Ads Research Lab
A Project Moosed research lab for ad design and the future of ads and marketing.

The Model

One advertiser. One site. One budget. Two lanes.

Living page, version 1.0, dated 2026-11-16

This page is versioned and never silently edited. When an experiment changes the model, the change lands as a new version with a dated note saying what changed and why.

Defined terms

Human Ads
Units whose payload is human perception: image, slogan, video, brand creative. Served in a human-visible page to a human browser, and bought and sold on impressions, viewability, and clicks.
Agent Ads
Units built specifically for an agent, whose payload is machine-readable structured value: a verifiable price, offer, availability, or term. No creative as payload. Delivered to the agent, evaluated by the agent, and summarised by the agent to its person in the agent's own voice, with disclosure.
The Take
The logged event in which an agent accepts an Agent Ad's value into its task context: it pulls the feed item, calls the offer API, or holds the bound token. The Take replaces the impression as the auditable unit in the agent lane.
Agent Channel / Direct Line Feed
The agent layer of a site or network (manifest, feeds, connector responses, and the structured data in the page the agent actually parses) through which Agent Ads are served in the agent's browsing path, the way the agent browses the site. Hidden from the human eye, because the person hears only the agent's summary. Never hidden from the system: delivered, taken, presented, and redeemed events are logged at both ends, so the channel is auditable and revocable.
Agent-as-Shopper
The condition in which the agent holds the person's cart, wishlist, price ceiling, preferences, and task context, and therefore, not the advertiser and not a tracking cookie, makes the match between an offer and a held intent.

The two lanes, side by side

ElementHuman laneAgent lane
VisitorA human browser arrivesA verified agent arrives, on its person's task
ServingHuman creative served in the visible pageAgent layer served: manifest, feed, connector response, structured data
PayloadVisual, brand, emotionStructured, specific, takeable value (offer, verified price, real inventory, terms)
DecisionThe person sees, and may clickThe agent evaluates the offer against held context (price ceiling, cart, preferences, merchant trust, market price check), then presents it to its person in its own voice, with disclosure
Key eventImpression, then clickThe Take (logged, identity-bound), then presentation, approval, and redemption
MeasurementValidated impressions and clicksTakes, presentations, approvals, and redemptions
Billing basisValidated display or clickValidated take or redemption, from the signed log

Simple in sense: give the agent something valuable to take; the agent is the shopper; the take closes the line from ad to sale. Complex in fact: identity, binding, verification, disclosure, anti-gaming, and logging must all hold at once. Flawless execution is the moat, because trust is the product.

Propositions

Stated in Paper #1 as numbered, testable claims. Evidence status below is the lab's own status: the lab has not run experiments yet, so lab status is Untested unless noted. External evidence, where it exists, is reviewed and graded inside Paper #1.

IDPropositionEvidence status
P1Creative throughput collapse. Interruptive creative with no machine-usable value yields near-zero agent throughput: agents block, strip, summarise past, or never render it.Untested by the lab. Directional external evidence, graded in Paper #1.
P2Takeable value is necessary. An Agent Ad's throughput is a function of verifiable, task-relevant value, not of creative quality.Untested
P3Agent-as-shopper matching beats advertiser targeting. Agent-side matching of a pool-scoped offer, using context the advertiser should not have, produces higher redemption per take than advertiser-side personalised targeting of the same pool.Untested
P4Instrument attribution. A single-use, identity-bound and product-bound offer token turns attribution into a logged chain (issued, taken, presented, approved, redeemed) instead of a modelled probability. The remaining question, whether the sale would have happened anyway, is testable with holdouts.Untested
P5Trust is single-failure fragile. One wrong-person presentation, one leaked or double-redeemed token, or one price that fails verification measurably reduces future acceptance, because the presenter was the trusted agent itself.Untested
P6Gaming equilibrium. If holding off reliably extracts offers, agents optimising for their people will learn strategic hold-off. Sustainable systems must price this in: late, capped, randomised, or threshold-bound offers.Untested
P7Lane divergence over time. As the agent-mediated share of a category's transactions rises, transactional spend in the agent lane rises with it, while human-lane spend concentrates in brand and entertainment. This is a prediction to be tested, not a result.Untested by the lab. Directional external evidence, graded in Paper #1.

Boundary conditions

The model applies where a purchase or booking decision is delegable and comparable on price and specification, where the agent holds real context with the person's permission, and where redemption is digitally verifiable. It applies weakly to pure brand, status, taste, and impulse categories, to informal commerce, and where regulation requires human-visible advertising.

Version history

VersionDateChange
1.02026-11-16First published version, drawn from Paper #1 (in review). All propositions marked Untested by the lab.