Who Owns the Receipt?: Personal Agents and the Attribution of AI Revenue

Personal agents remove the click and keep the sale

Google is the main source of verified, attributable sales today because attribution was built around the click. Between August and October 2026, five personal agents (xAI's Grok Bot, Meta's Muse, Instinct, Wajo's Fo and OpenAI's Dots) began completing purchases for consumers on payment rails that record every transaction. AI revenue is about to become attributable. The record of it, however, will sit with the agent and the payment processor before it reaches the brand, and that changes what a brand has to measure.

The argument in this note is narrow. The sceptics on LinkedIn are correct about the current numbers. They are using an instrument designed for referral traffic to measure a channel that is moving to completed transactions. When the agent buys, the only part of the sale a brand can still influence is whether it was selected.

The change is easiest to see as two journeys side by side.

Step	Search era	Agent era
1	Person types a query into Google	Person gives a request to an agent
2	Person scans a results page	Agent compares options, out of the brand's sight
3	Person clicks through to a website	Agent selects one option
4	Person buys; cookie or referrer records the click	Agent buys; payment rail records the transaction
What can be attributed	The click	The selection and the transaction

What the sceptics are measuring

By every referral-based measure available in October 2026, AI is still a small source of sales. Similarweb's State of Ecommerce 2026 puts generative AI at 0.4% of retail ecommerce visits. A Search Engine Land panel of 75 ecommerce sites found AI assistants contributed about 1.6% of search traffic and revenue. Visibility Labs found non-branded organic search generated about 68 times more revenue than ChatGPT referrals.

The agentic checkout numbers are thinner still. The State of AI Report predicted that a major retailer would attribute more than 5% of online sales to agentic checkout, and its own tracker says that bar has not been cleared. Merchants also cannot see much of what does happen. In PYMNTS Intelligence's October 2026 Merchant AI Report, no merchant selling through its own site or app could fully and consistently connect AI-driven visits to completed purchases.

On these figures, calling AI de minimis and unattributable is accurate. The weakness is in the instrument. Each figure counts a human who clicked from an AI answer to a merchant site and was then tracked to a sale. That is the search model of attribution applied to a new referrer. It cannot see a purchase that an agent completes, and agent-completed purchases are the part of the channel now growing.

Five agents in eight weeks

The consumer personal agent went from private beta to mass market between August and October 2026. Each of the five agents below can complete a purchase, and at least three of them pay through a single-use card or a wallet that records the agent's part in the transaction.

Date	Agent	How it buys	Scale signal
29 Sep 2026	Dots (OpenAI)	Own cloud computer and browser; user sets which actions need approval	ChatGPT Pro and Business Premium only at launch
28 Sep 2026	Fo (Wajo)	Own phone number, email address and single-use virtual cards	Used in 107 countries; 10x growth since launch
8 Sep 2026	Muse (Meta)	Link by Stripe, one-time card per purchase, with Link's purchase protections	10.9m downloads, 2.5m US daily users by 16 Oct (Sensor Tower via TechCrunch)
28 Aug 2026	Grok Bot (xAI)	Stripe Link wallet; human approval on every payment	Beta since 11 Aug 2026
Aug 2026	Instinct	By text message or phone, across connected apps	Approaching $1bn annual transaction volume, per its founder

Instinct's figure is the clearest early signal of transaction volume. Founder Noah Shinn told Fortune the company was approaching $1bn in annual transaction volume while still invitation-only and with no marketing spend. Muse is the clearest signal of reach. Sensor Tower counts it as only the 12th app since 2015 to pass 10 million US downloads within 40 days, helped by Meta directing 39% of its October advertising to it.

Why agent-completed revenue is attributable

When an agent completes a purchase, three parties hold a record that a click-based funnel never produces. The payment layer knows which agent paid. The agent knows what it was asked, what it considered and what it chose. And under the new identity protocols, the merchant can know that an agent was present.

The payment record comes first. Muse checks out with Link by Stripe, which issues a one-time card for each purchase so the agent never handles the user's real card details. Grok Bot completes purchases through the same Link wallet. Wajo gives Fo its own single-use virtual cards. A card issued for a single agent purchase is an attribution event in its own right: it ties one agent, one merchant and one amount together at the moment of sale.

Agent identity is arriving in parallel. On 6 October 2026, Sierra and Meta announced the Personal Agent Protocol (Poppy), with Genesys, Instinct, NiCE, Rocket, Shopify, Stripe and Walmart as founding participants. Under Poppy, the personal agent identifies itself so a company always knows when it is acting for a customer. A further 35 design partners joined when the draft was published on 9 October.

Taken together, these mechanisms make an agent sale more attributable than a ChatGPT referral has ever been. The referral could be lost to a stripped header, a second device or a later visit through Google. The agent purchase is recorded at the rail.

Who holds the record

Attribution becoming possible does not mean the brand can see it. Today the record sits mainly with the agent platform and the payment processor. In a Visa-commissioned PYMNTS survey of 1,185 merchants, only 23% could clearly identify both AI-driven traffic and the purchases that follow, and a further 21% could see the traffic but not connect it to orders.

The agents themselves add to the gap. Profound's research found that Muse and Instinct do not identify themselves by name when they browse merchant sites. This is the reason Amazon gave when it blocked Muse from Amazon.com on 20 September, twelve days after launch: Meta had given no notice and the agent did not identify itself while browsing. In financial services, JPMorgan has barred outside agents such as Muse from customer accounts.

The brand's position is therefore asymmetric. When it wins, it sees an order, perhaps flagged as agent-originated. When it loses, it sees nothing. The shortlist the agent built, the criteria it applied and the competitor it chose never reach the brand's analytics. A referral funnel at least recorded the visitor who left. An agent funnel records only the winner.

Access, Choice, Execution

The record also shows where a brand can lose. AIVO Journal's "Chosen, Then Blocked" divides an agent purchase into three gates (Access, Choice and Execution), and each failed in public within weeks of the agents launching.

Access failed when Amazon blocked Muse. Execution failed in Axios's head-to-head test, where Muse stalled at Target's sign-in page after choosing the product. Choice is harder to observe. Muse keeps a shopping profile of each user's tastes by product category, so two people making the same request can receive two different winners.

Choice is also the gate a referral funnel can never see, because the comparison happens inside the agent. It has to be measured directly, as the share of relevant decisions in which an agent selects the brand. AIVO calls this measure AI Win-Rate™.

Every category, not only retail

The referral debate has been conducted almost entirely in ecommerce terms. The agents are not confined to retail. Instinct's early users report planning road trips, buying weekly groceries and concert tickets, and cancelling subscriptions. Fo books appointments, coordinates vendors and calls businesses on the user's behalf. At the Poppy launch, Muse obtained a mortgage preapproval letter from Rocket within moments.

Merchants already expect the shift to show up in their own numbers. In the same Visa-commissioned survey, 38% expect agent purchases to exceed 15% of sales within two years. That expectation sits uneasily beside a measurement base in which three merchants in four cannot yet see agent purchases at all.

For a brand in any category, three questions follow. Can each major agent reach and read the brand, and does it identify itself when it does? When an agent compares the brand with its competitors for a real customer request, how often does it select the brand, and on which criteria does it lose? Once selected, does the purchase or booking complete? Referral analytics answers none of these. Each needs to be tested directly, agent by agent, before the revenue arrives in volume.

Limits of the argument

The timing is the weakest part of this thesis. Consumer trust is the binding constraint: Mastercard research found only 10% of consumers currently accept an agent completing a purchase autonomously, and Muse and Grok Bot both require approval before they pay. Reach is also narrower than the download figures suggest. Muse is US-only, Dots launched for paying Pro and Business Premium subscribers, and Instinct remains invitation-only. Early momentum is contested too: Muse's daily downloads fell 8.1% in the first week of October week on week, a dip Hunterbrook cited in disclosing a short position in Meta.

The infrastructure is incomplete. Poppy does not yet handle payments, and Sierra lists them as a future extension. Payment processors hold the cleanest records, and nothing yet obliges them to share agent-level data with merchants or brands.

Several of the figures above are company statements or third-party estimates. Instinct's transaction volume comes from its founder, and Muse's user numbers are Sensor Tower estimates rather than Meta disclosures. The 2026 holiday season will be the first real test. Adobe forecasts AI-referred traffic to US retail sites up 130% year on year, but does not say what share of traffic or revenue that represents.

Conclusion

For twenty years, marketing attribution has answered one question: which click led to the sale. Personal agents remove the click and keep the sale. What is left to measure is the moment the agent chose, and the brands that measure it first will know why they are winning or losing AI revenue while their competitors are still counting referrals.