The Real Take Rate of Agentic Checkout Is Your Customer Record

Take an $80 order. On your own site, acquired through paid social, the unit economics run like this: $32 in landed cost of goods, $9 to pick, pack, and ship, $2.62 in payment processing, $28 in blended acquisition cost. You keep $8.38, roughly 10 percent of revenue. Now route the identical $80 order through ChatGPT Instant Checkout. No ad dollar touched it. Model the platform fee at 2 percent, since OpenAI says only that merchants pay "a small fee on completed purchases" and has never published the rate. You keep $34.78.

Per order, the agentic sale looks four times better than paid. That framing is where most of the bad decisions of 2026 will get made.

The line the table hides

The first six rows price the order. The last two price the customer, and that is where agentic checkout either earns its keep or quietly bleeds you.

An on-site buyer hands over an email address, a shipping address, and marketing consent. Assume 30 percent of those buyers purchase again within twelve months, each repeat order contributing $36.38 before marketing. Expected future contribution: $10.91 per first order. An agentic buyer completes the purchase inside the chat. Under the Agentic Commerce Protocol (ACP), the open standard OpenAI and Stripe published, you remain the merchant of record and receive what you need to fulfill and support the order. The conversation, the browsing intent, and the ongoing relationship interface stay with the agent. Marketing consent does not ride along. If your reachable repeat rate drops to 10 percent, expected future contribution falls to $3.64.

That $7.27 gap equals 9.1 percent of the order. Add the modeled 2 percent fee and the effective take rate on an agentic sale is roughly 11 percent. The correct comparison set is suddenly Amazon's 15 percent referral fee, and it deserves the same cold-eyed evaluation. If this arithmetic is unfamiliar, start with what contribution margin actually measures, then run your own numbers through our CAC and LTV calculator. Winning back the opt-in after an agentic sale is a first-party data problem, and our first-party data playbook covers the mechanics.

One more turn of the screw. Everything above assumes the agentic order is incremental, meaning the buyer would never have reached you otherwise. If the agent merely intercepted a customer already headed to your site, the agentic order is strictly worse: $8.87 lighter per order in this model, because you paid a fee and forfeited a customer record on demand you already owned. The entire decision hinges on incrementality, which is precisely the question a per-order dashboard cannot answer.

Attribution breaks in a specific way

An Instant Checkout purchase never loads your site. No pageview, no pixel, no session. GA4 records nothing. Last-click assigns nothing. Meanwhile, if that buyer saw your Meta ad last week, Meta may still claim a view-through conversion on an order it did nothing to close, inflating platform ROAS at the exact moment you need it honest.

The fix is structural. ACP orders arrive through your commerce backend carrying their own order source, so treat "agent" as a channel from the first order: separate P&L line, separate repeat-rate cohort, separate margin target. Above the channel level, this is a textbook argument for media mix modeling over touch-based attribution, because MMM never depended on the click trail that agents just deleted. And if your reporting stack still assumes every conversion has a referrer, revisit how attribution models work and where they fail before agent volume grows past rounding-error size.

The bull case, taken seriously

The demand signal is real and measurable. Adobe reported that traffic to retail sites from generative AI tools rose 693 percent year over year through the 2025 holiday season, and its analytics team has since found that AI-referred shoppers convert better than the site average. Salesforce measured AI and agents driving 20 percent of retail sales, about $262 billion, over the 2025 holidays. The rails are consolidating fast, too: Google's rival Agent Payments Protocol launched with dozens of payments partners, including Mastercard and PayPal, and has since moved into an open standards body. This is infrastructure now.

There is also an absence penalty. When an agent can complete a purchase from a competitor in one turn but can only describe your product, the default recommendation drifts away from you. And the raw margin math is forgiving: at $34.78 contribution versus $8.38 on a paid order, an agentic sale can absorb a substantial LTV haircut before it underperforms paid acquisition. For a brand whose marginal CAC runs 35 percent of AOV, an 11 percent effective take rate on genuinely new demand is cheap volume.

The bear case, from the people with leverage

Walmart signed on to Instant Checkout in October 2025. Within months, trade coverage tracked a pivot: the retailer rebuilt its ChatGPT presence around its own Sparky assistant rather than routing transactions through the platform's checkout. Read that carefully. The retailer with the most negotiating power on earth chose to own the experience layer inside the agent rather than hand over the transaction. Most merchants will never get that option, which is exactly why the precedent matters.

The rest of the bear case is straightforward CFO material. The fee is undisclosed and changeable at the platform's discretion. Ranking is unauditable; OpenAI states that checkout participation does not influence product results, and you have no way to verify it. Price integrity erodes when agents comparison-shop across every participating merchant simultaneously. And every point of GMV that migrates into an agent is a point of concentration risk on a counterparty that can rewrite terms in a product update.

When opting in beats protecting the funnel

The decision reduces to three variables: your effective take rate (platform fee plus LTV haircut, expressed as a share of AOV), your marginal CAC, and your honest read on incrementality.

Where the numbers in the $80 model come from

The table is an illustrative model rather than observed data. The order value, the COGS share, the fulfillment cost and the processing fee are round assumptions chosen so the arithmetic is easy to follow. The repeat rates are assumptions too: 30 percent on-site and 10 percent for agentic buyers are plausible planning values, and the 2 percent fee is a placeholder because OpenAI has published no rate. The only external figures in this post are the Adobe traffic-growth and Salesforce holiday-sales numbers cited beside them, and both are platform-level aggregates rather than results from any single merchant. On engagements we start from a version of this table and replace every row with the account's own numbers: actual COGS by SKU, the repeat rate measured from the order file, and marginal rather than blended CAC from the last quarter. What the model cannot tell you is whether a given agentic order was incremental. Only a holdout or a geo test answers that.

Where the opt-in rule gives the wrong answer

The decision rule is sound, but four inputs go wrong often enough to flip the verdict.

• Blended CAC in the marginal CAC slot. The most common first-timer mistake. Blended CAC folds in cheap brand and organic traffic, so it sits well below the cost of the next paid customer; compare the effective take rate against it and agentic checkout looks worse than it is, and a funnel that should have been opened stays shut. Use the cost of the last cohort you actually bought.
• A repeat rate that cannot see repeats. If you measure agentic repeat by matching emails in your ESP, you will undercount, because a repeat order placed through the same assistant never shows up as a known customer. The LTV haircut then looks larger than reality. It depends on whether ACP orders carry a stable customer reference into your backend; check that before you trust the cohort.
• Return-heavy categories. The fee is refunded on returns, but pick, pack and return shipping are spent either way, and the model's fulfillment line assumes the order sticks. In practice apparel and gift-led categories need a return-adjusted contribution row before the comparison means anything.
• Testing at rounding-error volume. A geo holdout cannot detect lift when agentic orders are a fraction of a percent of volume, so the "provably incremental" condition is unmeetable for months. The trade-off: opt in as a small, tagged pilot while volume is too low to test, and set the decision date in advance rather than waiting for a result that cannot arrive.

The exception to all of this is the replenishment brand. In our work with Gold Collagen, a luxury DTC brand selling a daily liquid collagen, unit economics hinge on repeat purchase; email and lifecycle automation alongside paid growth and performance media returned 4x on ad spend and lifted customer retention by 30%. A business in that position hands over exactly the relationship its lifecycle program depends on when the first order completes inside an assistant, which is why the "protect the funnel" verdict above is the default there.

Sources

• OpenAI: Buy it in ChatGPT, Instant Checkout and the Agentic Commerce Protocol
• Adobe: AI traffic surges across industries, retail sees biggest gains
• Salesforce: 2025 holiday shopping data