ChatGPT product discovery: how OpenAI killed Instant Checkout (and what replaced it)

Less than three weeks ago, OpenAI made an announcement that quietly rewrote the rules of AI commerce. At ShopTalk 2026 on March 24, the company unveiled a richer, more visual product discovery experience inside ChatGPT — and in the same breath, killed off Instant Checkout, the in-chat purchase flow it had launched just six months earlier.

This isn’t a feature update. It’s a strategic pivot — and the signal it sends about where AI commerce is heading is more important than the feature itself. Lumio works with brands navigating exactly this shift; this post unpacks what OpenAI changed, why retailers wanted Instant Checkout gone, and what every brand serious about AI readiness should do next.

The Instant Checkout experiment

To understand the pivot, you need the backstory.

In September 2025, OpenAI launched Instant Checkout alongside the Agentic Commerce Protocol (ACP), a new open standard co-developed with Stripe.¹ The pitch was ambitious: shoppers could discover, configure, and buy products without ever leaving the ChatGPT interface. Walmart, Shopify, Etsy, and PayPal were the launch partners. The press coverage was breathless — TechCrunch called it a direct assault on Google and Amazon.

Six months later, the numbers told a different story. Only a handful of Shopify brands had activated Instant Checkout. Walmart’s EVP of product and design, Daniel Danker, told Wired that in-chat purchases through Instant Checkout converted at roughly one-third the rate of click-through traffic to Walmart’s website — a 3x conversion gap that retailers couldn’t ignore. Product selection remained limited. Pricing wasn’t always current.

OpenAI’s official statement to Adweek was diplomatic but clear: “We’ve found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery.”

In plainer terms: OpenAI tried to own the checkout. Retailers didn’t want them to. They’re backing off.

The new model: discovery-first commerce

What replaced Instant Checkout is more interesting than what was removed.

The new product discovery experience — rolling out to all ChatGPT Free, Go, Plus, and Pro users — includes visual product browsing within chat, image-based search (upload a photo to find similar items), side-by-side comparisons with prices, reviews, and features, and conversational refinement of results. When you’re ready to buy, ChatGPT redirects you to the retailer’s website or their in-ChatGPT app. OpenAI no longer owns checkout.

The retail partner list is substantial: Target, Sephora, Nordstrom, Lowe’s, Best Buy, The Home Depot, Wayfair, Walmart, and every Shopify brand via Shopify Catalog. Several retailers have built dedicated in-ChatGPT apps: Walmart’s app supports account linking, loyalty programs, and Walmart payments. Sephora’s app, piloted March 25, integrates Beauty Insider profiles for personalized recommendations.

OpenAI’s framing of the shift: “For users, this turns shopping from a fragmented, time-consuming process into a single, seamless experience. For merchants, it brings higher-intent shoppers who are closer to making a decision.”

The Agentic Commerce Protocol: what’s under the hood

ACP is the connective tissue. It’s an open standard (Apache 2.0 license, 1,300+ GitHub stars) co-developed by OpenAI and Stripe. The developer documentation lays out the full technical spec. Here’s what brands need to know.

Feed specifications

Attribute	Requirement
Formats	CSV, TSV (recommended), XML, JSON
Encoding	UTF-8
Max file size	10 GB
Compression	gzip
Update frequency	Every 15 minutes

Required fields per product: id, title (max 150 characters), description (max 5,000 characters), price (with ISO 4217 currency code), availability (in stock / out of stock / preorder), and image_url (at least one high-resolution image). Recommended fields include GTIN/UPC/MPN codes, customer reviews, rich media, shipping options, and performance signals.

Three-API architecture

1. Feeds API — full catalog uploads (daily snapshots)
2. Products API — individual product upserts and updates
3. Promotions API — promotion management

Brand onboarding happens at chatgpt.com/merchants. Currently US-only for onboarding, though discovery operates globally.

The critical design distinction

Here’s the line that should reframe how you think about ACP feeds: “ACP feeds are designed for AI reasoning, not indexing. Every field can become an argument the agent uses to explain why a product is relevant.”

This is the fundamental difference from Google Shopping feeds. Google feed attributes are indexed for keyword matching. ACP feed attributes are structured for an LLM to reason about — to construct an argument for why a specific product answers a specific question. A well-populated description field in an ACP feed isn’t metadata. It’s ammunition the AI uses to make your case to the shopper.

The Shopify angle

Shopify’s role in this ecosystem is worth calling out separately.

On the same day as OpenAI’s announcement, Shopify launched Agentic Storefronts — out-of-the-box access to ChatGPT, Microsoft Copilot, Google Search AI Mode, and Gemini, managed centrally from Shopify Admin. The partnership makes hundreds of millions of Shopify products discoverable through ChatGPT.

More notable: Shopify introduced an Agentic Plan — a new tier for brands not currently on Shopify that want to add products to Shopify Catalog specifically for AI channel sales. No separate integrations, no apps, no additional transaction fees beyond standard processing rates.

Launch brands include Glossier, Spanx, Vuori, Away, Stanley 1913, Steve Madden, Fenty Beauty, and Skims.

Shopify VP of Product Mani Fazeli framed it directly: “Agentic commerce isn’t something we’re reacting to; it’s a vision we’re bringing to life.”

The subtext: Shopify wants to be the default catalog layer for every AI commerce surface. If your products are in Shopify Catalog, they’re automatically available across every agent that supports ACP, UCP, or Shopify’s integrations. If they’re not, you’re managing each channel separately.

The numbers behind the shift

The traffic data explains why OpenAI is investing here — and why brands can’t ignore it.

ChatGPT processes roughly 50 million shopping queries per day. That’s about 2% of its 2.5 billion daily prompts. Stackline reported 84 million shopping questions per week from US consumers alone, and ChatGPT has grown from less than 1% to over 8% of Amazon’s weekly search volume in under a year.

The referral traffic is already material. Per Similarweb data from August 2025, ChatGPT accounted for 20% of Walmart’s referral clicks, 20%+ for Etsy, roughly 15% for Target, and about 10% for eBay. Amazon — which has blocked AI crawlers from accessing its marketplace data — sat below 3% and was declining.

The quality of that traffic matters more than the volume. Stackline’s research reports that AI-referred shoppers are less likely to bounce and convert at higher rates than shoppers from other referral sources — Stackline’s specific figures are in the linked article and worth reading directly.

Still early in absolute terms — AI traffic accounts for only about 0.2% of total ecommerce sessions per the Stackline reporting linked above. But the trajectory is unmistakable.

How OpenAI plans to make money

Sam Altman has confirmed OpenAI will charge approximately 2% commission on purchases made through ChatGPT, mirroring standard affiliate network rates. Product recommendations are explicitly not ads.

Altman’s public position on this is unusually direct: if ChatGPT accepted payments to rank products differently, “that’s probably catastrophic for your relationship with ChatGPT.” The stated vision is genuine best-fit recommendations funded by standard transaction fees — not pay-to-play.

OpenAI has begun testing ads in ChatGPT on Free and Go plans, but Plus, Pro, Business, Enterprise, and Edu accounts won’t see them, and ads do not influence recommendations. The commerce layer and the advertising layer are separate systems.

The three-ecosystem race

OpenAI’s pivot lands in the middle of a three-way competition for agentic commerce:

OpenAI ACP — 900 million weekly active users, 15-minute feed updates, conversational discovery with visual browsing, retailer-app integrations. Open protocol, early-stage, US-only for brand onboarding.

Google UCP (Universal Commerce Protocol) — built on the Shopping Graph’s 50+ billion product listings, full in-surface purchase journey with no redirect, backed by Walmart, Target, Etsy, American Express, Mastercard, Stripe, and Visa. The deepest existing brand infrastructure because it builds on Google Merchant Center.

Amazon — more than 250 million customers have used Rufus this year, with monthly active users up 149%, in a closed ecosystem that actively blocks OpenAI crawlers from accessing marketplace data. Amazon’s roughly 600 million product listings (per AMZScout, mid-2024) are absent from AI shopping results outside its own ecosystem. Its counter-strategy is fully closed: Rufus for in-app search, Alexa+ for voice commerce, and “Buy for Me” for agentic purchasing.

The competitive picture is still forming. But the directional bet is clear: brands who invest in structured product data that works across multiple AI surfaces — ACP feeds, UCP-ready Merchant Center feeds, JSON-LD on product pages — compound that investment across every agentic channel. Brands who optimize for one surface at a time will rebuild every time a new protocol launches.

The accuracy problem nobody’s solved

Amid the optimism, a reality check. ChatGPT’s product recommendations still have serious accuracy issues.

Testing by TechBuzz.ai (citing a WIRED investigation) found ChatGPT fabricating recommendations — citing specific products as “WIRED top picks” that reviewers had never endorsed.

OpenAI’s own Shopping Research model — a specialized GPT-5 mini variant post-trained with reinforcement learning for shopping tasks — scores 52% accuracy on multi-constraint queries. That’s better than the 37% baseline for standard ChatGPT Search, but it means nearly half of complex shopping queries return imperfect results.

OpenAI acknowledges the gap and encourages shoppers to verify on brand sites. The structured feed approach through ACP should improve accuracy — the model is working from brand-provided data rather than crawled web content — but the track record demands healthy skepticism.

For brands, this cuts both ways. Your product data quality directly impacts whether ChatGPT represents your products accurately. Bad data in, bad recommendation out.

What brands should do now

1. Get your products into ACP. If you’re on Shopify, you’re already there through Shopify Catalog. If you’re not, register at chatgpt.com/merchants and submit feeds via the Feeds API. The window for early-mover advantage is closing as more brands onboard.

2. Treat feed descriptions as sales arguments. ACP feeds are designed for AI reasoning. Your product descriptions should give the model specific, structured reasons to recommend your product for a given query. Not marketing copy. Not keyword stuffing. Concrete attributes, use cases, and differentiators.

3. Invest in cross-protocol data quality. The same product data improvements that make you visible in ChatGPT — attribute density, identifier hygiene, real-time inventory, rich media — also improve your performance in Google UCP and every other agentic surface. This is compound-interest work. The JSON-LD fields most merchants miss is a useful starting checklist.

4. Monitor your representation. Start asking ChatGPT shopping queries in your category and see what comes back. Are your products showing up? Are the prices correct? Are the descriptions accurate? If ChatGPT is misrepresenting your products, the fix is almost always in your data.

5. Don’t ignore the accuracy gap. If you’re a brand, monitor how AI surfaces describe your products. Hallucinated claims, incorrect specs, and fabricated review citations can damage trust. Better data reduces (but doesn’t eliminate) these risks.

The signal in the noise

OpenAI killing its own checkout after six months isn’t a failure story — it’s a correction that reveals where the real value is. Discovery, not transaction processing, is the hard problem. Getting the right product in front of the right shopper at the right moment, with enough context to make a confident recommendation — that’s the capability worth building.

The brands who win in this environment are the ones whose product data is rich enough to survive the reasoning process. Not rich enough to pass a feed validator. Rich enough that an LLM can construct a persuasive, accurate argument for recommending their product over every alternative.

If you want to know where your catalog stands — across ACP, UCP, and the broader agentic commerce landscape — get in touch. Lumio scores your product data against the attributes that AI surfaces are documented to read and shows you the specific gaps to close.

Footnotes

1. openai.com ↩