Volume Commitment Thresholds in Conversational AI Ad Contracts

Volume commitment thresholds in conversational AI ad contracts cannot be set using the methods that work in search or display, because the inventory, the targeting unit, and the billing model all behave differently, and that mismatch has already produced measurable failures in early pilots. Advertisers who signed CPM-based commitments could not spend down the budgets they had agreed to, because the impressions needed to exhaust that spend simply weren't available at the volume the contract assumed, even though demand was strong. The basic math buyers use to size a commitment in search or display does not transfer to a conversational interface, and any team that signs a volume deal in this channel without understanding why is accepting a risk that looks familiar but is not.

Volume commitments in conversational AI ad contracts as a different problem than in search or display

Search advertising rests on a foundation buyers have had two decades to study. A keyword carries a known query volume, a known depth of competing bidders, and a known CPM or CPC floor, which lets a buyer model, before signing anything, roughly how fast a given budget will get consumed. Conversational AI advertising offers no equivalent foundation yet. The unit being targeted is not a keyword but a conversational context, what OpenAI's early ad product describes as context hints rather than individual keyword queries, and that shift means the depth of inventory behind any given topic or category is not stable and cannot be looked up the way search volume can. Layered on top of that is a billing model that started out CPM-based and disconnected from performance, which made it extraordinarily hard for early advertisers to burn through a committed budget when the impressions themselves were scarce and went largely unmeasured. A third complication has no parallel in search at all: OpenAI's ad rollout confines ads to the Free and Go tiers, leaving Plus, Pro, Business, Enterprise, and Education subscribers entirely ad-free, so no commitment, however large, can ever buy access to the full user base. Each of these three factors, the targeting unit, the billing model, and the tier structure, would be enough on its own to break a commitment formula imported wholesale from search. Together, they explain why early pilot advertisers hit a wall that search buyers never encounter: money they were contractually obligated to spend, with no inventory path to spend it on.

Targeting inside an LLM conversation and inventory depth

Placement inside a chatbot conversation gets decided through relevance-weighted contextual matching, not a keyword lookup, so neither the buyer nor the platform can say, with any confidence, how many eligible impressions a campaign will generate over a given stretch of time. Visiblie's 2026 guide to ChatGPT ads makes the point: a keyword layer is absent. Advertisers bid for placement against topic categories and conversational context, not against the kind of discrete, countable query a search marketer has used for sizing campaigns since the earliest days of pay-per-click. Research on LLM advertising systems lays out two competing architectures for handling this. One is dynamic bidding, where the advertiser receives the query and its surrounding context in real time and returns a bid on the spot, which is about as precise as targeting gets but demands heavy computation and raises obvious privacy concerns, since it requires passing live conversational content to outside bidders. The other is static bidding, where the advertiser commits a bid against a category in advance, without any per-query exchange of information. Static bidding is coarser, but it is operationally realistic, and it maps closely to what platforms have actually built so far. A middle path in the research is a "genre" model, built around stable semantic clusters such as hotels or airlines, which cuts down on both the computational load and the privacy exposure of dynamic bidding while giving buyers something that feels closer to the keyword targeting they already know. Genre-level targeting is still a coarser signal than a keyword, and nobody has yet established how deep the inventory runs within any given genre. You can see this directly in OpenAI's own decision to expand ad exposure to logged-out users. That expansion was an operational response to an inventory shortage pilot advertisers had already run into, and it confirms that the addressable inventory available at launch was narrower than the volume commitments written against it assumed.

CPM-to-CPC billing shift and the meaning of "commitment

If you make a CPM-based commitment, you take on all of the inventory risk, because spend gets consumed by impressions delivered, whether or not those impressions ever convert into anything. That arrangement explains why pilot advertisers struggled to exhaust their budgets: they were paying for impressions too infrequent to reach the volume floor they had committed to, with nothing measurable coming back in return. Artefact's 2026 state-of-play analysis documents a shift away from that model. OpenAI's billing is no longer expected to run on CPM alone: CPC bidding is becoming available too, and a newly announced tracking pixel now lets buyers track conversions across events such as order creation, leads, and subscriptions. In the span of a short window, the platform moved from billing on impressions to billing on performance, and that move changes what a volume commitment actually represents. Under CPM, the buyer bears all of the performance risk, because spend drains regardless of outcome. Under CPC, spend only drains when a user clicks, so thin inventory or poorly matched creative can leave a commitment just as hard to exhaust, only through a different mechanism. A contract signed under CPM assumptions, in other words, can end up badly mispriced if the platform shifts to CPC partway through the term, because the effective reach a fixed budget buys changes with the billing model itself. The tracking pixel and the move toward CPC are real improvements for advertisers trying to measure what they are buying, but both arrived after the earliest commitments were already signed. Buyers locked in spend before the measurement infrastructure existed to justify the number they agreed to.

The platform tier split as the most underappreciated clause in any ChatGPT ad commitment

Ads are permanently excluded from every paid subscription tier. Any volume commitment on ChatGPT is, by construction, a commitment against a slice of the user base rather than the whole of it, and that slice skews toward users who engage less deeply than the subscribers sitting behind the paywall. This is not a flaw in OpenAI's design. Keeping paid tiers ad-free is a legitimate and defensible product choice, one that plenty of subscribers likely pay for directly. But it creates a sizing problem for anyone setting a commitment threshold, because the users excluded from ad exposure, the Plus, Pro, Business, Enterprise, and Education subscribers, skew toward exactly the high-intent, research-ready buyers that performance advertisers most want in front of their ads. Free-tier users tend to sit earlier in the research funnel: they generate more impressions in raw terms but convert at lower rates. Sizing a commitment against total user volume rather than the volume actually eligible for ad exposure overshoots what the campaign can realistically deliver, and sizing it against conversion outcomes instead proves hard to hit, because a paid subscription has removed the users most likely to convert from the pool. The contrast with Anthropic makes the stakes concrete. Artefact's analysis confirms that Anthropic has committed to keeping Claude entirely ad-free, and positions it explicitly as a protected space for productive work. Any advertiser building a commitment strategy across multiple AI platforms has to treat that as a fixed constraint: one of the major conversational surfaces is permanently closed to paid placement, not temporarily underdeveloped.

The platform's unilateral right to restrict placements and the unquantified delivery risk inside a committed contract

A volume commitment assumes the buyer controls the one variable that matters most, spend, but conversational AI ad contracts hand the platform a broad discretionary right to decline, restrict, or remove advertisers and ad content for reasons that cannot be fully listed in advance. OpenAI's Ad Policies reserve exactly that right, stating that the company can decline, restrict, or remove advertisers and refuse to display ad content and website links for any reason, including where they conflict with its advertising principles, business interests, or competitive position. Any buyer reviewing contract language needs to look closely at that last phrase, competitive position. It means that if an advertiser's product competes with something OpenAI builds or partners on, the platform can withhold placements without that withholding counting as a breach of its obligations under the contract. Brand safety adds another layer of uncertainty on top of this. In standard display advertising, the content sitting around an ad can be classified and screened before the impression ever serves. In a generative interface, the surrounding content gets created at the moment of the response. It cannot be pre-screened the way a static web page can. Advertisers in regulated categories such as finance and health may choose, on their own, to suppress placements whenever a model's response carries low confidence or touches sensitive territory, and that shrinks the inventory actually deliverable against a committed floor even further. Buyers negotiating these contracts should be asking for specific protections: delivery guarantees that trigger if the platform's own restriction rights suppress a material share of committed volume, contract language treating platform-side refusals the way force majeure clauses treat other uncontrollable events, category exclusion rights that let an advertiser opt out of adjacency risk without being penalized, and audit provisions that let a buyer verify why impressions were or weren't delivered against the commitment.

The publisher revenue-share structure and the inventory a buyer is committing to

If you set a volume commitment on a conversational AI platform, you commit against an inventory pool whose editorial character and brand suitability are shaped by publisher revenue-share agreements that are, in nearly every case, undisclosed to you. OpenAI has struck content deals with News Corp, Vox Media, the Financial Times, and the Guardian Media Group, and those arrangements bring trusted, brand-safe editorial material into the ChatGPT surface, creating ad-eligible inventory that sits next to credible journalism. The terms behind those arrangements, the specific revenue share, the exclusions, the renewal conditions, are not public. That opacity matters directly for anyone sizing a commitment. The buyer has no way to know which publisher surfaces, which content categories, or which response contexts actually make up the eligible inventory they are bidding into. If a publisher partner later exits an arrangement, adds new content exclusions, or renegotiates its terms, the inventory pool a commitment was priced against can shrink without any visibility into why, and a commitment floor set against an assumed level of inventory quality has no way to enforce itself if the underlying supply mix has already changed. There is a supply-side pressure running in the other direction too. If conversational interfaces increasingly synthesize answers instead of sending users out to click through to original sources, publishers lose referral traffic, so some may scale back their participation in ad-eligible supply deals, and that compresses the available inventory further over the life of a multi-year commitment. Buyers negotiating these contracts should press for transparency into supply composition and for substitution protections that spell out what happens to the commitment if a named publisher's inventory disappears mid-term.

Vertical context and whether a given commitment threshold is realistic

The right commitment threshold has to be set vertical by vertical, because the rate at which conversational context actually matches ad eligibility, the regulatory limits on where ads can run, and the stage of the funnel at which conversational intent turns into a purchase all vary sharply by industry. E-commerce is the most mature vertical for conversational AI advertising right now. OpenAI built shoppable integrations and launched built-in checkout in September 2025 that let users buy products directly inside ChatGPT, with the platform taking a commission on the transaction, before pulling that feature back in March 2026. Commitments in this vertical can reasonably be set higher, because the path from conversation to purchase is short and the inventory eligible to carry ads is comparatively broad. Finance carries the opposite profile. Regulatory exposure is highest here, and conversations that touch financial advice generate real reputational risk, often triggering strict suitability exclusions that shrink the inventory actually deliverable against a commitment floor, so buyers in this category should build exclusion rights directly into the contract language rather than treat them as an afterthought. Health carries its own distinct risk, tied to model confidence rather than regulation alone: when a response gets generated with lower certainty, health advertisers may reasonably choose to suppress their own placements rather than risk sitting next to an inaccurate answer. Buyers in this category should negotiate for category-level exclusion rights, so suppressed impressions don't count against the commitment threshold. Across every vertical, the same underlying distortion repeats. Research-stage conversations generate more eligible impressions but convert at a lower rate, so a commitment sized purely on impression volume can look healthy on paper even as it produces thin business results. Purchase-ready conversations run the opposite way: fewer eligible impressions, but conversion rates high enough that a commitment built around business outcomes needs a meaningfully lower impression floor than a research-stage campaign of equivalent value would require. A threshold measured only in impressions or clicks, regardless of which vertical it sits in, tells a buyer very little about what the campaign is actually worth.