In Part One, I wrote about the shift in power, and in Part Two, about data contamination. Now we reach a question that will eventually land on the negotiating table, with no one ready to answer it.
A user tells their assistant: "Find me a lightweight laptop with good battery life, below this price, that will arrive by the end of the week. If you find a good option, buy it."
The agent searches several stores, opens a sponsored result, and completes the purchase within its delegated authority. The advertising platform records a click.
Should the advertiser pay for it?
Direct answer
No, but probably not for the reason you think. An agent's click cannot be counted as a human click, but that does not make it worthless. Its economic impact is entirely real; it simply needs to be measured in a different unit, tied to a verifiable outcome rather than an impression or a click.
An agent's click cannot be counted as a human click. Does that make it worthless? No. The agent's economic impact is entirely real. It simply needs to be measured in a different unit.
Not All Non-Human Traffic Is Fraud
The US Media Rating Council has a standard for Invalid Traffic: traffic that should not count toward advertising metrics. It separates non-human activity from valid impressions, clicks, and outcomes. IAB Tech Lab also excludes legitimate crawlers, the tools search engines send to index pages, from billable metrics. The logic is simple: useful work was done, but no human saw the message.
That logic still holds. A shopping agent is not human either, and its interaction should not be sold as human attention.
But it differs from a crawler in one major way: it carries real purchase intent, acts under valid user authority, and produces a real transaction.
We therefore need to separate two things we have treated as equivalent for years:
- Human media activity: a person had an opportunity to see or interact with a message.
- Legitimate economic activity: an interaction is connected to a real mission and a valid economic outcome, even if the actor is not human.
An interaction can be non-human from a media perspective and entirely legitimate from an economic one. The mistake is recording both in the same ledger.
The Click Was Built for a World Where People Saw the Page
The click was never a perfect metric. It was an acceptable proxy for attention. An ad appeared in front of a person, and that person chose to open it.
An agent does not behave that way. In seconds, it can open dozens of pages, read hundreds of products from a data feed, and call several services to eliminate unsuitable options. Opening a page indicates neither interest nor preference, nor even that the product made the final shortlist.
Many agentic experiences involve no click at all. Instead of opening a page, the agent reads structured data and proceeds directly to checkout. OpenAI's Product Feed and Agentic Checkout specifications illustrate precisely this: search, advertising eligibility, and checkout are three independent capabilities.
In this world, the click is more a remnant of the web's old architecture than a meaningful unit of value.
So Where Does the Real Impact Happen?
Suppose an agent receives a sponsored offer. We still do not know whether the ad influenced the decision. Even making the shortlist is not enough, because that event usually stays inside the AI provider's system, where neither the advertiser nor the store can independently verify it.
The right measurement chain is:
Retrieved → Shortlisted → Human Approved / Valid Mandate → Purchased → Settled- Retrieved: the offer reached the agent.
- Shortlisted: the offer entered the shortlist of options.
- Human Approved: a person approved that specific choice.
- Valid Mandate: the agent acted without contemporaneous human approval, but within previously granted authority.
- Purchased: the order was placed and paid for.
- Settled: the order was delivered and was not canceled, returned, or identified as fraudulent.
The two alternative authorization states in the middle are the most important part. Sometimes the agent shows the options to a person, who makes the decision. Sometimes the user has already set a rule: "If this product drops below this price, buy it." In the second case, there is no contemporaneous approval. Validity rests on whether that prior authority can be proven.
Which Events Can Be Sold?
The events in this chain have neither equal value nor equal reliability.
Observed event. The store or media platform records it directly: receipt of an offer, order placement, payment, delivery, or return. This is the strongest level of evidence.
Provider-reported event. The AI provider says a product was considered or shortlisted. This is useful for analysis, but without valid attestation and auditability, it should not be treated like an independently verifiable, billable click.
Outcome eligible for settlement. A valid, delivered order after cancellations, returns, and fraud have been excluded. This is the most sensible basis for pay-per-sale and revenue-sharing models.
Now consider what happens if an agent reviews hundreds of offers in each mission and the provider charges for every review. We have simply created a new version of click inflation, with a more modern appearance and a better name.
A billable unit must be one of three things: valid human exposure, a verifiable economic outcome, or a machine event whose generation and auditing methods are specified in the contract from day one.
An Attributed Sale Is Not an Incremental Sale
This is the most important point in the article, because I think it will be the most expensive mistake of the coming years.
The agent saw the sponsored offer and made a purchase. Did the ad create the sale?
We do not know. The user may have wanted that exact brand and product from the start, with the agent merely shortening the path.
This is an old weakness of attribution, now becoming worse. The AI provider says it included the offer in its response. The media platform says the order was placed in its environment. A third advertising network claims the last human interaction. All three are telling the truth, and none answers the question.
Attribution tells us which platform was recorded along the path. Incrementality asks what would have happened without that platform. They are not the same, and never have been. The attribution gap
There are established statistical methods for measuring this: a control group that does not see the ad, a randomized test, or a comparison with the counterfactual in which the channel is absent. Without these, outcome-based pricing effectively means selling a sale that would have happened anyway, all over again.
IAB's advanced Retail Media measurement guidance emphasizes the same three things: incrementality, randomized testing, and data collaboration. Agentic Commerce did not create this problem. It has simply made last-click attribution less credible.
We Need Two Separate Ledgers
The advertising platform of the future must not mix humans and agents in one report. It needs at least two independent ledgers.
Human media ledger: human impressions and viewability, human clicks, engagement, frequency, unique reach, and sales following human exposure.
Agentic Commerce ledger: offers received, agent provider and type, authentication status, human approval or a valid mandate, orders, deliveries, cancellations, returns, and incremental outcomes against a control group.
If agent traffic enters human metrics, it contaminates click-through rates, conversion rates, audience behavior, frequency caps, and all your predictive models. Worse, you may not notice for a long time, because the charts may actually look better.
The Contract Must Be Explicit
The basic principle is one sentence: no non-human interaction should be sold as human attention without an explicit agreement.
Models worth putting on the table include:
- Payment per impression or click only for valid human exposure
- Diagnostic reporting on machine consideration, with no billing
- A separate fee for referrals from authenticated agents, subject to an auditing standard
- Payment per valid order or revenue sharing
- Settlement after delivery, net of returns
- An additional fee for incremental impact only when the testing method has been agreed in advance
The market is moving in this direction, but the transition is incomplete. Shopify's pilot charges a commission on attributed purchases. That is a step beyond raw pay-per-click, but its attribution still relies on the last click and a seven-day window. The incrementality problem has not been solved; it has only moved.
Five Decisions for an Agentic Model
1. Classify traffic. Humans, crawlers, unknown automation tools, trusted agents, and malicious bots should not be treated as the same type of actor.
2. Keep human metrics isolated. No machine interaction should enter human impression or click metrics without a clear label.
3. Follow the outcome all the way through. An order being placed is not enough. Delivery, cancellation, returns, and fraud must also be reflected in settlement.
4. Separate attribution from incrementality. An attributed sale is not necessarily an incremental sale. If you are paying for the latter, you must measure the latter.
5. Write auditability into the contract. If consideration or shortlisting is the basis for pricing, generation methods, attestation, deduplication, and inspection rights must be clear from day one.
One practical point for the Iranian market: have this discussion at the beginning, not later.
Before agent clicks appear on an invoice, defining them is a technical discussion in which everyone can be reasonable. Once they appear on the invoice, that same discussion becomes a negotiation over revenue one party stands to lose. Nobody wins that negotiation.
From Clicks to Provable Impact
If an agent has clicked an ad, we still know nothing. That click may simply be part of a machine search.
But if a clearly labeled sponsored offer entered the decision, led to a purchase within valid delegated authority, resulted in a delivered order, and a test showed that the outcome was genuinely incremental, then we can talk about economic impact.
The future of advertising will not be built by whoever reports the most agent clicks. The winning platform will be the one that keeps the distinction between human attention, legitimate agent activity, and incremental outcomes auditable.
Take the next step