Agentic shopping was slated to be the next big thing to revolutionize the way people discover and buy products online. So far it hasn’t lived up to its promise.
For instance, ChatGPT launched Instant Checkout in late 2025 and discontinued it only five months later.
Where many go wrong is thinking that’s the end of agentic shopping.
The retailers who understand why it stumbled and can see how it’s being regrouped for the future will be the ones best positioned when it finds its footing.
In this post, we cover:
- How agentic shopping actually works today
- What our testing of 120+ prompts revealed about agent behavior
- Why most ecommerce sites aren’t ready (based on 1,100 agentic readiness audits)
- How you can prepare your store for agentic purchases from now
What is agentic shopping and how does it work
Agentic shopping is a form of online shopping in which an AI agent handles the product discovery and purchasing process on a user’s behalf.
Rather than surfacing a list of options and leaving the rest to you, an AI agent:
- Interprets your request
- Searches for products
- Evaluates options against your constraints
- Navigates to a retailer’s website
- Adds products to cart
- Initiates (and sometimes completes) the checkout process
The promise of agentic shopping is that AI agents will be able to complete end-to-end transactions on behalf of users. However, neither the market nor the technology is developed enough to support agentic shopping right now.
Agentic shopping already had its first false start
OpenAI’s Instant Checkout was a feature that let users discover and purchase products directly inside a ChatGPT conversation. No need to visit a retailer’s website at all.
Walmart, an early adopter, put it to the test with 200,000 products, only to find its conversions were three times lower than when shoppers completed the transaction on its website instead.
Only around 12 of Shopify’s millions of merchants ever went live before OpenAI pulled the feature in March 2026. Three structural problems drove the discontinuation of Instant Checkout:
- Retailers lost control of their conversion environment. A chat interface strips away everything that closes a sale, such as product imagery, reviews, upsells, trust signals, and loyalty benefits. Moving checkout off-site doesn’t just reduce conversion; it removes every tool retailers have built to increase it.
- Users aren’t ready to hand over the decision. Forrester found that completing a purchase inside an answer engine is the least-adopted use case among regular AI platform users. Our own testing from 120+ agentic prompts found agents making at least seven assumptive decisions about products to add to the cart without asking for the user’s confirmation. Trust in agentic purchasing is being built incrementally, and AI agents that act before asking are delaying that process.
- Live product data at scale is an unsolved problem. Prices change, stock runs out, and shipping costs vary by location. Keeping thousands of product listings synchronized across many merchants in real time proved far more difficult than anticipated. Without uniform, live product feeds, agents confidently recommend products that are out of stock or mispriced.
Despite these challenges facing agentic ecommerce, the concept hasn’t been abandoned. Rather, AI platforms have pivoted to a discovery-and-redirect model, allowing users to find products inside the chat window and then complete purchases on the retailer’s own site.
For instance, ChatGPT and Google AI Mode frequently guide users on how to purchase the product on a retailer’s website.
For retailers, that’s actually a better outcome than losing the transaction to a chat interface entirely.
The current state of agentic ecommerce readiness
For an AI agent to shop on a user’s behalf, it needs to be able to find a store, read its content, understand what’s for sale, and complete a transaction. Most ecommerce sites can’t support even the first of those steps.
Cloudflare’s AI Insights tool analyses agent-readiness signals across the top 200,000 scanned domains. It shows that very few agentic protocols are adopted at the moment.
When looking specifically at the agentic readiness of ecommerce stores, they are significantly below the average.
For instance, only 15% of ecommerce sites have a robots.txt file (compared to 84% of the top domains scanned). Only 13% have a sitemap.
These aren’t advanced protocol requirements. They’re the foundational infrastructure the web has run on for decades, and the majority of ecommerce sites still don’t have them in place for agents to use.
Beyond the basics, adoption of other agentic protocols and standards drops sharply:
- Markdown negotiation (5.6%): the ability to serve content in a clean, structured format that agents can parse efficiently, rather than rendered HTML
- Universal Commerce Protocol (5.6%): an emerging standard that gives agents a structured way to query product data, pricing, and availability
- OAuth discovery (5.4%): allows agents to establish authenticated sessions with retailers
- MCP Server Cards, A2A Agent Cards, and x402 (effectively 0%): the most advanced layer of agent infrastructure, covering tool discovery, agent-to-agent communication, and native payment protocols, respectively
In our own agentic readiness scans of 1,100 ecommerce brands, the picture was consistent. The vast majority scored at Level 1 with basic web presence only. Few sites reached Level 2. None reached Level 3 or above.
Perhaps the most telling finding is that 41% of the sites we scanned blocked the agent readiness scanner outright with bot protection.
These sites didn’t score poorly on agent readiness; they refused to be assessed at all. Real purchasing agents could therefore also be blocked from accessing the website.
What 120+ agentic shopping prompts taught us about agent behavior
To understand how AI agents currently behave during a shopping journey, we ran 120+ prompts across ChatGPT and Google AI Mode covering both product discovery and directed purchase tasks.
Prompts spanned 16 product categories, including electronics, fashion, homewares, and beauty, and were tested with and without specific constraints like price caps, delivery requirements, and product specifications.
These were the stand-out insights.
Agents act as logistics coordinators, not just search engines
Agents go beyond finding products toward resolving the full purchase equation in a single step. They answer questions like who has the product you’re looking for in stock:
Where the product is available in your area:
Who delivers the fastest:
And who’s cheapest:
Many real-world purchase failures happen because logistics don’t align with a buyer’s requirements. For instance, the right product is available but delivers too slowly, or is cheaper elsewhere but out of stock.
Agents collapse that multi-tab research into a single response.
As agentic protocols become standardized, the pipelines that connect retailer inventory, pricing, and delivery data to AI platforms will mature, making it far easier for both online and local retailers to surface accurate, real-time availability to nearby shoppers.
The retailers who invest in keeping that data clean and current will be the ones agents recommend when a customer needs something asap.
Agentic shopping is a local experience by default
Because ChatGPT and Google AI Mode are primarily used as logged-in experiences, their agents have access to persistent location context.
In our testing, ChatGPT pulled location from memory unprompted, serving AU-specific retailers and pricing by default.
When a VPN was active, the agent’s thinking logs revealed confusion, often referencing both the user’s known location and the VPN location simultaneously before making a call.
For instance, in this conversation, my VPN was set to the USA, and yet the agent still made comments about purchases in Australia.
Google AI Mode went further, embedding live maps directly into shopping responses, showing store locations, distances, and real-time pickup availability alongside product recommendations.
A search for a rain jacket returned a Kathmandu store marked “Immediate In-Store Pickup” at 5km away.
A skincare search returned retailer pins at 200m scale.
Agentic shopping results are inherently localized, meaning local inventory data, delivery zones, and click-and-collect availability aren’t optional extras for retailers to include on their websites and in product feeds.
They’re a primary data layer that agents query when deciding what to recommend.
Constraint handling is surprisingly sophisticated
In our tests, both AI agents we tried handled constraints and multi-condition queries, including price caps, spec filters, delivery windows, and availability requirements reasonably well.
They also held the line when no perfect product match existed.
For example, when we asked both Google AI Mode and ChatGPT to find a refurbished iPhone 14 Pro, 256GB, Space Black, under $600, neither platform hallucinated a match.
Google addressed each constraint raised in the prompt and explained how it affected the purchase price.
It also offered to tailor its search parameters to help the user find the best compromise for the product. However, the next steps were largely left in the user’s hands to monitor marketplaces like eBay for their ideal product.
ChatGPT simply admitted it couldn’t find what the user was looking for and went straight to recommending similar products.
However, it ended with an assessment of the realism of the user’s constraints and an offer to monitor the market and alert the user if a product matching these conditions becomes available.
That kind of constraint-aware rejection is more useful than a forced match.
Rather than recommending a product that doesn’t meet the prompt, both platforms surfaced the gap, explained it, and gave the user a path forward. For ecommerce teams, this means agents are actively filtering out products that don’t meet a user’s stated requirements, making accurate, up-to-date product data and pricing more critical.
Agents make assumptions when they should ask questions
When ChatGPT navigated a major furniture retailer’s website on a user’s behalf, it made at least seven decisions without asking, at times because they “seem to match the user’s interest”:
Or because “the user didn’t specify”:
Or because it simply decided it “seemed appropriate” and therefore didn’t need to be changed:
Many assumptions were reasonable in isolation, but collectively represented a significant trust gap.
- Cookie consent: Auto-accepted the cookie popup without considering whether the user would prefer to manage their settings
- Color: Defaulted to white with no explanation or prompt from the user
- Size: Selected the 140x65cm variant without consulting the user
- Configuration: Chose a drawer on the left, noting it “seemed to match the user’s interest” despite no context being provided
- Account status: Assumed the user didn’t have an account and selected guest checkout automatically
- Postcode: Pre-filled the delivery postcode and noted it “seems appropriate for delivery, so we don’t need to change it”, without confirming with the user
- Delivery method: Selected “deliver to door” without asking whether click-and-collect might be preferred
The agent completed the task. But it did so by making the user invisible in the process.
This is one of the core UX challenges AI designers need to solve before agentic commerce becomes a genuinely smooth experience.
Ask too many questions, and the agent becomes more friction than the checkout it’s replacing. Make too many assumptions, and users end up with the wrong color, wrong size, and a delivery slot they don’t like.
The sweet spot — an agent that knows when to act and when to check — is still very much a work in progress.
How to optimize for agentic shopping across three levels
Agentic shopping is still early, but the retailers who structure their data well now will have a compounding advantage as the protocols mature. There are three levels worth optimizing for.
1. Product level: Specs, attributes, and facets
Agents don’t read between the lines.
When a user asks for a fragrance-free moisturizer for rosacea under $40, the agent matches products against attributes such as ingredients, skin-type suitability, and price.
Vague copy like “great for sensitive skin” doesn’t give an agent enough signal to make a confident recommendation. Structured, specific product data does.
That means complete attribute sets, accurate facets, and descriptions that answer the questions your customers actually ask, because those are exactly the queries agents are fielding on their behalf.
- Attributes are the core product specs: dimensions, weight, materials, compatibility.
- Facets are the filterable characteristics shoppers use to narrow choices: color, size, price range, skin type, dietary requirement.
- Features are the functional benefits that answer “What does this actually do for me?” This is the layer that connects specs to real-world use.
Agents draw on all three when matching a product to a query.
Auditing your product catalog for gaps in these three layers is the most immediate action ecommerce teams can take to improve their product’s visibility in agentic results.
2. Retailer level: The logistics and trust layers
In agentic search, the retailer is as recommendable as the products it sells.
Agents don’t just match products to queries. That’s only half the job. The other half is evaluating which retailer to send a user to based on a combination of logistics and trust signals.
On the logistics side it’s about assessing:
- Who has the product in stock
- Who delivers the fastest
- Who covers the user’s area
- Who’s cheapest right now
On the trust side, signals like return policies, warranty terms, seller ratings, and whether the retailer is a recognized name in the user’s market take priority.
What’s striking is how explicitly agents surface these evaluations.
In our testing, retailer recommendations frequently included a one-line qualifier that served as a trust summary (effectively a mini review written by the agent to help the user decide who to buy from).
Elements like budget compatibility and the retailer’s brand identity also played a role.
Agents use positioning and origin as recommendation signals, describing one retailer as offering “contemporary Australian designs with excellent engraving options” rather than just listing products. Retailers whose brand story, values, and specializations are clearly communicated online give agents a better signal to characterize them accurately and favorably.
Retailers whose data across all of these dimensions (inventory, logistics, trust signals, and brand positioning) is current, accurate, and structured will win recommendations.
Those whose data is stale, incomplete, or hard for agents to parse will lose out to a competitor who isn’t.
3. Audience level: Patterns, preferences, and personalization at scale
The product and retailer layers are about being findable. The audience layer is about being chosen.
As agents get better at personalization, they increasingly match products and retailers to users based on inferred context. Not just what someone asked for, but what the agent infers they typically care about.
A regular trail runner asking for shoes carries different implicit requirements than a beginner. A buyer who always filters by sustainable materials or Australian-made products shouldn’t have to say it every time. A parent shopping for a child’s desk has different priorities than a design-conscious professional furnishing a home office.
This is where the agent’s memory and the retailer’s understanding of the audience converge. Agents build a picture of a user over time from signals such as purchase history, stated preferences, recurring constraints, and lifestyle preferences. Retailers who understand their audience deeply enough to reflect those patterns in their product data, content, and positioning give agents a better signal for making the match.
In practice, that means thinking about:
- Jobs to be done: What problem is your customer actually solving, and is that reflected in how you describe your products?
- Common constraints: What are the recurring requirements your audience brings, such as budget ranges, size needs, dietary requirements, and technical specs? Are those filterable and structured in your catalog?
- Taste and values clusters: What does your typical customer care about beyond the product itself — sustainability, local manufacturing, brand ethics, aesthetic style — and is that part of how your brand is represented online?
Agents can only personalize to the patterns they can see.
The more clearly your product data and brand positioning reflect the specific audience you serve, the more accurately an agent can recommend you to the right person at the right moment.
The technical and UX barriers standing between browsers and buyers
Agents experience the friction human shoppers already deal with, just with fewer workarounds available.
The barriers to agentic ecommerce that we observed across our tested purchase journeys and confirmed at scale through Cloudflare’s data include:
- Cookie consent walls add steps and failure points before an agent can even begin browsing
- Login walls and guest checkout modals interrupt agent flow at decision points, and with near-zero OAuth adoption across ecommerce, agents have no way to authenticate smoothly
- Rendered HTML pages force agents to interpret content rather than read it cleanly. Only 5.6% of ecommerce sites offer a structured alternative
- Multi-step delivery and postcode UX created the most back-and-forth in our observed purchase journey. Unstructured logistics inputs are a significant failure point
- Aggressive bot protection blocks legitimate purchasing agents alongside malicious crawlers. 41% of sites in our scan refused to be assessed at all
- Front-loaded personal information requests push the agent’s natural handoff point earlier than necessary, shortening how far it can get before returning control to the user
Many of these are worth fixing regardless of agentic commerce because they’re friction points for human shoppers too.
Agentic commerce stumbled in its first chapter because the infrastructure, the trust, and the data pipelines weren’t ready. All three are being built right now.
The retailers who understand what went wrong (and why the next iteration will be different) are the ones who will be ready when it lands.
