Native Ad Formats vs Display Formats Inside AI Interfaces

Display advertising was built for a world of fixed slots on static pages, and that world does not exist inside an AI conversation. This piece examines why display formats fail structurally once advertising moves into conversational AI interfaces, and why native formats are the only format logic built to survive there.

Banner and display ads are a structural bad fit for generative AI interfaces, not a creative or executional one. Display advertising is built on three assumptions that conversational AI eliminates entirely: a static page with fixed positions to carve up, a single keyword query to bid against, and a slot defined before any user shows up. Research on ad insertion in LLM-generated responses identifies three resulting failures that better creative cannot fix: contextual coherence fails because the ad has no mechanism to track intent that keeps shifting mid-conversation; computational efficiency fails because any insertion mechanism that adds meaningful latency degrades a real-time exchange users expect to feel instant; and social welfare fails because there are no static keywords left to auction, which is the pricing mechanism legacy ad markets depend on. Beyond those technical failures, a banner or interstitial dropped inside an LLM response lands either as a disruption that breaks conversational flow or as something a user mistakes for part of the AI's own output, a kind of trust damage display advertising on a webpage never had the power to inflict.

Why are banner and display ads a bad fit for generative AI interfaces?

Display advertising assumes a page that holds still long enough to carve out a rectangle for a banner, a sidebar unit, an interstitial. That assumption built the entire discipline: a slot gets defined in advance, a bid gets matched to a keyword, and an ad renders in a spot the page owner decided on before any user showed up. Research on ad insertion in LLM-generated responses shows that none of this holds once advertising moves into a chat interface. There are no static keywords to bid against and no fixed positions to fill, because user intent moves throughout a conversation rather than arriving as a single query against a stable results page. A banner dropped into an LLM response does not sit beside the content the way it would on a webpage. It sits inside the answer itself, which leaves it only two ways to land: as a disruption that breaks the flow of the conversation, or worse, as something a user mistakes for part of the AI's own output. The same body of research frames this as three separate technical failures display has no way to answer. Contextual coherence fails because the ad has no way to track intent that keeps shifting mid-conversation. Computational efficiency fails because if an insertion mechanism adds meaningful latency, it degrades a real-time exchange that users expect to feel instant. Social welfare fails because there are no static keywords left to auction, which is the mechanism legacy ad markets depend on to price inventory and allocate it efficiently. None of these are presentation problems that better creative or smarter frequency capping could fix. They are structural mismatches between a format built for fixed pages and an environment that has no pages. The experiential cost compounds the technical one. A flashing or static unit inside a conversation a user trusts to be helpful does more than get ignored: it breaks the premise that the assistant is giving an honest, unmediated answer, which is a kind of damage display advertising on a webpage never had the power to inflict.

What the native format is

Native advertising gets defined by its integration with the form and function around it, not by any particular look. In established digital channels, native units get assembled from modular pieces, a headline, an image, body text, a logo, and rendered to match whatever placement they sit in; they carry a disclosure label, but they read as part of the surrounding content instead of interrupting it. That distinction carries directly into an LLM context. A native ad inside a conversation is part of the conversation itself: contextual, showing up only when something in the exchange indicates a real need for it; valuable, adding information rather than just pushing a marketing line; and non-interruptive, extending the flow of the exchange instead of cutting across it. None of this is coincidental overlap. Native was not retrofitted to solve how advertising fits into AI systems; its defining traits happen to be the exact traits that demand.

Native-in-AI gets mistaken for two things it is not: AI-generated ad creative, a production tool for making ad assets faster that has nothing to do with where or how an ad gets placed, and search advertising repackaged with an AI summary running alongside it, which still maps onto keyword-triggered slots sitting next to results; native-in-AI is placements living inside a conversational response.

How conversational intent signals change the targeting logic entirely

What native-in-AI gains access to is not just a better-fitting format. It runs on a fundamentally different signal than anything search or display advertising has worked with. A search query tells an advertiser only what someone typed, in a single moment. A conversation can tell an advertiser what someone actually needs, often before the person asking has worked it out. Research on LLM ad insertion frames the mechanism this way: in a search engine, intent gets encoded by a static keyword; in an LLM chat, intent moves throughout the exchange. An exchange about a family vacation to New York might move from attractions to accommodations to transportation to budget within a single session, and each of those segments carries its own targeting signal rather than one fixed query standing in for the whole conversation.

The advantage runs deeper than volume of data. A search query tells an advertiser what someone typed. An AI conversation can carry intent, stated limitations, expressed preferences, and a sense of where the user sits in their decision-making process. If a user asks about hotels but mentions a budget constraint and two young children, that gives an advertiser a sharply different signal than the same surface-level question with no context attached, even though a keyword-matching system would treat both queries the same.

This is the reasoning behind the genre-based bidding approach proposed in the academic literature. Rather than bidding on individual user queries, which raises privacy concerns and gets computationally expensive fast, advertisers bid on stable, high-level semantic clusters, referred to as genres, that group conversations by shared context. So you can target without exposing raw, real-time user data to anyone bidding on it. OpenAI's live implementation reflects the same logic in practice: advertisers write context hints, prose descriptions of the conversations and topics where an ad applies, instead of exact-match keywords, and the matching runs on conversational context and landing page content rather than a keyword auction. Placing an ad at the right point in a response, surrounded by text actually about the topic the ad addresses rather than some other topic from earlier in the same session, both lifts effectiveness and cuts disruption in a way page-level display targeting was never built to achieve. User prompts reveal intent at a depth that keywords and demographic profiles cannot match, so targeting in conversational AI runs on a genuinely different signal, one that appears in preferences, constraints, and decision stage all in the same exchange. Real-time prompt analysis, the mechanism behind conversational AI ad platforms, turns into a real advantage here: it reads the full context of what someone is asking rather than just the words they typed.

Requirements for a technically sound native insertion framework inside an LLM

Getting native advertising to work inside an LLM response means solving three problems at once: transparency, timing, and coherence. No one of these can be solved in isolation, because a fix for timing that ignores coherence just produces a well-timed but irrelevant ad, and a fix for coherence that ignores transparency produces exactly the deceptive blending the FTC prohibits. The approach that resolves all three at once is decoupling ad insertion from response generation, and researchers from the University of Michigan, Peking University, the University of Hong Kong, and Tsinghua University's BIGAI lay out that framework in a paper published on arXiv.

The first piece of the framework decouples ad insertion from response generation itself. The model generates its answer first, with no commercial influence shaping what the answer says, and only afterward does the system insert ads at appropriate points in the completed response. That ordering removes the risk that the model hallucinates ad content or that advertising pressure quietly reshapes the substance of the answer. The second piece decouples bidding from individual user queries. Advertisers bid on genre-level semantic clusters instead of specific queries, which cuts computational load since no latency gets added to the live response, reduces privacy exposure since no sensitive per-query data reaches advertisers, and produces close to optimal social welfare under VCG auction mechanics, the same auction logic that underpins much of programmatic advertising already.

Transparency sits in the framework as a legal requirement. FTC policy holds that advertising not identifiable as advertising counts as deceptive if it misleads a consumer into believing it is independent or impartial rather than paid placement from a sponsoring advertiser. The framework requires ads to be marked wherever they appear, and OpenAI's live implementation does this with visually distinct tinted boxes and clear sponsorship labels, the same disclosure logic that governs native advertising across established channels. Coherence measurement is the hardest problem the framework takes on, because you cannot settle whether an inserted ad actually fits the surrounding conversation with a simple rule. The researchers built an automated model-based scoring metric to estimate contextual coherence, and it shows a Spearman's ρ of roughly 0.66 against human ratings, outperforming most individual human evaluators. That gives the system a scalable proxy for the judgment that decides whether a native insertion actually helps the conversation or quietly damages it. Without this kind of decoupling, any commercial incentive tied to topic categories risks distorting which topics get the most complete answers, not necessarily inside any one response, but across many responses over time.

How the first real implementations handle format in practice

OpenAI's ChatGPT advertising launch is the most fully documented implementation of native-in-AI advertising available, and it confirms the format logic described above while surfacing exactly where execution still lags theory. Ads appear in visually distinct tinted boxes inside the conversation interface, clearly labeled as sponsored, without interrupting the model's actual answer, a direct expression of the decoupling principle. Advertisers write prose context hints describing the conversations where an ad applies, and OpenAI states that these are not exact-match keywords and do not guarantee delivery into any specific conversation; matching draws on conversational context, landing page content, ad title, ad copy, the context hints themselves, and, only when a user has ads personalization turned on, signals from that user's broader ChatGPT activity. The buying model includes cost per click, where advertisers pay only when a user actually clicks, so spend ties to action taken after seeing the ad rather than to raw impressions. A conversion-optimized cost-per-click option lets advertisers run campaigns against specific downstream events, a purchase, a lead submission, a sign-up, with predicted conversion likelihood feeding into the auction and delivery decision. Confirmed launch partners include Target, Adobe, Williams-Sonoma, and Albertsons.

A newer format points toward where the native logic goes next. OpenAI's global ad solutions lead, Colleen Coulter, named sponsored agents as the future of AI-native advertising, a format that lets a user chat directly with a brand-sponsored agent after clicking into an ad, pushing the conversational format past the click and into an ongoing exchange with the brand itself. Amazon's Ads Agent marks a different but related direction: it brings agentic AI capabilities into advertising workflows and shows that the native-in-AI pattern is spreading past pure chat interfaces into AI-assisted buying and creative tools.

The honest gap in all of this sits in measurement. Buyers have flagged that prompt-level reporting is not yet available, so advertisers cannot see what a user was asking when an ad got served, and they are still asking for negative targeting capabilities. That is a limitation of early-stage infrastructure, not evidence against the format logic itself.

Advertising format across the major AI platforms

Format compatibility solves only half the problem for an advertiser. Beyond format compatibility, advertisers also need to know where inventory actually exists, and the major AI platforms have landed in genuinely different places. ChatGPT has launched advertising using the native, embedded format described above, and it has confirmed sponsored agents as a named future format. Gemini carries no ads inside the chatbot interface itself, but Google AI Overviews is expanding across search results pages; advertisers whose campaigns qualify for placement there get no opt-in or opt-out, though publishers can now opt their own content out of AI Overviews through Google Search Console. So participation in AI-adjacent placement may already be happening for some advertisers without any explicit decision on their part. Claude has taken the firmest opposite position: Anthropic draws its revenue from enterprise contracts and paid subscriptions and ran a Super Bowl campaign arguing against AI advertising outright, a position carrying real weight given the size of the company making it, not a marginal objection from a smaller player.

No single AI surface currently matches the reach of legacy digital channels, and the surfaces that do carry native advertising each run their own format and targeting implementation. Because of that fragmentation, advertisers trying to operate at any scale in this environment need the capacity to buy across multiple surfaces rather than build toward one platform's idiosyncrasies.

The trust architecture of AI interfaces and the stakes for format choice

An AI assistant holds a different relationship with its user than a webpage holds with its reader. Users treat what a chatbot tells them as advice from something that knows what it's talking about. That difference makes the choice between native and interruptive formats an ethical one.

A banner on a webpage sits visibly outside the editorial content surrounding it, and everyone reading understands the distinction without needing to think about it. An ad inserted into an LLM response sits inside what the user experiences as an answer coming from an assistant they trust. The gap between a placement that reads as clearly native and labeled, and one that reads as interruptive or unlabeled, decides whether a user takes the ad as a useful, contextual suggestion or as a manipulation of a source they believed was giving them a straight answer. Native ads in conversational AI earn that trust by staying integrated into the conversation itself: surfacing only when relevant, carrying a clear disclosure label, and adding to the exchange rather than breaking it. Platforms built around this principle match ads to the semantic intent of a given moment in the conversation rather than to a pre-sold slot or keyword tier, disclosing them as advertising while keeping them inside the flow of a genuinely helpful exchange. Display never had to answer to this standard, because display was never mistaken for the thing it sat beside. Native-in-AI only works because it accepts the standard rather than trying to avoid it.