The Media Companies and DSPs Staking Out Conversational AI Advertising Before the Channel Matures
Early backers shape platform formats and targeting models that competitors must now inherit.
Conversational AI advertising places paid units inside AI answer experiences. ChatGPT and Microsoft Copilot offer conversational placements, while Google monetizes AI Overviews through existing campaigns and Perplexity has retreated from sponsored answers. The companies staking out this channel are betting on more than access to a new audience. Durable advantage will depend on whether format, targeting, and inventory access fit together without compromising trust.
Why conversational AI advertising is being claimed now, before its rules exist
Two pressures are bringing the channel into existence before its norms are settled. AI platforms need more revenue than subscriptions alone can bring in, while AI-generated answers are weakening the paid-search channel that advertisers already use. OpenAI spent three years ruling out advertising before launching ads on February 9, 2026, against the arithmetic of a free-majority user base and operating costs in the billions. At the same time, paid-search click-through rates fall steeply when a Google AI Overview appears. Existing distribution is weakening as new inventory opens.
This is not an orderly handover between channels. Companies with different ways to make money are all deciding at once on format, targeting logic, inventory access, and the rules that protect trust. A platform’s decision about where an ad sits also shapes what buyers can purchase and what users understand to be independent advice. Those choices can become defaults before the industry agrees on standards. Early positioning is consequential because later entrants may inherit an architecture already defined by others.
What do ads inside AI conversations actually look like?
The defining format choice is where the ad sits relative to the answer. ChatGPT places a labeled sponsored card below the response, explicitly separated from the model’s text. Its “Answer Independence” principle makes the distinction clear. The response itself is not for sale. Google AI Overviews and Microsoft Copilot likewise place units beside or below responses, use existing campaign assets, and serve ads automatically when campaigns qualify, with no way for an advertiser to opt out.
University of Michigan researchers Jingyi Qiu and Qiaozhu Mei describe a broader ladder of commercial influence in Generative AI Advertising as a Problem of Trustworthy Commercial Intervention. Its tiers are product mentions, information framing, behavioral redirection, and long-term preference shaping. Deployed systems concentrate on the most visible, easiest-to-govern end of that ladder, while influence through framing and source selection lacks comparable disclosure frameworks. Separation is therefore a first-generation constraint. Mechanism-design research already proposes token-level auctions, RAG-based segment auctions, and sponsored LLM summaries that would move commercial influence into answer generation. Choosing a separate card is a trust bet that users can distinguish the paid placement from the advice around it.
How targeting works without keywords, and why that asymmetry favors early specialists
Conversational targeting changes the unit of relevance from a search term to a described situation. The context-hint model asks advertisers to express the conversational circumstances in which an offer belongs. Relevance weighting can let a precise advertiser with a small budget outbid a vague advertiser with a large one. That opening narrows as competitors learn to describe intent more accurately.
Microsoft Copilot extends this logic across the entire session. A user who asks about software integrations, expresses frustration with an existing solution, and then requests pricing comparisons has supplied a sequence of signals that no isolated keyword captures. The Michigan paper cites a study of nearly a thousand e-commerce websites over twelve months that found higher conversion rates and revenue per session from organic ChatGPT referrals than from paid social. That finding supports the intent-quality thesis, but it is not evidence of paid-placement performance. Learning which conversational signals point to commercial value is a skill that will stay useful. Early specialists can accumulate that knowledge while competition is thin, even though the initial auction advantage will not last indefinitely.
The platform map as of late 2026 — who opened, who retreated, and what each position signals
ChatGPT’s February 9, 2026 launch required a substantial upfront commitment and a premium CPM, with Target, Adobe, Williams-Sonoma, and Albertsons among launch partners. Self-serve Ads Manager opened May 5, 2026 with no spend minimums, first for US businesses and then for businesses around the world over the months that followed. OpenAI reported a significant annualized revenue run rate by August 31, 2026.
OpenAI’s September 16, 2026 announcement added an Ads Manager plugin for campaign creation and management through natural-language prompts, with Shopify and HubSpot as its first commerce and CRM partners. It also announced a test of “Sponsored Agents” with select US advertisers. OpenAI’s documentation identifies Free and Go users as the advertising audience, including the paid Go tier, while Plus, Pro, Business, and Enterprise remain ad-free. The boundary is both editorial and commercial. Advertising is separated from answers and allocated to particular subscription experiences.
Google takes a different route through existing Search, Performance Max, and Shopping campaigns, with no separate buy and no opt-out for qualifying campaigns. As commercial-intent AI Overviews expanded between late 2025 and mid-2026, Google kept monetization within its search ecosystem rather than establishing an independent buying channel. Perplexity reached a more restrictive judgment, retreating because even labeled, separated sponsored answers could make users suspicious of the whole response. Its retreat and OpenAI’s acceleration expose competing beliefs about the same trust boundary. Anthropic makes the strongest counter-position with no advertising or announced plans, revenue from enterprise contracts and paid subscriptions, and a Super Bowl campaign framed against AI advertising.
How DSPs and buying infrastructure are repositioning around the new channel
Microsoft closed Invest DSP on February 28, 2026, explicitly framing the decision as a pivot toward conversational advertising built on its AI investments. It named Amazon DSP as the preferred migration partner for displaced customers. Microsoft treated legacy display buying and conversational advertising as needing separate architectures, even as other companies began to connect them.
Amazon’s September 2026 unBoxed announcement merged its DSP and Ads console into Amazon Ads Agent. Its conversational layer, in expanded closed beta, supports campaign planning, management, and performance analysis through natural language. Amazon also introduced “Branded Conversations,” a multi-turn in-chatbot shopping unit, and “Sponsored Prompts,” which surface products in response to searches. US tests showed lower cost per impression and more unique customers reached.
The Amazon–OpenAI partnership makes the distribution bet concrete. Amazon Ads announced on September 10, 2026 that selected US advertisers could extend existing Amazon campaigns into ChatGPT through Amazon DSP, with Delta Vacations among the first pilot brands. This is the first major cross-platform programmatic bridge into ChatGPT inventory. Yet access alone does not settle whether established DSPs can accommodate the contextual precision that conversational placements require.
Viant introduced Outcomes, described as a “fully autonomous advertising product” built around its “AI Lattice Brain” optimization engine. Google released an open-source MCP server for its Ads API and introduced Ads Advisor and Analytics Advisor. These tools help run campaigns with AI, which is a different use of AI from buying ads inside conversations. Whether extended programmatic systems can combine cross-surface reach with the specificity of a 280-character context hint, or whether purpose-built buying layers will be needed, remains unresolved.
What the publisher side of the market reveals about inventory access and the traffic bargain breaking down
Conversational inventory is becoming contested as the old supply of publisher-page impressions weakens. AI-driven discovery reduces referrals to publisher-owned pages, eroding the inventory programmatic buyers have historically purchased. Reach plc, the UK publisher behind the Daily Express and the Daily Star, reported lower year-on-year digital revenue and a substantial decline in indirect programmatic revenue in its Q3 2026 trading update. Its figures illustrate the pressure without isolating every cause. The IAB’s April 2026 survey of US publishers found that a large majority reported moderate or significant business effects from AI-driven discovery, while a majority of small and mid-sized digital publishers reported reduced discovery referrals.
OpenAI signed deals with publishers including Dotdash Meredith and Condé Nast, compensating creators while legally incorporating their materials into ChatGPT responses. The publisher can earn licensing revenue even when no one visits its ad-supported page. For buyers, the structural consequence is greater concentration of conversational inventory within AI platforms, making direct supply relationships more valuable than access to the underlying publisher content alone.
What the trust and transparency problem means for structural bets made now
The Michigan researchers frame generative AI advertising as trustworthy intervention in the generative process, which changes what transparency must explain. Commercial influence can operate through framing, source selection, or behavioral redirection without producing a labeled sponsored object. Empirical research cited in the paper finds that ads woven into chatbot responses often go undetected, can match or outperform ad-free outputs on perceived helpfulness, and are susceptible to adversarial amplification. Helpfulness is therefore not a reliable disclosure mechanism.
The visible boundary preserved by major deployed systems as of mid-2026 is a governance response to that problem. It makes a sponsored unit identifiable without treating the organic response as commercial inventory. But Perplexity’s retreat challenges the sufficiency of that protection. Its judgment was that suspicion could spread from the ad to the answer itself, even when the two were separated. If that judgment generalizes, conversational advertising faces a higher trust cost than search because the assistant relationship depends more heavily on perceived neutrality.
If placement-level transparency is sufficient, early buyers can accumulate learning in a high-intent environment before competition thickens. If users want something closer to Anthropic’s advertising-free approach, systems built around sponsored conversations may struggle to grow. Neither outcome is established by launch momentum alone. Clear labeling and no influence on the organic response remain an ethical baseline and a practical prerequisite, but they are not proof that users trust the answer’s independence from the ad.
What early-mover positioning in this channel actually requires to be durable
A durable position requires compatibility across format, targeting, and supply. A labeled, separated format establishes the intended boundary of commercial influence, while precise context hints match a placement to the conversation. Inventory access then determines where that combination can operate. None of those capabilities substitutes for another. A buying relationship without contextual precision offers distribution without the channel’s distinctive targeting advantage, while targeting expertise without supply access cannot produce paid reach.
Automation also leaves part of the program outside any single platform’s control. Advertisers are still responsible for brand standards, the range of angles, and landing pages that match across channels. The Amazon–OpenAI bridge shows how buyers can get inventory by extending a DSP relationship, while a purpose-built buying layer needs its own direct supply relationships. Organic visibility and content licensing support participation in AI discovery, while a paid conversational advertising position requires inventory access.
Measurement is still the weakest link between signs of intent and proof of value. When people find information through an assistant, the AI may answer their question without a click, so attribution remains unresolved. Incrementality experiments, including the kind OpenAI has begun testing with Haus, Measured, and WorkMagic, are the current best approach, but the field is early. Launch dates and reported revenue cannot close that gap. The durable early-mover advantage is coherence across format, targeting, and inventory, constrained by trust and tested against incremental outcomes, rather than simply being first to buy an emerging surface.
