Agent-Native Commerce: What Happens When AI Does the Shopping
Something is happening that most of the commerce industry hasn’t fully registered yet. The shopping journey — the process of discovering, evaluating, and choosing products — is being automated. Not in the future. Now.
When a user asks ChatGPT “what’s the best project management tool for a 10-person remote team?”, they’re not searching. They’re delegating. They’re handing the discovery, comparison, and initial recommendation to an AI agent, and they’re trusting the answer.
That trust is warranted more often than not. And it’s reshaping how products get found and bought.
This is already happening
Consider what’s live today, in early 2026:
ChatGPT has over 900 million weekly active users. Its search and browsing features let it recommend specific products, link to merchant sites, and compare options across categories. OpenAI reports that commercial queries — “best X for Y”, “compare A vs B”, “where to buy Z” — are among the fastest-growing query types.
Perplexity processes 500 million queries per month and is specifically designed to answer questions with sourced citations. It already generates comparison tables, product recommendations, and buying guides. Its shopping features allow users to browse and purchase directly within the interface.
Google AI Overviews appear for over 30% of search queries and reach 1.5 billion users monthly. When a user searches for a product category, they increasingly see an AI-generated summary before any organic or paid results. The click-through to merchant sites is declining because the AI already provided the answer.
Agent frameworks like AutoGPT, CrewAI, and LangChain allow developers to build autonomous agents that can research products, check prices, read reviews, and make purchasing recommendations — all without a human touching a browser.
This isn’t experimental technology. It’s mainstream consumer behaviour.
What “agent-native” actually means
A product, service, or system is agent-native when it’s designed from the ground up for AI agents to interact with — not retrofitted from a human-first design.
Think about it this way:
Mobile-native didn’t mean shrinking a desktop website to fit a phone screen. It meant rethinking the interface, the interaction patterns, and the information architecture for how people actually use phones — touch, vertical scrolling, app-based workflows.
Agent-native is the same shift, but for AI. It means:
Structured data over visual layouts: Agents don’t see your beautiful product page. They parse data. If your product information isn’t available as structured, machine-readable data, agents can’t work with it.
Protocol-level attribution over cookie tracking: Agents don’t carry cookies or click tracked links. Attribution has to be embedded in the data exchange, not in the browser session.
Programmatic access over web scraping: Agents that scrape merchant websites are doing it because there’s no better option. Scraping is brittle, expensive, and often blocked. Agent-native commerce provides clean data through structured feeds or APIs.
Recommendation tracking over click tracking: The moment of influence in agent commerce is the recommendation, not the click. If an agent recommends your product and the user buys it, that recommendation needs to be trackable and attributable.
The three layers of agent-native commerce
Agent-native commerce isn’t a single product or feature. It’s an infrastructure stack with three layers:
Layer 1: Offer data
Merchants need to publish their offers — products, pricing, commission rates, terms — in a structured format that agents can discover and consume. This is the equivalent of a product feed, but designed for agent consumption rather than ad platforms.
Today, merchants publish product feeds to Google Shopping, Amazon, and affiliate networks. Each platform has its own format, its own requirements, and its own walled garden. Agent-native commerce needs a single, open format that any agent can access.
Layer 2: Attribution protocol
When an agent recommends a product and a sale happens, there needs to be a reliable way to trace the conversion back to the recommendation. This can’t depend on cookies, browser sessions, or pixel fires. It has to work at the protocol layer — embedded in the transaction data itself.
This is the hardest problem to solve and the most important. Without reliable attribution, merchants can’t measure which agents drive value, and agents can’t be compensated for their recommendations. The incentive structure that powers affiliate marketing collapses.
Layer 3: Commission distribution
Once a conversion is attributed to an agent’s recommendation, the commission needs to be calculated and distributed. This includes multi-touch attribution (what if multiple agents influenced the purchase?), commission rules (percentage vs flat rate, category-specific rates), and settlement (when and how agents get paid).
Why the incumbents haven’t built this
The major affiliate networks — Awin, CJ Affiliate, Impact, ShareASale, Rakuten — are well-resourced companies with deep industry relationships. They could, in theory, build agent-native infrastructure. They haven’t, for three reasons:
Revenue model conflict: Existing networks earn fees on every tracked transaction. Their infrastructure is optimised for the click-cookie-convert model. Rebuilding for agent attribution means replacing the system that generates their revenue — while it’s still generating revenue.
Publisher relationships: The networks’ primary customers are publishers — bloggers, comparison sites, influencer platforms — who create content and place tracked links. These publishers are threatened by AI agents, which do the same job faster and at lower cost. Building tools that empower agents could alienate existing publishers.
Technical debt: Most affiliate platforms are built on tracking technology from the mid-2000s. Server-side tracking was added later, but the core architecture still assumes browser-based sessions. Retrofitting for agent commerce requires rethinking the foundation, not adding a feature.
This is the classic innovator’s dilemma. The incumbents are too successful with the current model to build the next one.
The Hyperoptic parallel
When Hyperoptic launched in the UK in 2011, the broadband market was dominated by BT, Sky, and Virgin Media. These companies had millions of customers, national infrastructure, and decades of brand recognition. Nobody gave a full-fibre-only challenger a chance.
But the incumbents had built their networks on copper. They were invested in sweating existing assets, not replacing them. Full fibre was technically superior but required building new infrastructure from scratch — something the incumbents could afford but had no incentive to do while copper still worked.
Hyperoptic built for where broadband was going, not where it was. Today, the UK government’s broadband strategy is built around full fibre, and the incumbents are spending billions to catch up.
Agent-native commerce is in the same position. The current infrastructure works for browser-based shopping. It doesn’t work for agent-mediated commerce. The companies that build the new infrastructure now — before the incumbents are forced to — will define the next era.
What happens next
Over the next 24 months, AI agents will go from recommending products to completing transactions. The technology is already there — agent frameworks can browse, evaluate, and execute purchases. What’s missing is the infrastructure layer: structured offer data, protocol-level attribution, and automated commission distribution.
The companies that provide this infrastructure won’t just participate in agent-native commerce. They’ll be the foundation it runs on.
Previous: Affiliate marketing wasn’t built for AI agents. Next: Why agent attribution needs an open standard.