Viewability and Attention Standards for AI Chat Ad Formats

Viewability, as the industry has used the term since 2014, measures whether an ad had the chance to be seen. It says nothing about whether anyone actually looked at it, and in AI chat interfaces, where an ad appears inline with a conversational answer rather than on a scrollable page, that gap stops being a footnote and becomes the whole problem. Attention, not viewability, is the only measurement currency that fits an environment built around reading a response rather than scrolling past one.

The original IAB standard set a low bar on purpose: an ad counts as viewable if 50% of its pixels sit on screen for at least one second, two seconds for video. That threshold was built to solve a specific, narrower problem: ads loaded far below the fold that nobody ever scrolled down to see, counted as delivered impressions regardless. Compared to raw impression counts, requiring some minimum of pixel-and-time was a real improvement. But the standard carries an assumption baked into its foundation: an ad lives on a page, spatially apart from editorial content, and "opportunity to be seen" can be approximated with geometry and a stopwatch. Nobody designed it to handle an ad sitting inline with a conversational answer, where there's no fold, no scroll race, and no clean line between the ad and the content around it.

The mismatch shows up even in the environments viewability was built for. Campaigns routinely clear 70 to 80% viewability while brand recall stays flat, quarter after quarter. The metric confirms delivery. It was never built to confirm whether the ad landed, and treating it as a proxy for attention is a category error the industry has mostly tolerated because there was nothing better to use.

What makes AI chat a structurally different ad surface

A user typing a question into a chat interface and reading the answer that comes back is doing something different, attentionally, than a user scrolling a feed. The posture is active, sequential, and focused on one thing at a time, not the ambient half-attention of a passive scroll. That difference matters for measurement because most attention theory built for display and social assumes a baseline of low engagement that chat simply doesn't share.

Ads currently show up in AI chat in several places: an inline sponsored card after the answer finishes, a unit running alongside the conversation, a sponsored follow-up suggestion chip offering the next question, and a brand mention grounded directly in the generated response. Each has a different relationship to the answer itself. Some sit independent of the response stream; others are woven straight into it. That distinction changes both how much attention a user actually gives the ad and, separately, how technically feasible it even is to measure that exposure.

Woven-in formats raise a problem with no real precedent in classical ad auctions: an inserted ad can change the flow, tone, specificity, and length of the entire generated response around it. Call it a generative externality. A banner sitting next to an article doesn't rewrite the article. An ad grounded in an LLM's answer can reshape the answer itself, and no pixel-based measurement approach has any way to detect that it happened.

There's also no fold to speak of, no spatial separation to instrument, and no stable, persistent DOM element sitting there the way a display banner sits on a page. JavaScript viewability tags were built around exactly those assumptions, and none of them transfer cleanly. Latency adds a further wrinkle: the working default for AI app publishers is an after-answer card rendered independently of the response stream, which behaves nothing like a page load and its associated timing events. Put together, the issue isn't that viewability standards perform poorly in chat. It's that they're measuring an object that doesn't exist in this environment.

Where the IAB/MRC Attention Measurement Guidelines land and what they actually specify

The IAB and the Media Rating Council published the Attention Measurement Guidelines in November 2025, built with input from more than 200 members spanning advertising, media, and measurement companies. The framing inside the document is deliberately modest: attention is a complementary signal, not a replacement for viewability or conversion metrics. It's meant to sit as a layer between the two, adding information rather than displacing what already exists.

The guidelines define distinct measurement tiers. Exposure-based measurement asks whether the ad appeared on screen in a viewable position, which is essentially viewability under a different name and functions as the floor, not the ceiling. Engagement-based measurement covers whether the viewer actually did something: clicks, time in view, scroll depth, and interaction patterns. Outcome-based measurement looks past the ad entirely, at the business result it produced.

Four measurement methodologies get validated under the guidelines. Data signal-based measurement covers JavaScript tags, the Open Measurement SDK, and server-to-server integrations, capturing things like time-in-view, scroll depth, audibility, interaction patterns, and screen orientation. Visual and audio tracking covers eye tracking, gaze tracking, facial coding, and presence monitoring. A third category covers physiological and neurological observation. A fourth covers panel- and survey-based methods, the older but still useful tools of asking people directly what they remember.

These guidelines now form the basis for MRC accreditation audits of attention vendors, which is the mechanism meant to fix years of vendor fragmentation. One number from the guidelines is worth sitting with: most digital ads fail a 2.5-second attention-memory threshold, because they were built and optimized for delivery, not for holding anyone's focus. The Playbook gives the industry a consistent framework to start measuring against that threshold rather than each vendor picking its own bar. A companion resource, the CIMM and IAB Attention Measurement Playbook for Marketers, drawn from more than 40 industry interviews, walks through how to operationalize these metrics across live campaigns. And the guidelines are explicit on one caution that matters more than it might first appear: attention is not a currency, and it is not a binary yes-or-no outcome. It's probabilistic, and any measurement approach that treats it otherwise is misreading its own data.

Which parts of the attention framework map onto chat formats and which do not

Some of this transfers into chat without much friction. Engagement-based signals, click-throughs on a sponsored card, how long the card stays visible, whether a user actually taps a suggestion chip, are all things a chat interface can log directly. Panel- and survey-based methods transfer just as cleanly, since brand recall and recognition studies can run after a session regardless of what surface the ad appeared on. Outcome-based measurement gets a foothold too: conversion tracking that OpenAI has started rolling out gives advertisers a downstream signal to work with, even though mid-funnel attribution is still thin.

Data signal-based exposure measurement is where the framework breaks down. Scroll depth and pixel-on-screen math both assume a page a user can scroll through. In a response that streams in token by token, pixel coverage at any given instant tells you almost nothing about whether the user actually read the text sitting next to the ad. The whole infrastructure built around the OM SDK and JavaScript tags exists for display units and video players, and none of that instrumentation was designed for a conversational stream. Audibility metrics, meanwhile, are simply irrelevant to a text-based chat format, and screen orientation is a weak signal at best on a single-surface text interface where there's not much else going on.

No current method closes the more interesting gap: whether a user's attention on the answer spills over onto an adjacent sponsored card, and whether a brand mention woven into the response even registers to the user as an ad at all rather than as part of the answer itself. Research in this space has been candid that the industry hasn't yet answered whether LLM ads can be held to the same measurement rigor as the rest of a media plan.

Readiness varies sharply by surface. The after-answer sponsored card is more tractable than other formats, since it's spatially separate from the response and behaves enough like a display unit to instrument with adapted versions of existing tools. The sidebar sits in similar territory, close enough to a display sidebar that existing methods partially apply. The sponsored follow-up suggestion chip offers a clean engagement signal once tapped, but almost nothing on pre-interaction exposure, and Perplexity pulled this format in February 2026, in part over trust concerns tied to exactly this ambiguity. The response-grounded brand mention sits at the bottom: woven into the text with no clear pixel boundary, and neither attention measurement nor basic ad recognition has a real answer for it yet.

Why pre-standard vendor fragmentation makes the gap worse and what accreditation is meant to fix

Before November 2025, every major attention vendor built its own definition of attention, and none of them lined up. Adelaide runs an AU score. Amplified Intelligence uses a neurometric approach. IAS and DoubleVerify each run multi-signal models blending exposure and engagement data. None of these scores were built to be comparable across vendors, which meant a high score on one platform carried no guaranteed relationship to a high score on another.

For a buyer, that meant no reliable way to tell whether attention scores meant the same thing from one platform to the next, let alone whether LLM inventory could be judged against search or social inventory on any common basis. MRC accreditation is the structural answer: vendors seeking accreditation have to disclose methodology, meet the guidelines' minimum standards, and submit to audits. That creates, for the first time, something like a comparable baseline across vendors.

The same fragmentation is about to happen again, one layer up, in AI chat specifically. Without chat-specific measurement standards, every AI publisher and every attention vendor will end up defining "attention in a conversation" its own way, repeating the exact problem accreditation was built to solve, just in a new environment.

Some of the infrastructure to prevent that is already underway. IAB Tech Lab's AAMP, the Agentic Advertising Management Protocols, was formally named on February 26, 2026, and is building out transaction, delivery, measurement, and privacy standards for agentic and AI-driven environments; AAMP 2.3 is the current release, adding enterprise-ready deployment support and firmer pricing integrity controls. Separately, IAB Tech Lab formed the AI Content Monetization Protocols Working Group in August 2025, driven partly by publishers watching their traffic drop as AI experiences strip their content of surrounding context. Measurement and attribution problems are tangled up directly with that publisher-side crisis: an underlying content and traffic problem, sitting one step upstream of the ad measurement problem.

The honest read: accreditation infrastructure already exists and works reasonably well for legacy formats. Extending that infrastructure to chat-specific environments is still early work, and buyers should not assume it's further along than it is.

What conversational intent signals offer that pixel measurement never could

A prompt typed into a chat assistant carries more than a keyword ever could. A multi-turn conversation about a purchase reveals category interest, sure, but also circumstances, hesitations, and the actual criteria someone is weighing, information a search query never surfaces on its own.

A growing share of users now start their pre-purchase digital journey inside an AI chat interface, with some verticals like travel seeing particularly strong adoption. A user arriving at an LLM has often already searched around and built a shortlist, which means the LLM is functioning less as a discovery layer and more as a validating sounding board for a decision already in motion. That changes what an ad impression at that exact moment is worth, and it's worth more than a comparable impression earlier in the funnel.

Prompt-level context targeting, matching ads to topic, conversation history, and inferred intent, is already running in prototype systems today. The targeting logic has outpaced the measurement infrastructure meant to validate it, which creates an odd inversion: advertisers can aim at high-intent moments with real precision but can't yet measure attentive exposure inside those moments with anywhere near the same rigor. The channel's strongest asset sits only partly monetizable until measurement closes that gap.

None of this means the channel lacks value. It means the industry built its measurement infrastructure for pages, and now has to build a parallel version for conversations, a genuinely different engineering problem, not a smaller version of the old one.

What buyers and publishers should actually do while the standards are still forming

Buyers should treat the engagement-based tier from the IAB/MRC guidelines as the working measurement layer for chat ads today: click-through on sponsored cards, whether a session continues afterward, downstream conversion wherever OpenAI's conversion tracking is live. Porting display viewability KPIs directly into a chat buy is a mistake. A campaign optimized against the old 50%-pixels-for-one-second standard will optimize toward the wrong target entirely in a streaming-response environment, and the money will follow a number that doesn't mean what it used to mean.

Panel- and survey-based brand recall studies are worth running alongside any chat campaign, since they transfer cleanly across surfaces and give buyers something genuinely comparable across channels. Any attention vendor applying scores to AI chat inventory should be made to disclose methodology up front. The accreditation framework exists, but the chat-specific version of it isn't finished, and buyers shouldn't assume a vendor's badge covers ground it hasn't actually covered. LLM exposure belongs in the same measurement view as search, social, and CTV from day one, because without a consistent cross-channel picture, there's no way to tell whether AI advertising is growing demand or just redistributing it from somewhere else.

For publishers, the after-answer sponsored card is the highest-readiness format on offer right now, and other surfaces carry more measurement uncertainty and, with it, more friction when demanding buyers start asking questions. Instrumenting what's instrumentable today, card visibility duration, interaction rate, post-ad session behavior, builds a baseline that can be audited later even if the signals themselves are imperfect. Engaging directly with AAMP and CoMP working groups matters too, since the standards being drafted now will decide what measurement infrastructure publishers are required to support down the line.

Buying and measuring consistently across all these surfaces gets a lot harder without some platform-side layer built specifically for it, something that reads conversational context and connects to AI publisher supply directly, letting advertisers apply one measurement logic across surfaces instead of negotiating a separate approach with every AI publisher one at a time. That's the structural role conversational-AI advertising DSPs are built to play. Attention measurement for AI chat will likely develop the way it developed for display: through a mix of vendor competition, industry standard-setting, and accreditation pressure. It only gets there, though, if buyers treat measurement as a condition of spend rather than a detail to sort out later.