Why Architecture Decides Which Platforms Belong Inside AI Conversations and Which Do Not

A conversational AI advertising platform matches paid placements to the intent of a live exchange rather than a page, keyword, or stored profile. ChatGPT has a dedicated way to buy ads, while Google and Microsoft place campaigns that qualify within their AI surfaces. This is not the same as adding a chatbot to a website. The decisive filter is architecture, because serving conversational ads requires systems that can read, place, and measure against a changing session.

Why page-based ad systems cannot simply move into conversational environments

The page has a URL, an indexable topic, and a placement slot that exists before the user arrives. A conversation has no equivalent fixed object. It is constructed turn by turn for one person, so the relevant content and the placement decision emerge together, against what the user just said.

That difference changes the unit of targeting. Keyword bidding assumes a discrete query that can be matched against a string, while conversational intent develops across exchanges and can change during the session. Cookies and profiles gather signals across sites and visits, while the live exchange can show intent without cross-site tracking or a persistent profile. The distinction is between a system that allocates space against established content and one that must interpret content as it unfolds. Porting the former into the latter exposes a foundational design mismatch.

How the conversational environment actually processes an ad request

The conversational targeting model begins with a description of where an ad belongs. Its primary control is a natural-language context hint, a freeform description capped in length at the ad group level. The advertiser describes the relevant conversation. A comma-separated set of topics does not express the relationship between a user's goal, constraints, and decision stage.

At serving time, the unit of value is the conversational state. That includes whether the user is researching, comparing, or deciding, how specific the request is, and how far the session has progressed. Topic, vocabulary, depth, and apparent goal contribute together. In the relevance-weighted, second-price model described here, contextual precision can allow an ad to beat a higher bid, while the winner pays just above the next competing bid. Spending power is not the only deciding factor.

Placement then has to preserve the boundary between advertising and the answer. ChatGPT uses a visually distinct, labeled card below the response rather than inserting the advertiser's message into the answer itself. OpenAI calls this principle “answer independence” and makes the separation of ads from the answers ChatGPT gives part of the serving architecture. Though contextual matching is more like programmatic display than keyword bidding, reading a live exchange takes tools that page-based display infrastructure was not built to offer.

Which platforms let brands place ads inside live AI conversations, and why do their approaches differ?

ChatGPT built a dedicated route into conversational inventory. After three years of stating that it would not carry advertising, OpenAI launched ads on February 9, 2026. The pilot required a $200,000 commitment and quoted a premium CPM, with launch partners including Target, Adobe, Williams-Sonoma, and Albertsons. Access subsequently widened beyond the managed pilot. A self-serve Ads Manager opened on May 5, 2026, with no minimum spend.

Only logged-in adult users on Free and Go are eligible to see ads, while Plus, Pro, Business, Enterprise, and Education subscribers see none. OpenAI also blocks ads in mental and personal health conversations, though it began allowing them in general health advice contexts in April 2026. Suitability is therefore not solely an advertiser preference. The system excludes some conversations before campaign choices can determine placement.

AI Overviews and AI Mode bring the LLM experience into search, keeping the auction close to where answers are put together. Google's Ads Help page says that existing Search, Shopping, and Performance Max campaigns can qualify for placements within AI Overviews if they meet the requirements for AI-powered targeting. A separate conversational buying system is not required. The lesson for advertisers is to audit qualifying campaigns before assigning a new budget, because those campaigns may already serve in AI answers.

Microsoft Copilot similarly enrolls eligible campaigns and ad types automatically, without an opt-out or a guaranteed placement. Microsoft states that targeting considers the whole conversation within a session. Its “ad voice” serves a separate purpose. That term names the transitional summary that acknowledges the organic response and introduces advertisers.

Other surfaces reveal the filter through their commitment to trust. Perplexity built its proposition around answer credibility, making sponsored placement a potential threat to the reason users rely on its answers. Claude's absence from the advertising pool is a design commitment. These differences show why conversational reach cannot be assumed from a surface's popularity or technical capability.

Why the tier and reach constraints matter as much as the targeting model

Even a well-matched ad is useless if the buyer it is meant for cannot see it. ChatGPT's Free and Go eligibility creates a different addressable audience from its overall user base. Go is a paid subscription that still includes ads and whose users have chosen to invest in AI tooling, while Free users tend to be earlier in research or decision-making.

For B2B advertisers, the exclusion of Pro and Enterprise users creates a more consequential boundary. Buyers on those plans cannot see ChatGPT ads, however clearly their conversations show commercial intent. Targeting cannot cross that boundary. The surface consequently skews toward consumers and early-stage research rather than providing unrestricted access to professional and enterprise buyers. A strong contextual model does not remove the audience constraint.

Google and Copilot invert that trade-off through automatic eligibility without a tier exclusion. Campaigns that qualify can reach people across many conversations, but advertisers have less say over whether their ads appear there than they would with a dedicated way to buy conversational ads. Format imposes another boundary. Rufus's main format is built for a shopping assistant whose sessions already have transactional intent, so it cannot simply move to a general-purpose chat interface.

What prompt-level signals reveal that keyword and profile signals cannot

A keyword asks the advertiser to reconstruct intent. From a three-word query, a search marketer may need to infer the user's goal, budget, and position in a decision process. A conversation can state those things directly. It may contain the goal, the spending constraint, the shortlist, and the options already rejected, all in the user's own words. The distinction is not simply a larger quantity of targeting data, but a different kind of evidence.

Google assembles audience signals from prior behavior, while a ChatGPT exchange can reveal what someone wants now without requiring a historical profile for that signal to be strong. Session depth adds a temporal dimension as well. A request made after several comparisons has a different context from the same words at the start of research. A system that reads only the latest query misses how it fits into the session, even if it gets the topic right.

This richer input allows a conversational system to work with information beyond what a keyword or profile alone contains. The commercial effect can also occur elsewhere. Users may move from an AI exchange to the open internet to complete a purchase, making the session an upstream source of intent. That position explains both the targeting opportunity and the measurement problem that follows.

Why attribution breaks in a conversational environment and what that means for buyers

Page-based attribution is organized around a discrete, timestamped click on a known URL. That click is an anchor for linking an ad someone saw to a later conversion. Conversational influence need not produce that anchor. A user can encounter a brand while researching, close the chat, and return days later through a direct visit or branded search. The eventual purchase may leave no traceable connection to the exchange that helped shape it.

The resulting gap extends beyond missing clicks. Conversational interfaces also constrain ad load, typically showing zero to two sponsored placements per session, with ads appearing in a minority of commercial queries. When ads do appear, ChatGPT shows one placement, while the average for Google AI Mode is ~1.71. Because trust depends on users not perceiving answers as manipulated by advertising, platforms limit commercial exposure, leaving fewer observations for statistical optimization.

ChatGPT excludes its ad-free subscription tiers, while Copilot matches against a session. These are different exposure conditions. Combined with the absence of a standard click-to-convert path, they prevent conventional CTR and ROAS benchmarks from transferring cleanly. Early data can therefore be sparse because of the product's design.

Even with more campaign history, some of that mismatch will persist. The interface is built to serve the user's session, which can end without creating a trackable route to purchase. Last-click evaluation will consequently tend to undervalue upstream influence by crediting only the later, observable interaction. The practical implication for buyers is to design tests around that role. View-through attribution, branded search lift, and incrementality testing belong in the measurement approach.

What the architectural filter means for platforms that want to operate across more than one AI surface

Each AI environment has its own context-reading logic, formats, audience boundaries, and auction mechanics. Compatibility with one does not establish compatibility with another. A system that understands a conversation on its own surface may still lack access to the inventory where the same user's intent appears elsewhere. Its targeting ability and its reach are separate capabilities.

Generalist demand-side platforms approach that divide from the opposite direction. They have cross-surface reach, but page-level targeting does not equip them to interpret live conversational state. Adding that capability changes the underlying architecture. An AI ad network on a single surface may read sessions well but reach only that surface's users. Neither breadth without contextual understanding nor contextual understanding without breadth provides the infrastructure for the full conversational channel.

The system needs infrastructure built for conversations to bring those abilities together. It must read and bid against conversational signals across more than one AI surface, with direct supply relationships that give it access to the relevant inventory. This defines a distinct infrastructure category. Conversational AI advertising consists of paid, targeted commercial messages matched to live intent, requiring capabilities beyond a chatbot added to a brand's website or a page campaign given a conversational label.

For publishers operating AI interfaces, the same filter works in reverse. Their inventory represents an unfolding exchange rather than a pre-existing page slot, so monetization infrastructure must be able to interpret that exchange. A native conversational ad layer is therefore a structural requirement. Search and display technology can participate where the surface's own architecture supports it, as the integrated approaches demonstrate, but that does not make those systems portable across every AI interface. The boundary is what the infrastructure can actually read, reach, and measure.