food for a senior Lab?" She knows she is shopping. The assistant knows she is shopping. The answer is a ranked list, and she chooses.
A different shopper asks her personal AI agent to book the vet for Friday. The agent finds a slot, sees in her email that the vet flagged weight gain at the last visit, notices the dog food reorder is due next week, and asks whether she wants to try the weight-management formula instead. She never asked a shopping question. The agent created one.
That second moment is new, and it is the one that matters. Across the three shoppers brands now serve — the Human Shopper (about 85% of traffic), the AI-Assisted Human (roughly 10 to 15% and rising), and the Autonomous Agent (under 1% today) — most AI shopping so far has started with a person who already knew what they wanted. Personal agents change where purchase intent begins.
This piece argues one thing: agentic commerce will not accelerate because shoppers finally trust an AI to click "buy." It will accelerate because personal agents start creating shopping moments the shopper never declared.
Agentic commerce stalled at the button, not at the reader
For most of the past year, "agentic commerce" meant one thing: checkout inside a chat window. That version has had a hard time.
OpenAI launched Instant Checkout in ChatGPT in September 2025 and ended it in March 2026, saying the first version did not offer the flexibility it wanted and that merchants would use their own checkout while it focused on product discovery. Checkout did not disappear. It moved. Walmart's Sparky now runs inside ChatGPT, taking shoppers from discovery into a Walmart environment with their own account, loyalty, and payment. Google announced Universal Cart in May with two paths to buy: check out on Google, or send the cart to the merchant's site. Microsoft made Universal Commerce Protocol feeds generally available in Microsoft Merchant Center in April, with account linking that surfaces loyalty benefits inside Copilot.
The reading layer, meanwhile, never stalled. Discovery is where the volume and the money went. OpenAI opened self-serve advertising in ChatGPT in May. NielsenIQ found that 42% of consumers used an AI tool while shopping in the past month, while only 5% of U.S. consumers had used a fully autonomous agent to place an order.
Read together, the lesson of the past year is not that shoppers reject AI in commerce. It is that the checkout button was never the constraint. What was missing was a reason to shop that the shopper did not have to bring.
Declared intent: where AI shopping lives today
Almost all AI shopping today runs on declared intent: the shopper arrives knowing they are shopping. It takes two forms.
Question-led. The shopper asks a general assistant — ChatGPT, Gemini, Perplexity, Copilot — a product question and reads a synthesized answer. The assistant is not a store, but it shapes the shortlist. (The AI shopping assistants field guide covers each in depth.)
Shop-led. The shopper is already inside a retailer and asks the retailer's own assistant. Amazon's Alexa for Shopping, which replaced the Rufus brand in May, was used by more than 350 million customers in the 12 months to July, with active users nearly doubling year over year, according to Amazon. Andy Jassy told investors that U.S. customers who use it spend 40% more per order on average than those who do not. Walmart's Sparky has been tried by roughly half of Walmart app users.
In both forms, the shopper did the first and hardest step: recognizing a need and deciding to shop for it. Content wins in this world by answering the question well — the right claims, the right attributes, consistent across every source the assistant reads. That is the world the answer engine optimization discipline was built for, and it is not going away.
Inferred intent: what personal agents change
A personal agent works differently. It is not a place the shopper goes to shop. It is a helper the shopper hands their week to, and shopping happens along the way.
| Declared intent | Inferred intent | |
|---|---|---|
| Where | ChatGPT, Gemini, Perplexity, Copilot; Alexa for Shopping, Sparky | Personal agents such as Muse and Instinct |
| Who starts the shopping moment | The shopper | The agent, spotting a need inside another task |
| What the shopper sees | A ranked answer or results page | A few options and an approval request |
| What wins | Answering the question well | Being the obvious fit for a need the shopper never typed |
Two products arrived this summer that bracket the category.
Meta's Muse launched September 8 in the U.S. Meta describes it as more proactive and longer-running than a chatbot. It connects to email, calendar, payments, and shopping, runs on its own cloud computer, and keeps working on a schedule or in response to events after the app is closed. It checks with the user before sensitive actions, including every purchase. Sensor Tower data shows more than 2.5 million downloads in its first two weeks.
Instinct, from Spear Street Technology, has no app at all. Users text or call it, and it acts across their connected accounts. It is invite-only, and its founder has described early users sending it on grocery runs and cancelling subscriptions. It raised $250 million at a $2.5 billion valuation in August and is reportedly in talks at a far higher one. Purchases now run through Stripe's Link, where users approve them; separately, early testers reported the agent sending an email without asking first.
Neither company has published how much of its usage is shopping. The volume is not yet measurable. The mechanism already is: the agent carries context the shopper never restates, and it turns that context into purchase moments. Two categories show how differently that plays out.
Snacking: impulse becomes occasion
Snacking impulse has always been won at the physical shelf — eye level, the endcap, the checkout lane. It does not travel well online. NielsenIQ research shows consumers are about half as likely to make an impulse snack purchase online as in a store, even as online accounted for roughly three-quarters of U.S. consumer packaged goods dollar growth in 2025.
An agent has no cravings and never walks past an endcap. What it has is the occasion: game night on the calendar, a road trip it is booking, a week of school lunches. When an agent builds a basket around an occasion the shopper did not spell out, impulse does not vanish. It moves from placement to occasion inference, and the agent chooses the snack unless the shopper named a brand.
There is a nuance worth watching. Our snacking evidence report found that occasion-framed prompts pull AI answers toward mainstream indulgent brands, while health-qualified prompts pull toward better-for-you challengers. An agent planning game night generates an occasion frame by default. Incumbents may benefit first — unless the agent also knows the shopper's constraints, such as a household managing GLP-1 appetite changes or a nut-free classroom.
Pet nutrition: the agent sees the autopilot break coming
Pet food runs on autopilot. Chewy reported that 84.6% of its net sales in its most recent quarter came from Autoship customers, a figure that includes their purchases outside the subscription itself. Our pet nutrition evidence report found subscription penetration across the category at about 52%. Agents do not add delegation to this category. It was delegated already.
What agents add is awareness. The autopilot breaks at life moments — a vet diagnosis, a new puppy, a move to senior life stage, a recall, a price jump. Those signals live in email, calendars, and photos. A retailer's subscription engine knows the reorder cadence. A personal agent can see the life around it, which means it can spot the switching moment before the shopper does and put a new formula in front of her.
Muse's design hints at where this leads. Its permission system lets users grant standing approval for many kinds of repeated tasks, while every purchase still requires sign-off. The reorder itself may someday run quietly; the switch is what will reach the shopper.
The gatekeeper split: wall, travel, open
Within 48 hours in September, three large retailers showed three different answers to the same question: what happens when someone else's agent wants to shop here?
Wall. Amazon blocked Muse from its store on September 20. Its stated reasoning is that third-party apps buying on customers' behalf should operate openly and respect a retailer's decision about whether to participate, as food delivery apps and travel agencies already do. Amazon said it had not agreed to Muse accessing its store; Meta has said Muse never sees users' passwords or payment details. Amazon runs its own agent path through Alexa for Shopping and Buy for Me. The legal backdrop is unsettled: an appeals court ruled in August, in Amazon's case against Perplexity's Comet browser, that it is the user, not the AI company, who accesses the site — leaving terms-of-service claims as the remaining route.
Travel. Walmart sends its own agent out. Walmart's Daniel Danker said in March that the Sparky experience would “travel directly into ChatGPT and Gemini and anybody else that we ever integrate with.” The shopper meets Walmart's agent inside someone else's assistant, and the account, loyalty, and checkout stay with Walmart.
Open. Shopify opened its merchants to Muse. On September 21, Shopify CEO Tobi Lütke announced agentic checkout with Shop Pay for Muse across Shopify stores, with merchant consent built into the existing checkout. PayPal announced a similar arrangement. Markets noticed: Shopify and Meta shares both closed sharply higher that day.
Each posture has coherent commercial logic, and none is the right one. For a brand, the implication is the same under all three: the same product is now read by Amazon's agent inside the wall, by Sparky when it travels, and by Muse on an open storefront. The brand does not choose which.
What content has to do, by persona and by task
One clarification makes the rest of this section simpler:
“the persona is a property of the task, not the tool.”
The same shopper, using the same agent, is an AI-Assisted Human for game-night snacks — the agent proposes, she approves. She is effectively an Autonomous Agent for the dog food reorder, where a standing rule replaces a new decision. Muse does not move from one persona to the other. Tasks do, one at a time, starting with the most routine. That is why autonomous traffic stays small even as agents spread, and why every persona still needs its own kind of content.
| Reader | What it is doing | What content must provide |
|---|---|---|
| Human Shopper | Scanning a results page and product page | Scannable titles, images, and highlights. Amazon's new 75-character title cap and Item Highlights field make this sharper, and the listing itself is still where most traffic lands |
| AI-Assisted Human, declared intent | Asking a question and reading a synthesized answer | Claims and attributes that answer real questions; review depth; consistency across retailer pages and brand sources of truth |
| Personal agent, inferred intent | Matching a need nobody typed against constraints, often on a schedule | Complete structured attributes — occasion, format, pack size, dietary and life-stage fit, replenishment cadence — plus live price and inventory in the catalogs agents can reach |
| The human at the approval step | Choosing between a few options, or confirming one, in seconds | A one-line reason to pick this product. Brand's right to win, compressed |
The last row deserves its own line.
“The approval card is the new PDP.”
The shopper still decides, but she decides from a two-line summary the agent wrote from your content. If the attributes and claims that make the case are missing, the agent cannot make it for you.
What agents weigh is starting to be measured. A peer-reviewed study from Columbia Business School and Yale, which ran AI shopping agents through controlled storefronts, found that agents weigh price, ratings, and reviews; penalize "Sponsored" tags; reward platform endorsements such as "Overall Pick"; and that modest edits to a product title can materially shift which products they select. It is one study, run in a simulated marketplace. It points in a clear direction all the same: the reader is paying attention to content, not placement.
Visibility is now a moving target
If agents read content, how visible a brand is inside their answers becomes a live number. It moves for reasons a brand does not control, and for reasons it does.
What a brand does not control. The same Columbia and Yale study found that model updates can reallocate demand across products even when the catalog does not change at all. Answers also vary from one run to the next. One independent tracking study of nearly 700,000 AI answers found that two repeat ChatGPT answers to the same question shared, on average, only about a fifth of their cited sources; Google AI Overviews shared about a third. That measures sources rather than product picks, and it blends randomness with real change. The direction still holds: a single check is an anecdote.
What a brand does control. Titles, attributes, reviews, and consistency. Each is an input agents read, and each can be changed.
What can now be watched. Google's Merchant Center now shows brands their AI shopping share of voice across AI Mode, AI Overviews, and the Gemini app, in a U.S. pilot that began in July. (Our Merchant Center AI insights piece covers what it shows and withholds.) Retailer-owned agents have announced no comparable report, and personal agents report nothing to brands at all. The full picture has to be assembled across surfaces.
The practical consequence is observational rather than prescriptive: a quarterly audit captures a snapshot of something that now changes week to week. The reader is always on. Visibility is always moving.
What to watch next
Whether personal agents secure retailer agreements. Shopify and PayPal chose to open. Amazon chose to wall. The next large retailer to choose will show which posture is becoming the norm.
Whether declared-intent assistants learn to infer. ChatGPT, Gemini, and Alexa for Shopping all add memory and proactive features. If they begin to suggest purchases the shopper did not ask for, the line between declared and inferred intent will blur from the other side.
Whether the approval step holds. Muse requires sign-off on every purchase today. That is a policy setting, not a technical limit. If it ever relaxes, routine replenishment is the likely first candidate, and pet nutrition is the category where it would show first.
Whether shoppers follow. Only 16% of shoppers say they are comfortable letting an AI assistant find and buy products for them, according to Coveo. Downloads are fast. Trust is slower. The <1% figure for autonomous traffic reflects that, and it is the number to watch.
For brands, none of these changes the underlying work. The reader has already changed. What is changing now is where the shopping moment begins — and the brands whose content is legible to every reader, in every posture, will be the ones the agents can choose.
References
Company and primary sources
- Amazon, CEO Andy Jassy on Amazon Stores growth, delivery speed, and AI shopping (Q2 2026 earnings, July 30, 2026). https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-stores-growth-ai-shopping-q2-2026-earnings
- Meta Help Center, How Muse works with your guidance and approval. https://www.meta.com/help/artificial-intelligence/1385290430137537/
- Google, Google Shopping introduces Universal Cart, agentic shopping (May 19, 2026). https://blog.google/products-and-platforms/products/shopping/google-shopping-cart/
- Chewy, Form 10-Q (fiscal 2026; definition of Autoship customer sales). https://www.sec.gov/Archives/edgar/data/0001766502/000162828026061018/chwy-20260802.htm
- Allouah, Besbes, Figueroa, Kanoria, Kumar, What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce (Columbia Business School / Yale; ACM Web Conference 2026). https://arxiv.org/abs/2508.02630
- NielsenIQ, 42% of consumers now use AI tools to shop (2026). https://nielseniq.com/global/en/news-center/2026/42-of-consumers-now-use-ai-tools-to-shop-niq-data-shows/
Reporting
- Retail Dive, Walmart brings Sparky to ChatGPT as OpenAI rethinks Instant Checkout (March 25, 2026).
- CNBC, OpenAI's first crack at online shopping stumbled. It's preparing for the next wave (March 20, 2026).
- Retail Dive, Amazon customers are embracing Alexa for Shopping (August 3, 2026).
- Axios, Meta debuts Muse, its long-planned personal AI agent (September 8, 2026).
- GeekWire, Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping (September 2026).
- Axios, Amazon kicks out Meta's Muse in split over AI shopping (September 21, 2026).
- Quartz, Shopify partners with Meta Muse for AI-powered checkout (September 22, 2026).
- CBS News, Meta's Muse personal agent is a hit, as AI commerce starts to take off (September 23, 2026).
- TechCrunch, Viral AI startup Instinct has raised $350 million at a $2.5 billion valuation (August 26, 2026).
- FoodNavigator, NielsenIQ reveals the future for impulse shopping (August 10, 2026).
- Parse, AI citation volatility by industry (2026). https://parse.gl/research/ai-citation-volatility-by-industry
Related Genrise insights