Platforms That Sell Native Ad Placements Inside Live AI Conversations Ranked by Inventory Type

Conversational AI advertising platforms fall into three distinct inventory categories: walled-garden single-surface platforms like OpenAI's ChatGPT and Google's AI Overviews, open-exchange cross-surface DSPs that aggregate buying across multiple AI assistants, and hybrid platforms that operate both a DSP and their own SSP with direct publisher supply. The platform a buyer chooses determines reach, how precisely ads can be matched to live conversation, and how much control the buyer retains over placement terms. This piece ranks and explains each category as the market stands today, so you can weigh inventory before you allocate budget, not after.

The inventory structure of an AI ad platform

Conversational AI advertising is a genuinely new channel, distinct from search or social. The surface is a live conversation. There is no results page, no feed slot, no keyword auction, and no persistent cookie, and that structural difference means the inventory a platform controls determines almost everything else about what a buyer can do.

Search advertising was built around a stable surface, the SERP, and it had a known targeting primitive, the keyword. Social advertising was built around a feed and an interest graph. Both models assume a fixed slot, a reusable creative, and a trackable click. Conversational AI removes every one of those assumptions at once: there is no page to rank on, no feed position to bid for, no keyword string to match against, and often no clean click to measure, because the assistant can answer a question, cite a source, and shape a user's decision without that user ever leaving the chat window.

Because the surface is conversational rather than a fixed slot, the inventory a platform controls, meaning which AI surfaces it can place into and on what contractual terms, sets the outer boundary of what any campaign can reach and how finely it can be targeted. If you evaluate these platforms on bidding features, creative formats, or reporting dashboards before you understand what inventory sits underneath them, you will find the real constraint only after you have already committed budget. Inventory structure is the first question to settle.

The three inventory structures that exist in the market today

Three structurally distinct inventory types operate in the conversational AI ad market today, and the differences among them are not cosmetic. They determine how far a campaign can reach, how precisely it can be matched to context, and how much control a buyer keeps over the terms of that match.

The first is walled-garden, single-surface inventory, where the platform owns both the AI assistant and the ad inventory running inside it. Buying is direct, contextual matching runs deep because the platform has full access to its own model, but reach is bounded by the audience of that one assistant. OpenAI is the defining example: ads began appearing inside ChatGPT on February 9, 2026, with self-serve buying opening at ads.openai.com on May 5, 2026, Criteo named as the first technology partner in March 2026, and StackAdapt adding programmatic access in May 2026. Google AI Overviews is a second walled-garden case, distinct in mechanism: ads inside AI-generated answers have been live since 2024 and run through existing Google Ads infrastructure, meaning Search, Performance Max, and Shopping campaigns qualify automatically, with no separate opt-in or opt-out.

The second category covers the surfaces where no inventory exists at all, and knowing which ones matters as much as knowing which ones are open, so budget isn't allocated against supply that doesn't exist. Anthropic has said publicly that Claude will stay ad-free by policy. Microsoft Copilot and Meta AI have not opened advertising and have announced no timeline for doing so. Google's Gemini chatbot, separate from the AI Overviews search surface, has no announced ad product either.

The third category is open-exchange, cross-surface inventory: a demand-side platform that aggregates buying across multiple AI surfaces, operating on top of an exchange and direct publisher supply relationships rather than owning any single assistant. Reach extends beyond one walled garden, and contextual matching works across surfaces rather than inside just one. Legacy media buying infrastructure, built for stable inventory and predictable formats, cannot access conversational AI surfaces at all, which is what makes a cross-surface DSP a structural necessity rather than a convenience. A fourth variant, hybrid, builds on this by operating both a DSP and its own SSP or exchange with direct publisher supply, allowing it to buy into walled gardens where they open programmatic access while also placing ads into independent AI publishers through its own exchange, giving it pricing control and supply guarantees a pure demand aggregator does not have.

How contextual matching works inside a live conversation, and why it is different from keyword or behavioral targeting

The targeting primitive inside conversational AI is the live meaning of the conversation itself, distinct from a keyword string or a behavioral profile assembled from months of past browsing. That distinction changes what counts as valuable inventory in this channel.

OpenAI's system for this is built around what it calls "context hints," natural-language descriptions that advertisers supply at the ad-group level to describe the conversations where a product genuinely belongs. The system uses these as semantic guidance, matching ads to the meaning and intent of a live exchange rather than to a fixed keyword. A user asking an assistant for travel recommendations right now is expressing purchase intent more precisely, and more immediately, than any keyword bid or demographic proxy ever could, because the signal comes directly from what the person is asking for at that moment.

That signal also disappears when the conversation ends. No durable user profile gets built or synced across sites the way a cookie-based system would do. So conversational targeting is privacy-first by architecture, not by policy choice, and that distinction matters if you are managing compliance in a post-cookie environment where persistent identifiers are no longer the default.

The practical consequence for evaluating inventory is this: platforms with deeper access to conversational context, typically walled gardens with full model access or hybrid DSPs with direct publisher integrations, can match ads to live intent with more precision than platforms that have simply attached a keyword layer to a chat interface. A keyword bolted onto a chat window still behaves like a keyword. It does not read intent the way a system built around the model's own semantic output can. The May 2026 addition of cost-per-click bidding alongside impression-based buying aligned advertiser spend more closely with user action, but the more consequential shift underway is conversation-depth bidding, which treats the engaged dialogue itself, rather than the single click, as the unit of value. The inventory type a platform controls is what determines how much of this conversational signal it can actually read and act on, which is the thread the rest of this evaluation follows.

Buying inside OpenAI's ChatGPT

ChatGPT is the reference implementation of walled-garden AI ad inventory. It offers the deepest contextual signal you can get inside any single assistant, but the same constraints that shape it define every single-surface model in this market.

Inventory opened to buyers on February 9, 2026, on ChatGPT's Free and Go tiers. Self-serve buying through ads.openai.com launched on May 5, 2026. Criteo was named OpenAI's first technology partner in March 2026, and StackAdapt added programmatic access in May 2026. You can enter either directly through ads.openai.com or programmatically through Criteo and StackAdapt, and if you want to fold ChatGPT into a broader cross-surface plan, that DSP layer is how you get in without negotiating a separate direct relationship.

Because OpenAI controls the model, the interface, the user relationship, and the ad policy all at once, it also sets every term on what can run and how. Adult content, dating, gambling, alcohol, tobacco, political ads, illegal drugs, cryptocurrency, credit repair and debt settlement, and weapons are excluded. Health and financial services face additional review, not a blanket ban, and OpenAI has widened eligibility for both through several 2026 policy updates. Legal services were added only as of August 31, 2026, and are now permitted in the US for advertisers licensed to practice law in the jurisdiction where the ad runs.

Brand safety inside this walled garden rests on machine-learning systems, including LLMs and classifiers, that review ads, landing pages, and advertiser signals before an ad becomes eligible to run, with some decisions escalated for human review. What those systems cannot do is pre-screen each individual placement context, because every conversation is generated dynamically and does not exist until the user creates it. That is the unresolved brand-safety problem sitting inside every walled-garden model that places ads into live, generative text.

The ceiling on this model is straightforward: reach is bounded by ChatGPT's own user base. If you want to reach users on other assistants, you have two options: open separate direct relationships with each surface individually, or route through a platform built to aggregate supply across surfaces. That second option is where the cross-surface and hybrid types covered later in this piece become relevant, as the structural answer to a ceiling that is true of every single-surface platform, OpenAI included.

Buying inside Google's AI inventory: a different walled-garden with different access rules

Google's AI ad inventory, spanning AI Overviews and AI Mode, runs through existing Google Ads infrastructure, and that means most buyers already running Search, Performance Max, or Shopping campaigns are advertising inside AI-generated answers whether they realize it or not.

AI Overviews has carried ads since 2024: the ads appear inside the AI-generated answer itself, and you buy them through the same Google Ads console you already use. Eligible campaign types include Search campaigns using broad match or AI Max, an optimization layer within existing Search campaigns rather than a separate campaign type, along with Performance Max and Shopping campaigns, and smart bidding is required across all of them.

No opt-in or opt-out control exists in this system. If a campaign qualifies under these criteria, it is already running inside AI answers, full stop on the mechanics: this stands in sharp contrast to every other platform covered here, where entering AI ad inventory is a deliberate, separate buying decision. Google's Gemini chatbot, worth distinguishing from AI Overviews, carries no announced ad product of its own; AI Overviews lives on the search surface, not inside the standalone chatbot.

The walled-garden model carries a risk that the Google case makes visible. When one company controls search, AI Overviews, and potentially future chatbot inventory all under the same roof, a buyer's leverage to negotiate placement terms, demand transparency into matching logic, or request granular data access is limited by the fact that there is only one counterparty controlling the entire stack.

The surfaces that remain closed

The surfaces buyers most want to reach are not uniformly open, and the distance between where AI attention currently sits and where buyers can actually place ads is the strongest argument for a platform built to aggregate supply across the AI publisher ecosystem rather than one that waits for a single walled garden to open its doors.

Microsoft Copilot remains closed to advertisers with no announced timeline, a meaningful gap given its penetration into enterprise software. Meta AI is closed as well, with no timeline announced, significant given the scale of Meta's consumer reach across its existing platforms. Anthropic has made it explicit public policy that Claude will stay ad-free, so no inventory exists there and none is expected.

The total addressable conversational AI audience is substantially larger than what any single platform, or even the two currently open walled gardens of ChatGPT and Google AI Overviews combined, can deliver on its own. AI-native advertising surfaces keep expanding past the handful of major consumer assistants into copilots, agentic experiences, and AI features embedded inside third-party apps, and much of that new supply sits with independent AI publishers, not with any walled garden.

The inventory a platform controls, meaning which AI surfaces it can place into and on what terms, sets the outer boundary of what any campaign can reach and how it can be targeted, a principle that separates genuine conversational AI advertising infrastructure from legacy ad tech retrofitted onto a chat interface. Platforms built as a DSP on top of direct publisher supply and their own exchange exist specifically to work against that boundary: they can read conversational context across several surfaces at once and offer reach beyond any single walled garden, closing a structural gap that generalist buying platforms were never built to close. A cross-surface DSP with direct publisher relationships can reach users on surfaces that have not opened a self-serve advertising portal of their own, and may never open one. If you treat AI as a genuine distribution channel rather than a single-surface experiment, that reach across publishers is what the decision turns on. The cross-surface model is the structural response to a channel that is, by its nature, spread across many surfaces at once rather than concentrated in one.

How to evaluate a hybrid DSP+SSP platform

A cross-surface DSP operating on top of an exchange and direct publisher relationships occupies a distinct structural position in this market: it aggregates buying across multiple AI surfaces while preserving the contextual precision that makes conversational advertising valuable in the first place, a combination that neither a generalist buying platform nor a single-surface walled garden can offer on its own. A hybrid platform structure combines its own SSP and direct publisher supply with programmatic buying across surfaces, which gives it pricing control and supply-quality guarantees that a pure demand-side aggregator, lacking owned inventory, cannot provide.

Evaluating a hybrid platform means asking a different set of questions than evaluating a walled garden. The first is reach: how many independent AI publishers does the platform have direct supply relationships with, beyond whatever walled-garden programmatic access (through partners like Criteo or StackAdapt) it has also secured? The second is contextual precision: does the platform's exchange carry genuine access to live conversational signal from its publisher partners, or does it rely on a shallower proxy layered on top of chat interfaces that were not built with advertising in mind? The third is control: because a hybrid platform owns inventory directly rather than only brokering access to someone else's, it can set and enforce supply-quality standards across its publisher network in a way a pure aggregator, dependent entirely on third-party terms, cannot.

Real-time prompt analysis is the parsing of what a user is actually asking an AI assistant at the moment they ask it, and it is the targeting primitive that makes this entire channel function, so building infrastructure to do that job requires an architecture designed for it from the start. A system built around keyword auctions or an interest graph assembled from past behavior cannot replicate what a real-time conversational matching engine detects about a user's immediate purchase intent, because the behavioral proxy and the live utterance are measuring fundamentally different things. If you are weighing a hybrid platform against a single-surface walled garden or a generalist DSP retrofitted for chat, test that distinction, reach across publishers combined with genuine access to live conversational meaning, against actual campaign data before you commit budget.