A growing share of product discovery now happens inside an AI assistant. A shopper describes what they want, the assistant searches, compares and shortlists, and in some cases it can complete the purchase. For ecommerce teams, that adds a new shelf to compete on, and it is built almost entirely from product data.
This guide covers where agentic shopping stands in late 2026, how assistants choose products, why their answers change from one market to the next, and what to do about your data.
Where agentic shopping stands
The last year brought fast launches and some quick reversals.
| Date | What happened |
|---|---|
| September 2025 | OpenAI and Stripe launch Instant Checkout in ChatGPT, built on the open Agentic Commerce Protocol, starting with Etsy sellers in the US |
| September 2025 | Google announces the Agent Payments Protocol (AP2), developed with more than 60 organisations |
| November 2025 | Google adds agentic checkout in Search for eligible US merchants: it can buy on the merchant’s site with Google Pay once the shopper confirms |
| January 2026 | Google launches the Universal Commerce Protocol, co-developed with Shopify, Etsy, Wayfair, Target and Walmart; retailers stay the seller of record |
| March 2026 | OpenAI moves away from Instant Checkout, letting merchants use their own checkout “while we focus our efforts on product discovery” |
| May 2026 | Google says its protocol-based checkout will roll out in Canada and Australia, and later the UK |
The pattern is clear. Checkout inside the assistant is still settling, but discovery inside the assistant is here. Whether the final click happens in a chat window or on your own site, the shortlist is increasingly made by a model.
Traffic is growing from a small base. Adobe reported that visits to US retail sites from generative AI tools grew 693.4% over the 2025 holiday season compared with the year before, while noting that the user base “remains modest”.
How assistants choose products
Assistants build answers from structured data more than from web pages.
Feeds and catalogues. Google describes its Shopping Graph as “more than 50 billion product listings, 2 billion of which are updated every hour”, fed largely by merchants. OpenAI accepts merchant product feeds, and its specification lists nine basic fields “required for a useful discovery feed”: item_id, title, description, url, brand, seller_name, image_url, availability and price. Its standard upload “currently targets the US”.
Freshness. Price and availability are the fields that change most and matter most. OpenAI’s guidance is to update the feed when a sale starts or ends. An assistant that quotes yesterday’s price, or recommends a product that is out of stock, sends a disappointed shopper to you.
Signals beyond the feed. Reviews, ratings, prices and stock all shape what appears. Google’s assistant runs several searches at once for complex requests, so a product can be found through questions you never targeted. Our guide to AI Mode and rank tracking explains how that fan-out changes measurement.
Answers change by market
The same question does not get the same answer everywhere. In a study updated in October 2026, Yotpo, which sells commerce marketing tools, ran identical purchase prompts in the US, UK, Germany and Australia. Even the most consistent engine shared “only about 21% of its citations across markets”, and on Google AI Mode the figure was 14.6%.
That fits what we see in ordinary ecommerce. Catalogues, sellers, stock and prices differ by country, and our study of the same product in 20 countries showed how far prices alone move. An assistant answering a shopper in Munich is working from a different shelf from one answering in Chicago.
So a single check from one location tells you little. Brands selling in several markets need to look at each market separately.
Who lets agents in
Not every store welcomes them. Amazon’s robots.txt disallows a long list of AI crawlers and agents, including GPTBot, ChatGPT-User, OAI-SearchBot, GoogleAgent-Shopping, PerplexityBot and Claude-User. Amazon has said that third-party applications that offer to make purchases “should operate openly and respect service provider decisions whether or not to participate”, and it sued Perplexity over its Comet browser in November 2025. A preliminary injunction against Comet was later vacated on appeal, and the case continues.
On the other side, the industry is building ways for good agents to identify themselves. Cloudflare’s Web Bot Auth signs agent requests with public-key cryptography, and Visa’s Trusted Agent Protocol, launched in October 2025 with Cloudflare, builds on the same standard so that merchants can tell a legitimate shopping agent from an unknown bot. Our piece on the bot wall AI agents face covers this in more detail.
For merchants, this is a decision to make on purpose: which agents can browse, which can buy, and how you will recognise them.
What to do about your product data
- Treat feeds as a storefront. Fill every required field, write titles and descriptions for a reader who asks questions, and keep brand and seller names consistent across channels.
- Keep price and stock fresh. Update feeds when prices or availability change, not only on a schedule. Our guide to monitoring availability covers the stock side.
- Match the feed to the page. Assistants and shoppers both check your product pages. A feed price that differs from the page price costs trust.
- Decide your agent policy. Choose which agents may browse and buy, publish it in robots.txt, and use signed-agent verification where your stack supports it.
- Prepare each market separately. Feed country targeting, currency, language and local reviews all matter, because answers differ by market.
- Measure what assistants say. Track whether you appear for real shopper questions, at what price and with which seller, market by market.
How to measure it
Measurement means asking the questions shoppers ask and recording the answers, in each market you sell in. Use a fixed set of questions, repeat them over several days, and store the full answer with every product, price and link. Then check those answers against your own product pages as local shoppers see them. Residential proxies help with that second step, loading public product and search pages from the country you are checking.
Keep to public pages and each platform’s terms, and respect the choices of sites that keep agents out. The point is to see what shoppers see, not to get around anyone’s rules.
The bottom line
AI assistants are becoming a shelf of their own, and that shelf is stocked from product data. Checkout inside the assistant is still being worked out, but discovery already happens there, and it changes by market. Keep feeds complete and fresh, make them match your pages, decide which agents you let in, and measure what assistants say about your products in every market you sell in.
Sources and references
- Stripe, Stripe and OpenAI launch Instant Checkout and the Agentic Commerce Protocol (29 September 2025): stripe.com
- CNBC, OpenAI revamps shopping experience in ChatGPT after Instant Checkout (24 March 2026): cnbc.com
- OpenAI, product feed specification: developers.openai.com
- Google Cloud, Announcing the Agent Payments Protocol (16 September 2025): cloud.google.com
- Google, agentic checkout and AI shopping for the holidays (13 November 2025): blog.google
- Google, the Universal Commerce Protocol (11 January 2026): blog.google
- Google, shopping updates at Google Marketing Live (20 May 2026): blog.google
- Adobe, 2025 holiday shopping season results (7 January 2026): news.adobe.com
- Yotpo, Is your AI visibility global? (updated 8 October 2026): yotpo.com
- Amazon, statement on Perplexity’s Comet: aboutamazon.com
- Amazon robots.txt: amazon.com
- Cloudflare, Securing agentic commerce (24 October 2025): blog.cloudflare.com
- Visa, Trusted Agent Protocol (14 October 2025): usa.visa.com