AI agents can manage frequency capping across several negotiated deals only where one system sees every impression against the same identifier and applies one cap to all of them. When deals run with different sellers, a cap set inside each deal limits exposure within that deal, not across the campaign. A buyer's agent can enforce a campaign-wide cap on the impressions it bids on itself. It cannot enforce one on impressions it never sees, on users it cannot recognise, or on a household it can only count as separate devices.

TL;DR. Frequency caps break across negotiated deals because each seller counts only its own impressions, and identifiers rarely match from one seller to the next. On connected TV, the household is the natural unit; on mobile and desktop, the device is, and the two do not convert cleanly. A buyer's agent needs a mandate that sets one campaign-wide cap, an identifier hierarchy and a rule for unrecognised users. Deal terms then have to say how each seller's cap interacts with the buyer's, and who bears any shortfall that capping causes on a guaranteed deal.

Why do frequency caps break across separate deals?

Frequency caps break across separate deals because each deal is counted in a different place. A cap applied by a seller limits how often a user sees the buyer's ad on that seller's inventory. A second seller applies its own cap to its own inventory, with no knowledge of the first. Neither is wrong, yet the campaign overshoots.

Take a hypothetical example. A buyer agrees three deals for a four-week campaign: a private marketplace deal with a news publisher, a programmatic guaranteed deal with a streaming service and a preferred deal with a sports site. The intended cap is three exposures per person per week. Each seller applies three per week inside its own deal. A person who reads the news site, watches the streaming service and checks sports scores can see the ad nine times a week, three times the plan, while every seller reports full compliance.

The buy side has its own version of the problem. If deals run through more than one buying seat, or an agency and a brand both buy the same audience, each seat counts separately.

How does household versus device capping change the result?

Household and device capping measure different things, and choosing one changes what a cap of three actually means. Connected TV inventory is usually identified at household level, because one screen is shared. Mobile inventory is identified by device, which usually means one person. Desktop identification has become patchier as browsers limit third-party identifiers.

A household cap of three on CTV allows three exposures on the living-room screen, however many people are watching. A device cap of three on mobile allows three per phone. A household of two adults with two phones and a television could therefore see the ad nine times under caps that each look sensible. Converting between the two needs a household graph, a model that links devices to households, and every graph is an estimate.

The buyer has to choose which unit defines the campaign cap and how other units convert to it. That choice belongs in the mandate; an agent left to infer it will use whichever identifier each deal supplies.

What should the mandate say about frequency?

The mandate should state the cap once, for the whole campaign, and tell the agent how to apply it across every deal. A working checklist:

Campaign cap: the maximum exposures per person or household per period, across all deals combined.
Unit and hierarchy: which identifier counts first (for example, household for CTV, device elsewhere) and how they are linked.
Unrecognised users: what the agent does when an impression carries no usable identifier, such as excluding it or limiting it to a set share of budget.
Priority: which deal gets the remaining exposures when a person nears the cap, for example the guaranteed deal before the preferred deal.
Reporting: a frequency distribution across all deals, not an average per deal.
Escalation: what overshoot triggers human review.

The unrecognised-user rule is the easiest to forget. Without it, an agent applying a strict cap to known users can push spend toward unknown ones, where no cap applies at all, and report a campaign that looks perfectly capped.

What do the deal terms need to say?

The deal terms record how each seller's cap and the buyer's cap work together. The mandate binds the buyer's agent; only the deal binds the seller. Agree five points explicitly:

Seller-side cap: whether the seller applies its own cap, at what level and in what unit.
Identifier passing: which identifier the seller includes in bid requests, so the buyer's agent can count consistently.
Shortfall: on a programmatic guaranteed deal, the buyer is expected to take the agreed volume. If the buyer's cap causes it to decline impressions, the terms should state whether the resulting under-delivery counts against the guarantee or against the buyer.
Multiple deals with one seller: whether the seller's cap applies across all of the buyer's deals with that seller or per deal.
Frequency reporting: whether the seller will share frequency data so the buyer can check its own counts.

Standards help with the surrounding steps. According to IAB Tech Lab, AAMP's support for the Deals API covers programmatic guaranteed, preferred deals and private marketplaces, and its OpenProposal specification standardises how publishers describe products and packages. AdCP, the Ad Context Protocol from AgenticAdvertising.org, includes a check_governance task that validates budget, brand safety and targeting rules before a buy runs. A protocol can carry a frequency term. It cannot count impressions nobody can match.

What can a buyer enforce alone, and what can it not?

A buyer's agent can enforce a cap on every impression it bids on through its own seat, using the identifiers it receives. That covers private marketplace and preferred deals, where the buyer decides impression by impression, and much of a guaranteed deal, subject to the shortfall terms.

A buyer cannot, on its own, cap exposures bought through another seat or by another party, placements the publisher serves directly without a bid on each impression, or users who arrive without a usable identifier. It cannot deduplicate households without a household graph, and it cannot verify a seller's cap without the seller's delivery data.

That is why frequency is a negotiated term, not a setting. Where the buyer cannot count, the seller has to commit, and both sides need the same record of the commitment.

How can a buyer check that caps held?

A buyer checks caps by comparing the frequency distribution it observed with what each seller reports, against the terms both agreed. If the records differ, the question is which record governs. On a marketplace such as Alkimi, which works within IAB Tech Lab's AAMP framework, the frequency terms sit on the DealSheet: written at agreement, timestamped and retrievable by buyer and seller independently, with people keeping approval rights at the deal level.

Frequently asked questions

Can a DSP frequency cap work across private marketplace deals?

Yes, for impressions bought through the same seat where the DSP recognises the user. It cannot count impressions bought elsewhere or served to users without a usable identifier.

Is household or device frequency capping better?

Neither is better in general. Household capping suits shared screens such as connected TV; device capping suits personal devices. The mandate should name one unit for the campaign cap and state how the other converts to it.

Why does capping cause under-delivery on programmatic guaranteed deals?

A guaranteed deal expects the buyer to take the agreed volume. If the buyer's cap declines impressions, delivery falls short, so the terms should say who bears that shortfall.

Can an AI agent enforce frequency caps on its own?

Only on impressions it bids on and can identify. Caps across sellers, seats and unidentified users depend on seller commitments written into the deal terms.

How Do AI Agents Manage Frequency Capping Across Several Negotiated Deals at Once?