Every ad platform I have ever spent money on runs on the same primitive: an auction that clears when a human is about to look at something. Search auctions clear on a query. Social auctions clear on a scroll. Video auctions clear on a play. The entire trillion dollar machine assumes that the scarce resource is a pair of eyes, and that the moment those eyes land on a surface, someone should pay for the privilege of being there.
That assumption is quietly breaking. When an AI agent researches a purchase on a customer's behalf, there is no scroll, no play, no glance. There is a retrieval pass, a comparison pass, and a decision. The scarce resource is no longer attention. It is consideration: the finite set of candidates the agent actually evaluates before it acts. And where there is scarcity plus commercial intent, an auction always forms. The only questions are who runs it, what the bid unit is, and how ugly the first version gets.
My position is simple and falsifiable. By the end of 2028, at least one major AI platform will operate a paid consideration mechanism inside its agent shopping flows, priced per evaluated candidate rather than per impression. By 2030, consideration auctions will be a standard line on media plans for ecommerce brands. If neither happens, this paper was wrong. I am comfortable with that bet because I have watched this exact movie before, in search, and the economics rhyme too hard to ignore.
Key Findings
- The auction is not dying, it is migrating. It moves from the moment of human attention to the moment of machine shortlisting, and the clearing logic changes with it.
- The bid unit shifts from the impression to the Consideration Slot: a paid guarantee that your offer is retrieved and evaluated, not that it wins.
- Quality score comes back with teeth. Agents verify claims at decision time, so a paid slot with weak substance converts worse than no slot at all.
- Second price logic breaks when the buyer is software, because agents can be interrogated about why they shortlisted something, and platforms will need auditable answers.
- Early consideration auctions will be private, negotiated, and invisible to dashboards. Self serve arrives two to three years after the pilots.
- Brands that treat this as another retargeting checkbox will overpay. The leverage is in the claim layer underneath the bid, not the bid itself.
Why an Auction Forms at All
Some people in my industry argue that agents make advertising obsolete: the agent just picks the objectively best product, so why would anyone pay? That argument misunderstands what agents actually do. An agent researching a category cannot evaluate every option on earth. It samples. It retrieves a candidate set, maybe eight options, maybe twenty, and reasons over that set. Everything outside the set loses with certainty.
Whenever a gatekeeping step compresses a large market into a small candidate list, the positions on that list acquire economic value. That is not cynicism, it is arithmetic. Organic retrieval will fill most of the list most of the time, exactly the way organic results fill most of a search page. But the marginal slot, the one that gets a challenger brand into a comparison it would otherwise miss, is worth real money to someone. Platforms know this. They monetized the search results page the same way, and I laid out the mechanics of that history in programmatic advertising explained. The consideration list is the new results page, minus the pixels.
The honest version of this product is not "pay to win the recommendation." No platform can sell the recommendation itself without destroying the trust that makes agents useful. The sellable unit is entry into evaluation. You pay to be examined, and then your product data, pricing, and reputation have to survive the examination.
The Consideration Slot
Let me define the unit precisely, because sloppy definitions are how buyers get robbed. A Consideration Slot is a paid guarantee that, for a defined class of agent tasks, your offer is included in the candidate set and evaluated against the user's constraints, with the evaluation logged. Not an impression. Not a click. Not a conversion. An evaluation.
Three properties follow from that definition. First, a slot has no value if your structured data cannot answer the agent's questions, which means the prerequisite spend is data infrastructure, not media. I made this argument from the demand side in marketing to AI agents, and the auction version is stronger: paying to be evaluated with weak claims is paying to lose in front of a judge. Second, slot pricing should vary with task specificity. Being evaluated for "running shoes under 140 dollars, delivered Friday" is worth far more than being evaluated for "shoes," and platforms will tier accordingly. Third, the natural billing event is the evaluation log, which finally gives advertisers a receipt that is checkable rather than claimed.
Back of the napkin, imagine a category where an agent shortlists ten candidates and organic retrieval fills seven. Three paid slots, thousands of tasks per day, and a conversion rate that depends entirely on offer strength. That is a small, brutal, high signal marketplace. It looks nothing like display and everything like a courtroom with an entry fee.
Quality Score Gets Fangs
Search taught platforms that pure highest bidder auctions poison the well, so they invented quality score to blend relevance into ranking. In consideration auctions, quality score stops being a discount mechanism and becomes a survival mechanism, because the evaluator actually checks.
A human clicks a mediocre ad and the platform still gets paid. An agent evaluates a mediocre offer, finds the shipping claim contradicted by the checkout page, and rejects it, and if the platform keeps selling slots to that advertiser, the agent's outcomes degrade and the user notices. Platforms therefore have an existential incentive to gate consideration slots on claim integrity: consistency between feed data and live site, honored prices, return policies that match reality. I call the resulting metric Evaluation Yield: the share of paid evaluations that survive constraint checking without a disqualifying discrepancy. I expect Evaluation Yield, under whatever name each platform invents, to function the way quality score did in 2005, quietly deciding who can afford to participate at all.
The strategic consequence: the cheapest way to lower your effective cost per consideration is not bidding tactics. It is fixing every place where your claims disagree with each other.
Who Runs the Auction
Three candidate operators, in descending order of likelihood. The AI platforms themselves, who own the agent, the logs, and the user relationship, and who need revenue lines that do not depend on human eyeballs. The retail and commerce platforms, who already run retail media auctions and can extend them to agent traffic hitting their catalogs. And a possible neutral layer, an exchange for consideration slots across platforms, which I would love to see and mostly do not expect before 2030, because the platforms have no incentive to interoperate while they are land grabbing.
What I do not expect is the open RTB model replicating cleanly. Real time bidding worked because impressions were a commodity and nobody audited much of anything. Evaluations are not commodities, they are contextual, logged, and interrogable. The auction after attention will look less like an exchange and more like a certification regime with prices attached. I cover the fraud pressure on these logs separately in Paper No.55, because fake evaluations are coming the moment real ones are billable.
The Uncomfortable Transition Years
Between now and self serve maturity sits an awkward period that will punish the unprepared. Pilots will be private. Pricing will be opaque and negotiated. Large brands will get access first, and the platforms will disclose as little as legally possible about how paid and organic consideration mix. If that sounds familiar, it should; early search monetization went through the same murky phase before disclosure norms hardened.
During this window, two failure modes dominate. Brands with strong data foundations but no platform relationships will be invisible to the pilots. And brands with relationships but weak data will burn budget on evaluations they cannot survive. The rare position, strong claims plus early access, will compound advantages for years, because agent memories and platform priors form early and update slowly.
What I Would Do About It
First, build the receipt habit now. Demand evaluation level logging from any partner selling you agent adjacent placement, and refuse to buy anything billed on proxies you cannot audit. The consideration auction only works for buyers if the billing event is verifiable, and buyers set that norm in the first two years or never.
Second, spend the next twelve months on Evaluation Yield before you spend a dollar on slots. Reconcile your product feed, your site, your policies, and your third party listings until an adversarial reader finds no contradictions. This is unglamorous and it is the whole game.
Third, assign clear ownership. Consideration bidding is not a search team job or a programmatic team job; it sits between data engineering and media, and orphaned functions get eaten by agencies and platforms. One named owner, one budget line, starting this fiscal year.
Fourth, negotiate your way into pilots even at unattractive prices. The learning value of seeing real evaluation logs before your competitors exceeds the media value by an order of magnitude. Pay tuition early, while tuition is cheap and the teachers still return your calls.