The Architectural Differences Between AI-Native Exchanges and Legacy SSPs Wrapped for Chat

AI app publishers need monetization infrastructure that can use conversational intent and generated responses as native auction inputs without turning them into display impressions. The architectural difference between AI-native exchanges and legacy SSPs wrapped for chat is the unit of trade. An LLM-native exchange must allocate influence over possible model outputs, with delivery and publisher controls designed for the conversation.

Why the impression is the wrong unit of trade for LLM advertising

An impression auction assumes that what is being sold exists apart from the content around it. Legacy real-time bidding infrastructure, built around OpenRTB, auctions individual impressions in milliseconds against creatives prepared before bidding begins. The placement is discrete, and the advertisement is already made. These assumptions define the transaction and its format. A buyer competes for an opportunity to display a known creative, rather than for influence over content that has yet to be generated.

In an LLM environment, the sponsored response does not exist before generation. The commercial opportunity lies among the possible outputs the model could produce, including different recommendations or references. LLM-Auction, identified by arXiv 2512.10551, describes this as a shift from discrete ad slots to distributions over LLM outputs. The advertiser is bidding to influence that distribution. That changes what allocation means, because the auction must account for how an advertiser’s inclusion affects the response the user receives.

The distinction also sets the limits of retrofitting. A chat interface can expose inventory to a legacy SSP without changing what that SSP understands as a tradeable asset. It still allocates an impression against a prebuilt creative. This can accommodate a conventional advertisement around a conversation, but it does not supply the abstraction required to auction influence within generated content.

What a legacy SSP actually does when it wraps a chat surface

A legacy SSP wrapper translates a conversational interaction into the nearest impression object the underlying system understands. It compresses context into bid request fields, fits the advertisement into a prebuilt format, and presents the resulting opportunity to campaigns organized around page-level signals. The translation preserves the old auction contract. Although the surface is conversational, the buyer still receives a representation designed for choosing a creative to occupy a placement.

That representation usually has a page URL, device type, geography, and perhaps a cookie ID. Those fields describe access conditions or proxies for interest. They do not, by themselves, represent the user’s stated constraints or what the assistant is about to say. Adding a context field leaves this underlying limitation intact. The field may enrich targeting, but the system still resolves the transaction as a bid for an impression.

The compression becomes especially restrictive when a surface offers an experience that does not fit the accepted creative formats. Those experiences must be sold as generic impressions or stay outside that auction ecosystem. This is a format consequence of the data model. A wrapper can translate a surface into the existing contract, but translation is not replacement. The alternative therefore requires a different auction object, one that represents how advertising can shape a response.

How an LLM-native exchange redefines the auction object

An LLM-native exchange allocates influence over a probability distribution of possible responses. Its central question is which advertiser’s value, weighed against user experience, should shift the output toward a particular recommendation or reference. Instead of choosing a finished advertisement for a fixed location, the mechanism must evaluate the relationship between advertiser value and the content the model may generate.

LLM-Auction proposes a learning-based generative auction mechanism that integrates auction and generation. It formulates allocation as preference alignment between LLM outputs and a mechanism objective that balances advertiser value with user experience. This makes the objective part of generation itself. The approach optimizes the model to account for allocation externalities without extra inference cost, rather than treating those effects as a separate problem after an advertiser has been selected.

Externalities are central because an advertisement does not affect only its own exposure. Its placement can change how other advertising appears and how the user experiences the whole response. Existing pre- and post-generation bidding approaches face a trade-off. They either ignore these interactions or incur prohibitively high inference costs. Folding allocation into generation is the paper’s proposed answer to that problem, allowing the mechanism to consider the response as a whole.

The paper also identifies allocation monotonicity and continuity as properties of the optimal rule, and proves favorable incentive properties for a first-price payment rule under that allocation. The significance is architectural rather than merely mathematical. It establishes a basis for defining auction behavior at the distribution level, instead of assuming that an auction requires a discrete slot. It does not, however, settle how that mechanism can connect to every application architecture, particularly where the publisher cannot modify the model producing the response.

The decoupling problem — why ad insertion architecture matters as much as auction design

An auction that understands generated outputs can still depend on an insertion method that does not fit the application. Most existing LLM-native advertising approaches change how generation works through fine-tuning or internal model mechanisms. That couples advertising to the model producing the answer. The coupling becomes a deployment constraint when the application relies on a closed-source service or a workflow whose behavior cannot be reduced to one model.

PILA, identified by arXiv 2607.25590, names three incompatibilities in this approach. Frontier LLMs are closed-source and API-only, leaving their parameters unavailable. Agentic applications increasingly combine heterogeneous components, so advertising must operate across the pipeline. There is also an objective conflict. Asking the same model to optimize response quality and advertising may interfere with its primary task of serving the user.

PILA reformulates insertion as conditional response rewriting in a lightweight sidecar module. Drawing on the sidecar pattern in cloud-native architectures, it attaches after generation while leaving the upstream model or workflow frozen. The delivery component is therefore model-agnostic. Its ad-intensity factor ρ controls the trade-off between user-side naturalness and ad-side exposure and gives downstream pricing and deployment an interface to use. Naturalness becomes an explicit control rather than an incidental result of winning an auction.

A legacy impression auction has no corresponding insertion layer that reasons about this trade-off. Its allocation ends with a winning bid. The two papers describe different needs in the architecture but stop short of showing a single system that brings them together. Allocation must understand distributions, while delivery must fit model-agnostic applications without entangling their primary response generation. Those requirements also change the information the exchange needs before it can form a useful auction.

What the bid request must carry when context is a conversation, not a page

A conversational bid request must describe the task being done as well as the surface where it takes place. Its relevant inputs include the user’s intent, expressed constraints such as price range, location, timing, and preferences, and the assistant’s next action. Recommending, comparing, booking, and summarizing describe different response contexts. Those distinctions help define what advertiser participation would mean within the answer, rather than leaving the buyer to infer relevance from a browsing category.

The contrast with legacy targeting is a difference in signal quality. A page URL or cookie ID may correlate with a need, but the conversation can express that need directly. The request can describe what is wanted now. This is why adding conversational traffic to an unchanged request format loses more than descriptive detail. It removes the task and constraints that give distribution-level influence its value, leaving the buyer with proxies for the very intent the application already understands.

ChatGPT’s ad system illustrates the resulting change in targeting vocabulary. It uses a short context hint, a freeform natural-language description of the conversations, topics, or keywords where the advertiser’s product is relevant. Relevance is not an exact-string contest.

A well-constructed AI-native bid request can summarize intent without transferring the underlying text, reducing dependence on personal identifiers. Privacy belongs in the request design itself. For publishers, this provides a way to make commercial intent legible while limiting what leaves the application. The same conversational understanding must also identify eligible advertising opportunities in the first place.

How do AI app publishers identify and represent inventory for monetization?

A legacy SSP manages a catalogue of stable placements, including banner slots, video pre-rolls, and in-feed units. Their dimensions and cadence are known before an individual interaction occurs. Conversational inventory has no equivalent fixed catalogue. An LLM SSP must dynamically classify monetizable moments as they emerge from exchanges that did not exist before the user began typing.

The inputs to that classification come from the interaction itself. What the user asked, the topic under discussion, and the content the publisher supplied define the potential opportunity. The sell-side data model must therefore represent an event whose suitability depends on the conversation. This complements the bid request’s task representation by establishing what supply is actually eligible for sale.

Publisher controls operate at that same granularity. Because sponsored content appears in a dialogue the user trusts, blocking rules and suitability filters must take the conversation into account as well as the surrounding page. A URL match is not enough here. A financial news publisher that blocks gambling ads needs that rule to cover the full conversational context. The control belongs to the opportunity being classified, so that eligibility reflects both the publisher’s rules and the discussion in which the advertisement would appear.

This model also identifies supply that conventional display infrastructure does not monetize. AI interactions that generated zero ad income because bots do not view display ads or follow referral links can become monetizable events. The revenue source is the interaction itself. In the model described here, that adds revenue while keeping the existing display or programmatic yield. Its representation must remain conversational, however, or the exchange returns to selling the generic impression that failed to describe the opportunity.

Why a platform needs both exchange-level context and cross-surface reach to serve either side well

Generalist buying platforms and single-surface AI ad networks solve different parts of this problem. Generalist platforms have cross-surface reach but get the flattened bid requests that legacy infrastructure produces. Reach does not restore the missing context. Without a representation of prompt-level intent and distribution-level influence, a buyer cannot bid meaningfully on those properties, regardless of how many surfaces the buying platform can access.

Single-surface AI ad networks have the reverse limitation. They can read conversational context within their own environment, but they cannot provide access beyond it. Their strength is bounded by the surface. An advertiser that wants relevant conversations across AI assistants, copilots, and embedded chat widgets needs more coverage than a single-surface integration can offer. Contextual depth and breadth of supply are separate capabilities, and neither substitutes for the other.

A DSP operating across AI surfaces on top of an exchange with direct publisher supply combines contextual depth with breadth of supply. In that setup, the exchange builds the conversational bid request where it has the context, while direct ties to publishers give it access across surfaces. Both layers have a defined role. The exchange makes the opportunity legible, and the buying platform makes it accessible across the ecosystem without reducing it to pass-through legacy inventory.

For AI app publishers evaluating monetization infrastructure, these distinctions establish concrete architectural criteria. The auction object must describe generated outputs, insertion must work with the application’s model or workflow, and inventory controls must apply to the conversation. Cross-surface demand then gives that supply broader access to advertisers. The boundary between an AI-native exchange and a wrapped SSP lies in what the system can trade.