AI shopping agents are assistants like ChatGPT, Gemini and Perplexity that research, compare and sometimes buy products for a shopper.
To win with them, a store needs complete, consistent product data (price, stock, shipping, returns, reviews, specs) that an agent can read, and a landing page that converts the high-intent visitors those agents send.
Most of that work is the same clarity work that already wins A/B tests with human shoppers.
This guide covers where agentic commerce actually stands in November 2026, how ChatGPT shopping picks products, and what our database of real A/B tests says about the product information agents and people both rely on.
What agentic commerce is, and where it stands
Agentic commerce means an AI completes shopping tasks on someone's behalf: finding products, comparing them, and in some setups paying.
Google describes it as commerce "where AI completes tasks on people's behalf" in its January 2026 announcement.
Three open protocols now compete to connect agents with stores:
| Protocol | Backers | Launched | What it covers |
|---|---|---|---|
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe | September 2025 | Product feeds and checkout inside ChatGPT (Stripe) |
| Agent Payments Protocol (AP2) | 2025 | Agent-led payments (ap2-protocol.org) | |
| Universal Commerce Protocol (UCP) | Google, co-developed with Shopify, Etsy, Wayfair, Target and Walmart | January 11, 2026 | Discovery, checkout and post-purchase support; compatible with A2A, AP2 and MCP (ucp.dev) |
The checkout half of the story has moved slower than the hype.
OpenAI launched Instant Checkout in ChatGPT in September 2025, then in March 2026 said the first version "did not offer the level of flexibility that we aspire to provide" and shifted its focus to product discovery, letting merchants use their own checkout (CNBC).
Google is still pushing checkout. In May 2026 it announced a Universal Cart that works across retailers in Search and Gemini, with the option to pay through Google Pay or hand the cart to the merchant's site.
In both cases the retailer stays the merchant of record (Google).
The practical takeaway: today most agent-driven sales still end on your own site.
The agent picks the shortlist; your product page closes the sale.
How ChatGPT shopping works
OpenAI's help center says ChatGPT chooses products using structured metadata from first-party and third-party providers, such as price and product description, plus other third-party content.
It states product results "are not ads, nor influenced by any OpenAI partnerships," and that ads are shown separately.
When several stores sell the same item, ChatGPT ranks merchants "based on factors like availability, price, quality, and whether they are the maker or primary seller of that item."
It may also rewrite your product titles and descriptions into simpler versions.
There are two ways to get products in.
Shopify stores are already connected through Shopify Catalog, with no extra work required.
Other merchants can apply to send a direct feed using OpenAI's product feed spec.
That spec is the clearest public description of what an agent wants to know.
Nine fields are required: item ID, title, description, URL, brand, seller name, image, availability and price.
Optional fields include shipping price, whether returns are accepted, the return window in days, review count and star rating.
Two details matter for stores. An omitted shipping price "means unknown, not free shipping."
And availability must be explicit; a missing value rejects the row.
Optimizing for AI shoppers: what to fix first
We can't A/B test ChatGPT's ranking, and every test in our database was run on human shoppers.
But the fields agents read line up closely with page elements that stores have tested.
Making them clear, specific and consistent has a track record with people, and it is what the feed spec asks for.
Shipping cost and delivery date
If your feed says one thing about shipping and your page hedges, both agents and shoppers lose confidence.
Dave Diederen removed a "Shipping calculated at checkout" line from a product page that already promised free shipping and saw +14.07% revenue per visitor.
Concrete delivery dates also beat vague claims. Dave Diederen added a shipping bar with a "Ships by" date below the add-to-cart button for +13.48% revenue per visitor. See more in our shipping messaging tests.
Returns and warranty
The feed has fields for whether you accept returns and how many days customers get.
On the page, Blend Commerce showed Aeromotive's 90-day return policy and 1-year warranty below add to cart, which lifted average order value 17.26% and conversion rate 4.5%.
The same test carries a warning. A variant with the return policy alone raised add-to-cart clicks 16.43% but cut conversion 16%.
How you present a policy still needs testing; see our guarantee tests.
Price people can compare
Agents compare prices across sellers, and ChatGPT labels the lowest one it knows "Best price."
Comparable pricing helps people too: Blend Commerce showed a sports nutrition brand's price per dose next to the pack price and got +17.76% revenue per visitor and +23% add-to-cart clicks over two weeks. More in pricing and discount tests.
Specs and attributes
Agents match products to detailed requests ("waterproof, under $80, size 10"), so missing attributes can drop you from a shortlist.
On the page, Oliver Kenyon added a comparison table of key attributes to product pages for +$32,884 per month.
ConvertPolo moved CellBell's key specs to the first fold on mobile in a bundled redesign that reported +22.62% revenue, with no sample size given.
Specs are not a free win for human layouts, though.
Mintminds ran four tests moving specs on Lampenlicht.nl's product pages, and all four significantly lowered add to cart and revenue per visitor.
Put every attribute in your feed and structured data; test where they go on the page.
Availability and reviews
Availability is a required feed field and a ranking factor.
On the page, Dave Diederen showed "In Stock" instead of "Almost Out of Stock" and saw +16.78% revenue per visitor, so plain, accurate stock status can beat scarcity copy.
Review count and star rating are optional feed fields that agents weigh.
On the page, Blend Commerce put a product-specific review high on the PDP for Singular Sound and got +19.14% conversion rate.
Placement still varies by store, which is why our review tests include losers as well as winners.
@oliverkenyon
@DaveDiederen@DaveDiederen
Blend CommerceBlend CommerceProduct descriptions written as answers
Google now offers conversational attributes in Merchant Center so descriptions match how people phrase questions to AI.
It announced attributes that cover answers to common product questions, compatible accessories and substitutes.
Write descriptions that state facts plainly; our product description guide covers the tests on tone and angle.
What research says about how agents choose
A 2025 study, "What Is Your AI Agent Buying?" by Allouah, Besbes, Figueroa, Kanoria and Kumar, ran randomized trials with AI shopping agents on a mock store.
Agents showed strong position biases that differed by provider and model version, penalized "sponsored" tags, and rewarded platform endorsements.
Their sensitivity to price, ratings and reviews "vary sharply across models," and model updates "can drastically reshuffle market shares."
A seller-side agent that tweaked descriptions to match each query drove significant market-share gains.
Treat this as a lab result, not a playbook.
It used simulated stores, and any trick tuned to one model version may stop working after the next update.
Accurate, complete data is the only approach that doesn't depend on a given model's quirks.
Converting the visitors agents send
AI referrals are still a small slice of traffic, but they arrive ready to buy.
Adobe found traffic from generative AI tools to U.S. retail sites rose 693.4% year over year in the 2025 holiday season (Nov.
1 to Dec. 31), while noting "the base of users remains modest" (Adobe).
Those visitors converted 31% more than other traffic sources, spent 45% more time on site and viewed 13% more pages per visit, according to Adobe's AI traffic report.
A shopper arriving from ChatGPT has usually compared options already, so the page should confirm what the agent told them: the same price, stock and delivery terms, near the add-to-cart button.
That makes the product page the main thing to test for this traffic.
Segment AI referrals in your analytics (see GA4 for CRO) and check whether your current winners hold for that group.
On-site AI tools can help too, but test them.
The Bee's Knees British Imports switched to Google Vertex AI-powered search with 36 Agency and Nimstrata and saw +20.6% conversion rate over five weeks at 99% confidence.
In contrast, Oliver Kenyon removed a live chat feature that was costing about $28,000 a month.
More in site search and our AI for CRO hub.
How to measure whether any of this works
You can't run a clean A/B test on what ChatGPT recommends, because you don't control who sees which version. Three practical options:
- Test the page, not the agent. Run normal A/B tests on product pages and read results for the AI-referral segment.
- Use holdouts for feed changes. Improve attributes for one product group and leave a comparable group unchanged, as in a holdout test. Results will be noisy because agents change between model versions.
- Watch share of voice. Google's new AI performance insights in Merchant Center compare your share of voice on AI surfaces against similar brands, rolling out in the U.S. and four other countries.
Expect attribution gaps: some shoppers read an AI answer and then type your URL directly.
Our attribution guide covers how to read that kind of mixed signal.