Begin with the event you are trying to count. An agent requests a page, takes the text, and leaves. No script runs, no pixel fires, no viewability threshold is crossed, and the human who prompted the agent may never see your URL at all. Every metric the display business rests on assumes a rendered impression in front of a pair of eyes, and none of those assumptions survive the retrieval. So measurement has to be rebuilt from two ends that do still exist: the read itself, which the publisher can observe in its own edge logs, and the outcome, which the advertiser can observe in its own sales data. Everything between those two points is currently unstandardised, contested, and controlled by platforms with little commercial reason to open it up.
Why the existing impression standard does not apply
A conventional display impression is defined by rendering and by opportunity to see. The Media Rating Council viewability thresholds, the counting rules, the invalid traffic filters and the whole audit apparatus all assume a browser, a viewport and a person. An AI agent has none of these. It is a non-rendering client, which is precisely why publishers have historically classified this traffic as invalid and filtered it out of their reporting rather than trying to sell against it.
That leaves three separate things people mean when they say they want to measure advertising in AI, and conflating them is the main source of confusion in the market:
Organic visibility. Whether an AI system mentions or cites a brand or publisher at all, and how favourably. Nobody paid for the placement.
Paid placement inside the answer. An advertisement bought from the platform and rendered in the assistant's interface, of the kind OpenAI now sells in ChatGPT.
Content encountered by an agent. Material a publisher places in the page that the agent reads, which may be sponsored, and which the user never sees in its original form. This is the newest of the three and the least standardised.
The first now has a framework. The second is measured by the platforms themselves. The third is where the argument is.
What the IAB has actually standardised so far
On 3 August 2026 the IAB published Measuring Visibility in the AI Era, a 36-page set of measurement guidelines for tracking brand and publisher visibility in AI-powered discovery. The stated problem was market fragmentation rather than absent effort. By the IAB's own count, more than 20 companies now sell AI visibility measurement tools, each using different methodologies that can produce different answers for the same brand or publisher. "Consumers are increasingly discovering and considering brands and products in AI platforms, but measurement frameworks haven't kept pace," said Caroline Giegerich, the IAB's vice president for AI.
The framework organises metrics into a causal hierarchy it calls the 4 P's of AI Visibility.
Presence. Does the brand or publisher appear at all? Metrics are Mention Rate, Citation Rate, Share of Voice and Visibility Momentum. The gap between Mention Rate and Citation Rate is treated as a signal in its own right: one measures whether you are spoken about, the other whether you are relied upon.
Prominence. Where and how prominently do you appear? A single metric, Position, measured against the formatted response a user actually sees rather than the model's raw output.
Portrayal. In what context, and with what accuracy? Sentiment, Framing, Hallucination Rate and Factual Inaccuracy Rate. This is the brand safety dimension that has no analogue in display, and the framework requires hallucinated and merely inaccurate mentions to be reported separately, per platform, with flagged examples surfaced to the client rather than quietly dropped.
Persuasion. Does any of it drive action? Recommendation Strength and Post-Citation Click-Through Rate. The IAB explicitly designates click-through as a bridge metric and defers the attribution methodology behind it to a separate framework still to come.
There is a parallel publisher track, which matters more to media businesses than the brand metrics do. It adds Citation Decay Rate, which measures the shelf life of a piece of content inside AI systems, Content Utilization Rate, which measures whether your work was substantively drawn upon or merely name-checked, and Attribution Clarity, which grades citations from a full linked reference with title and date down to a bare publication name. The IAB's own framing is that publishers can bring these numbers into licensing conversations.
Two things the framework deliberately does not cover, and they are the two that matter most for anyone selling to agents: it measures organic, non-paid visibility only, and it does not address agentic media buying specifications or commerce attribution. Paid placement measurement and publisher downstream attribution are named in the document as gaps for future work.
Directional versus decision-grade, and why buyers should care
The most immediately useful thing in the framework is not a metric. It is a two-tier quality classification that gives a buyer a way to ask whether the data they are paying for is fit for the decision they are about to make.
Directional measurement identifies patterns and signals trends. It supports early signal detection and competitive awareness. It is explicitly not sufficient for budget allocation, provider selection or executive strategy. Decision-grade measurement meets a higher bar across query volume, sample size, prompt type coverage, testing cadence, reproducibility, data validation, methodology documentation, platform coverage and multi-platform aggregation.
Some of the thresholds are concrete enough to use in a procurement conversation. A programme running fewer than 50 queries is exploratory rather than directional. Prompt coverage spans four intent types, informational, comparison, recommendation and transactional; directional work needs at least two, decision-grade needs all four with results segmentable by intent. Reproducibility at decision-grade means defining acceptable variation ranges within a seven-day window and reporting confidence levels. Cross-platform figures have to be reported per platform, not blended into a single number.
The reason this bites is that AI answers are non-deterministic. The same prompt does not return the same answer twice, and the framework's position is that single-response measurement is not measurement at all. Visibility on a query is a distribution rather than a value. Independent work supports the caution: Similarweb research published in November 2025 found citation sets changing by roughly 50% each month, with only 11% of citations overlapping between major AI platforms. Semrush, analysing 126 million United States AI search prompts collected between January and April 2026, found that only 36 of more than 1,200 tracked brands appeared in the top 100 most-mentioned list on every platform in every month.
Against that volatility, the framework's governing instruction on disclosure is the sentence to remember when a vendor demonstrates a dashboard: where a provider cannot or will not disclose against a required item, that absence is itself to be treated as a signal.
The harder problem: who gets credit
Visibility is the tractable half. Attribution is the half that has stalled, and it is the one being drafted now. Digiday reported on 24 August 2026 that the IAB has a framework due on 12 November covering how to attribute and credit conversions influenced by AI. Giegerich described the goal as creating "a shared framework for measuring and crediting AI's role in conversions, especially when traditional signals are minimized", and said the framework will likely separate AI impact into two categories: when AI serves something to a user, and when AI helps that user make a decision.
The reason this is difficult is easy to state and hard to solve. If an agent reads a product page, weighs it against competitors and buys on a user's behalf, the evidence trail that display advertising relies on does not survive the journey. UTM parameters and referrer data are stripped or suppressed. There is no cookie, no session and often no click. So the question becomes what evidence should count as proof that an ad encountered by an AI contributed to a conversion, and there is no settled answer.
There is also no settled view on whether the practice is legitimate. Giegerich put the disagreement inside her own working group plainly to Digiday: "What does it mean to advertise to an agent? One side might think that 'this is an interesting test,' and the other side is like, 'that's deception.'" Asked what had been hardest for the group to agree on, she said "everything". That is not a drafting problem. It is a market that has not decided what it is doing.
Publishers have an obvious position and are pressing it: their content informed the answer, and they do not want to be excluded from the attribution conversation. Jaime Schultheis, head of global data partnerships at Bombora, told Digiday the opportunity was one of overdue reciprocity, noting that many large technology partners have grown audiences off publishers with very little flowing back.
Why the platforms are the binding constraint
Even a good standard needs data, and the data sits inside systems that do not open. Michael Bishop, co-founder of the AI-native advertising platform OpenAds, set out to Digiday the four questions any framework has to answer: what is being measured, who is doing the measuring, how, and at what layer. His warning is that a standard can answer all four and still be defeated by access.
Bishop's precedent is instructive. Describing how measurement vendors operated inside Facebook, he said they "were not running their own JavaScript tags. Facebook actually coded the integration to measure their own homework effectively, and had the measurement vendors basically manually testing to rubber stamp that integration. So the black box was basically maintained, and the measurement vendors were effectively operating as like a third-party auditor trust layer." His verdict on that arrangement repeating here: "it does not look very good for anyone."
There is some movement in the other direction. Digiday has reported that OpenAI opened its ChatGPT ads manager in the United States while promising third-party measurement and cost-per-acquisition bidding, and that in June 2026 its global head of ads, David Dugan, described third-party measurement as a natural step for the platform. That is worth having. It is also, for now, a promise about the platform's own advertising products rather than about the agent traffic hitting publisher servers, which no platform currently reports on to the publisher at all.
What a publisher or advertiser can measure today
The standards are months away and partial when they arrive. Four things are available now.
Instrument the read at the edge. AI and agent retrievals are visible in server and CDN logs even when they are invisible in client-side analytics, which files most of them as direct or discards them entirely. Log-level classification of agent traffic by user agent, verified bot identity and behaviour is the only first-party record of what was actually consumed, and it is the raw material every downstream metric depends on.
Track organic visibility with the IAB vocabulary, and hold vendors to the disclosure list. Mention Rate, Citation Rate and Share of Voice are now defined terms. Ask any provider which of the four data collection architectures they use, whether live web retrieval was enabled during collection, how many queries and responses sit behind a reported number, and whether the figure is directional or decision-grade. Treat refusal as an answer.
Measure the outcome end honestly. Branded search lift, direct traffic composition, assisted conversions and post-citation click-through are all imperfect and all better than nothing. Report them as a converging set rather than as a single attributed number, which is exactly what the framework recommends for directional data.
Write measurement into the commercial terms. If you are selling anything into agent-read inventory, the reporting definition belongs in the insertion order, because there is no industry default to fall back on and no auditor to appeal to.
Where blankspace sits in this
blankspace operates at the layer the frameworks have not reached yet: detecting Live Search Agent retrievals at the CDN edge and treating the read as a measurable, monetisable event rather than an unclassified log line. The relevant point for this discussion is narrow. Edge detection produces a first-party record of which agent fetched which content when, which is the one piece of evidence in this chain that a publisher owns outright and does not have to request from a platform.
That is deliberately a smaller claim than the category sometimes makes. It does not establish that a model used what it read, it does not prove influence on an answer, and no vendor in this space, blankspace included, can currently demonstrate the causal link between a retrieval and a purchase that the IAB's November framework is trying to define. What edge measurement gives a publisher is a defensible count of the demand, which is the prerequisite for pricing it and for arriving at any attribution conversation with data of your own rather than data supplied by the counterparty.
Frequently asked questions
Does an AI agent reading my page count as an ad impression?
Not under any current standard. Conventional impression counting requires a rendered advertisement with an opportunity to be seen by a person, and an agent retrieval satisfies neither condition. In practice this traffic is generally filtered out as invalid rather than counted. The IAB's August 2026 visibility framework covers organic mentions and citations, not impressions served to agents, and explicitly leaves paid placement measurement to future work.
What is the IAB's AI visibility framework, in one line?
Measuring Visibility in the AI Era, published 3 August 2026, is a 36-page set of guidelines that defines a shared vocabulary for brand and publisher visibility in AI answers, organised as the 4 P's of Presence, Prominence, Portrayal and Persuasion, plus a quality standard separating directional data from decision-grade data. It does not rate vendors, prescribe tools or cover paid placements.
When will there be an attribution standard for AI-influenced conversions?
Digiday reported in August 2026 that an IAB attribution framework was due on 12 November 2026, drafted with a working group of technology companies, publishers, agencies, measurement vendors and brands. It is expected to distinguish between AI serving something to a user and AI helping a user decide. A published framework is not the same as adoption, and the underlying evidence problem, that platforms hold the signals, is not solved by a document.
How do I tell a serious AI visibility vendor from a weak one?
Ask for the disclosures the framework requires rather than for a demo. Which platforms and model versions are covered, how the prompt library was built and how large it is, whether queries are synthetic or drawn from real search behaviour, which collection architecture is used, whether live web retrieval was on or off, and how variation across runs is handled. A provider positioning itself as decision-grade should supply all of it in writing.
Should publishers be selling ads to AI agents before measurement exists?
That is a commercial judgement rather than a technical one, and the market is openly split on whether it is a legitimate format or a deceptive one. What is clear is that a publisher selling into this today is selling without an agreed counting standard, without third-party audit and without a settled position from the platforms. Digiday has reported that Perplexity blocked Time's agent-facing ads and called them deceptive. Instrumenting and pricing your agent traffic is sensible now. Selling guaranteed outcomes against it is not.