Agentic Commerce
Shopping Research in ChatGPT & Agentic Commerce Protocol
OpenAI wants to change the way we shop…and compared to how Google has been doing it traditionally, OpenAI is closing the loop.
And containing the experience within ChatGPT.
OpenAI has ambitious plans for ChatGPT, building on the conversational power of AI, much like an enhanced Google search, but with deeper personalisation.
Shopping Research turns product hunting into an interactive chat, while Instant Checkout closes the loop by enabling purchases right within the interface.
A future where AI agents handle end-to-end commerce.
But, not all features are fully rolled out, and limitations like regional availability persist…
How Shopping Research Works
Shopping Research, announced on November 2025, is a dedicated mode in ChatGPT designed for high-consideration purchases where details matter.
Conversational Discovery
Start with a prompt like “I’m looking for a new mobile phone” ChatGPT asks clarifying questions (e.g., budget, features) and refines in real-time based on your feedback.
Structured Outputs
Delivers a buyer’s guide with top recommendations, side-by-side comparisons, pros/cons and citations from trusted retailers.
It pulls in prices, availability, specs, reviews and images.
Personalisation
Leverages ChatGPT’s memory (if enabled) to tailor suggestions from past chats, like suggesting eco-friendly options if you’ve mentioned sustainability before.
Under the Hood (This interest me most!)
The flow below is not 100% official, but gleaned from documentation.
Obviously multi-model orchestration an integral part of the architecture. Similar to the OpenAI Deep Research API.
This development has introduced a new model, called shopping research model.
Powered by a version of GPT‑5 mini trained with reinforcement learning specifically for shopping tasks.
Shopping Research operates within ChatGPT’s layered architecture:
Input Understanding & Routing
Use models like text-embedding-3-large and GPT-5-mini-router to classify intent (e.g., shopping query) and route to the right tool.
Policy & Safety Layer
gpt-oss-safeguard checks for safe, non-harmful requests.
Orchestration Engine
GPT-5-Thinking-mini decides if a specialised tool is needed, balancing cost, latency and domain.
Model & Tool Selection
Invokes a post-trained GPT-5-mini variant optimised for shopping via reinforcement learning, plus web browsing for real-time data.
Execution Layer
The Shopping Research tool synthesises info from high-quality sources, ensuring citations and avoiding low-trust sites.
Response Synthesiser
Merges results into a clear, interactive guide.
OpenAI evaluations show 52% product match on multi-constraint queries (e.g., price + colour + specs), rising to 64% in domain-specific tests — outperforming standard ChatGPT Search (37%).
Instant Checkout
While Shopping Research focuses on research, Instant Checkout — launched September 2025 — handles the purchase.
This “Buy it in ChatGPT” feature lets U.S. users complete transactions without leaving the app, turning ChatGPT into a true commerce agent.
How It Works
During a chat (e.g., via Shopping Research or a direct query), eligible products show a “Buy” button. Tap it to review details, confirm shipping/payment, and checkout in taps.
Limitations
U.S.-only for now; multi-item carts and global expansion planned.
Not all products qualify — prioritised by relevance, availability, and Instant Checkout enablement.
Free for users; merchants pay a small fee per sale, but this doesn’t bias recommendations or raise prices.
This closes the loop the original Google experience lacks, reducing friction from search to cart.
Shopping Research
Instant Checkout
Protocols like ACP (live via OpenAI/Stripe) for payment
Safety layers ensure ethical use.
OpenAI’s Agentic Commerce Protocol (ACP)
ACP live and powering Instant Checkout.
Co-developed by OpenAI and Stripe, this open-source standard (Apache 2.0) enables AI agents to orchestrate full commerce flows from discovery, negotiation, payment to fulfilment.
Key Mechanics
Agents (like ChatGPT) interact with merchants via REST/MCP APIs, sharing secure tokens for payments without exposing credentials.
One-line Stripe integration for merchants; businesses apply via ChatGPT’s merchant portal. It’s PCI-compliant, with merchants as the record-keeper.
Fully operational since September 2025 in ChatGPT. AI platforms can adopt it (contact acp@stripe.com); docs at developers.openai.com/commerce.
Google’s AP2
September 2025, Google unveiled the Agent Payments Protocol (AP2), an open framework for secure, agent-led transactions — building on their Agent2Agent (A2A) protocol.
Backed by 60+ partners (e.g., Mastercard, PayPal, Etsy), AP2 aims to standardise payments in an AI-driven world, preventing fragmentation.
Open-sourced on GitHub with specs and reference code, AP2 is ready for developers.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
Announcing Agent Payments Protocol (AP2) | Google Cloud Blog
Learn more about the Agent Payments Protocol (AP2), an open protocol that builds on A2A, Agent to Agent Protocol. AP2…
Agentic Commerce Protocol
The Agentic Commerce Protocol is an open standard that enables programmatic commerce flows between AI agents and…
https://openai.com/index/buy-it-in-chatgpt/
https://openai.com/index/chatgpt-shopping-research/