No Eyes on the Ad

The headline post on Business of Fashion last Friday was How AI Shopping Agents Are Reshaping Fashion’s Ad Economy, a piece about how AI agents are distorting ad metrics.

I was quoted in the piece on emotional storytelling — that as AI search shortens the time shoppers spend on brand sites, the moments when a human does show up matter more.

But reading the rest of the piece made me realize I'd been missing something in my own coverage. I've spent some time now arguing that agentic commerce is a structural threat to retail media — ad surfaces disappearing, first-party data losing some signals. What I hadn't focused on is a more immediate, messier problem: AI agents are already out there clicking around. And brands are paying for it.

Here's how it works.

An AI agent navigates a retailer's website — looking up product availability, checking prices, sometimes even trying to complete a purchase
The site often can't tell it's not a human
The agent's activity triggers an ad impression
Under a CPM (impression) model, the advertiser gets charged. But no human ever 'saw' the ad.

This is the logical endpoint of a problem Jason O'Toole, Head of Connected Commerce and Media at Gildan, was describing just last week. "The quality of impressions is not standardized," he said on a podcast. "A lot of times if you scroll past an ad... the consumer may never see it, but it will still count in certain platforms as an impression."

At least a human scrolling past an ad could have seen it. A bot has no eyes.

Retailers know that a marketplace model can dramatically boost product assortment, shopper engagement, and total revenue.

But, to get the most out of your marketplace, you need an ad tech solution that can really engage sellers.

Mirakl Ads is powering the future of retail media for leading retailers — to activate both 3P sellers and 1P brands.

Learn more

This isn't hypothetical. BoF cites cybersecurity firm Human Security, who say that agentic traffic grew by more than 1,300% in the first eight months of 2025. It's still a small share of overall web traffic — only about 2.3% of users are actively using agents for checkout right now. But the direction of travel is clear, and the ad industry's measurement infrastructure wasn't built for this.

And DoubleVerify's VP of product management described a lab test where an AI assistant tried to buy shampoo that was out of stock. Rather than move on, the agent kept clicking — 10, 20 times — working through different sections of the site trying to complete the task. Every one of those clicks is a potential chargeable ad event.

But the ad spend problem is only half of it. Those clicks don't just trigger impressions — they feed the retailer's behavioral data pool. Search queries, product page visits, category browse patterns: all of it gets logged as member intent.

For a retailer with a logged-in user who has an agent working on their behalf, every one of those fruitless clicks looks like genuine consumer interest. That member now appears to be a heavy shampoo shopper. They get added to a specific audience segment. Brands buy against it. The targeting is wrong from the start.

Agent traffic doesn't just waste ad dollars. It poisons the data pool that justifies the whole model of both onsite and offsite ad dollars.

The response emerging from companies like Human Security and DoubleVerify is something called "cost-per-human" — essentially, only billing brands when a verified human viewed or clicked an ad. It sounds obvious. The catch is that distinguishing human from agent traffic is.... hard.

What does this mean for RMNs specifically?

Here are a few implications and things worth watching:

The measurement problem just got another layer. RMNs are already under pressure on attribution and incrementality. Now add a more basic question: are your reported impressions and clicks coming from humans at all? And harder still — is the behavioral data underlying your audience segments accurate, or has it been quietly corrupted by agent activity?
"Cost-per-human" could be good for the industry — but painful in the short term. If it takes hold as a billing standard, it forces a more honest accounting of real consumer attention. But it would almost certainly reduce reported metrics, which is uncomfortable for any RMN trying to justify its ad rates to brand partners.
The data problem is harder to fix than the billing problem. You can change how you charge for impressions. Cleaning up a corrupted behavioral data pool is a different challenge entirely — especially if you don't know how long agents have been muddying it.

Read more from me on related topics: