Choosing Happens in the Chat. Closing Still Happens on Your Site.

$
⏱ 7 min read
👤 Sokos Lee
#Agentic Commerce #AI Agents #Distribution #Checkout #Attribution #AI-Native Commerce #Merchant Strategy

Choosing Happens in the Chat. Closing Still Happens on Your Site.

Thesis: Agentic commerce is not a race to finish the purchase inside the model. Preference is forming in AI conversations while payment, identity, and disputes still settle on merchant rails. Operators who only chase in-chat checkout will lose to those who treat the agent-to-site handoff as a first-class product surface: intent captured upstream, trust closed downstream, attribution fixed so agent-sourced demand is measurable GMV — not a dark funnel.

I am building an AI-native commerce company. I want agents in discovery. I refuse to confuse “the model recommended us” with “we owned the close.”

The Signal: Instant Checkout Met Reality

The loud story for a year was end-to-end agent buy: ask the model, pay in the thread, never open a tab. The quieter story — the one operators should budget against — is that happy-path in-chat checkout did not become the default path.

OpenAI’s Instant Checkout / Agentic Commerce Protocol push was the clearest public experiment: product cards in chat, Stripe-powered pay, open protocol so merchants could plug in. Then the market forced a reshape. Merchant adoption lagged (on the order of a dozen live merchants against millions of storefronts). Product data went stale inside the conversation. Conversion inside the chat underperformed sending people to a real storefront. The pragmatic settlement that is showing up in operator chatter: discover in AI, buy on your own site — or inside a merchant-controlled app surface that still owns identity, inventory truth, and refunds.

That is not a defeat for agentic commerce. It is a separation of concerns the industry had to learn the hard way.

LayerWhere it is movingWho owns the failure mode
Intent + shortlistChat, shopping agents, “buy for me” threadsModel / agent host
Price, stock, policy truthMerchant catalog + feedsYou
Payment + fraud + disputeMerchant / PSP rails (or agent pay with your authority model)Whoever settles the charge
Preference formationUpstream conversationShared — but you only win if you show up correctly

A useful founder cut from the last day of discourse: in-chat purchase paths that tried to keep everything in the model converted roughly 3x worse than routing people to the merchant site, while ChatGPT-style referral traffic still converted materially higher than organic search because people arrived already decided. Choosing moved. Closing did not fully follow.

Founder translation: the agent is not your new checkout page. The agent is your new front door and sales floor. Your site (or app) is still the register.

Preference Is Upstream. Margin Is Downstream.

I already argued that default agent access is not default agent preference, and that machine-readable merchants eat brand-only merchants. Stack the next layer:

Once an agent can see you, the scarce work is not “can we enable ACP/UCP?” It is:

  1. Win the shortlist inside the conversation (structured truth, claim integrity, policy fields agents can trust).
  2. Survive the handoff when the human (or agent) leaves chat for checkout (deep links, stable SKUs, session continuity, not a cold homepage).
  3. Close without wrecking unit economics (fraud, returns, attribution, support load from agent-shaped intent).

China’s agent shopping rails — Alipay-class volumes measured in tens of millions of agent transactions per week in peak windows — prove demand can scale when payments, identity, and merchant systems are already coupled. Western discourse is still arguing demos of end-to-end auto-buy while the actual winning pattern looks more hybrid: agent as buyer brain, merchant as settlement and service system.

If you only measure “in-chat GMV,” you will under-invest in the path that actually converts. If you only measure “site conversion,” you will under-invest in the conversation where preference was formed — and you will not know which feed, claim, or SKU truth won the shortlist.

The Handoff Is a Product, Not a Redirect

Most merchant “AI commerce” plans still look like this:

Enable protocol → hope agents recommend us → celebrate a spike in bot traffic.

That is not a product. That is a toggle.

A real handoff product has five contracts:

1. Stable machine identity for the SKU.
Not a pretty PDP URL that breaks when marketing renames a collection. A canonical product ID, variant map, and inventory snapshot the agent can re-check at handoff time.

2. Intent payload, not just a link.
Size, color, quantity, budget cap, delivery window, gift constraint — whatever the chat already decided — must arrive as structured context. If the human has to re-answer three questions on your site, you burned the agent’s work and your conversion edge.

3. Authority boundary at money movement.
Deciding to buy is not the same as being authorized to pay. The timeline is full of agent wallets, x402 rails, and “policy gates before spend.” Even if the human pays on your site, you still need clear rules for agent-assisted carts: who set the cart, what was promised in chat, what can be refunded when the promise was wrong.

4. Dispute path that matches agent speed.
Payment rails and identity specs are getting loud. The quiet omission keeps showing up: what happens when two parties disagree that the work (or the goods) were as described. Agents do not wait 344 days for a civil dispute average. Your Monday-morning version is boring and mandatory: refund SLAs, chargeback evidence packs, and audit logs that reconstruct “what the agent claimed” vs “what the catalog said.”

5. Attribution that finance will accept.
Agent referral that converts 2x organic is worthless if finance cannot tag margin, returns, and CAC against it. If “ChatGPT / shopping agent” is a mystery blob in analytics, you will starve the feed quality work that creates the preference.

Think Big / Step Small / Do Smart

Think big.
The front door of commerce is becoming a conversation. Brands optimized only for human landing-page theater will still look premium — and lose the buyer who already chose elsewhere. The company that owns the preference-to-settlement bridge owns durable distribution in the agent era.

Step small (Monday morning).
Pick your top 20 SKUs by margin. For each:

  1. Fix agent-facing fields: price, stock, ship-by, return window, materials/claims that get challenged.
  2. Build one handoff URL pattern that accepts variant + quantity + optional agent session id (even if the session id is only logged for now).
  3. Tag that traffic in analytics as agent_referrer (or equivalent) and set a weekly review: conversion rate, AOV, return rate vs organic.

Do not rebuild checkout inside a model this week. Instrument the bridge.

Do smart.
Budget inference and agent experiments against accepted tasks, not demos. Kill any agent surface that cannot answer: where did preference form, where did money settle, and who is on the hook when it is wrong? Prefer merchant-controlled close until your verification stack (receipts, authority, dispute) is as strong as your catalog.

The Claim Worth Arguing

In-chat checkout is a feature. The agent-to-merchant handoff is the system.

If you disagree, show me a merchant where fully in-model settlement beat merchant-controlled checkout on conversion and return rate and dispute cost for more than a pilot SKU set — without starving catalog truth or authority controls. Best counterexample wins.

I am not anti-agent pay. I am anti-demo theater that skips the boring rails: inventory truth, intent continuity, spend authority, dispute, and attribution. Preference can live in the chat. Until your handoff is a product, closing still happens on your site — and that is an advantage if you design for it.

Sources