Search-era advertising priced keywords because the keyword was the most legible signal of intent the system could see. Agent-era surfaces produce a different signal: the conversational moment itself, with all the texture that the keyword stripped away. The unit of value is moving from the term to the context — and that move is what the intention economy names.
For two decades, the dominant ad signal was a keyword someone typed into a search box. The keyword was compressed, deliberate, and easy to auction. It worked because users had to translate their actual question into search-query language before the system could see it.
On agent surfaces, the translation step disappears. A user describing what they want to a model — in full sentences, across multiple turns, with context the model has accumulated — produces a signal that is no longer a keyword. It is the decision-forming context itself: what the user is trying to do, what constraints they have named, what they have already ruled out.
The economic question is not whether this signal is richer (it is) but what it is worth, and how it can be priced.
Keywords were never the thing of value. They were a compression of the thing of value into the form the search-era infrastructure could read. A user typing "running shoes for marathon training" was not really searching for those four words. They were in a specific decision-forming context — distance, surface, prior shoes, budget — that the keyword compressed because the system could not see the rest.
Behavioral ad systems tried to recover the compressed-out context by building profiles: who is this person, what have they bought, what cohort do they belong to. The profile was a workaround for the keyword's narrowness.
Agent surfaces remove the need for the workaround. The decision-forming context arrives directly. The compression artifact is gone.
The intention economy is the claim that economic value on agent surfaces accrues to moments of decision-forming context, not to the keywords or behavioral profiles that previous eras used to approximate them. The valuable unit is the conversational moment where a user has named what they want with enough texture that a sponsor's offer can match it precisely — without the system needing to know who the user is.
Two things change as a consequence:
Neither change is a moral claim. Both are mechanical consequences of the signal shape changing.
A sponsor in the intention economy buys access to decision-forming context — specifically, the right to be considered when a user's stated intent matches the sponsor's declared offering. They do not buy:
What they do buy is the alignment between their published offer (the covenant manifest, see /protocol-spec and agent-ads.org/abf) and a specific class of conversational moments. The unit is the moment, not the audience.
This is a different pricing object than a CPM or a CPC against a user profile. The accountability question — how a sponsor knows the moment actually arose and the placement actually happened — is the sister page's subject.
OpenAI Ads is the category signal that this thesis is operational, not theoretical. A major agent-surface platform shipping an ad system that selects against conversation, context, and ad details — rather than against user-behavioral history — is a visible live datapoint for the intention-economy claim.
The protocol described on contextual-ads.ai generalizes the same shift: it specifies how decision-forming context can be priced and matched across any agent runtime, not just one platform's. OpenAI Ads is a category leader making the category exist; the protocol makes the category portable.
The intention-economy thesis only works if the moment can be priced, which requires that it can be measured. Without measurement, "decision-forming context" becomes another marketing word that any system can rebrand around.
Measurement in this protocol cannot collapse back into behavioral tracking — if the only way to verify that a sponsor's bid was honored is to follow the user across sites, the thesis fails. The accountability surface has to be the protocol's own artifacts: manifest fetches, feed eligibility, pre-render verification outcomes, rendered-label proofs, non-user-identifying outcome windows.
That accountability surface is described on the sister page.