The IAB's AI Visibility Metrics: The 4 P's and the Decision-Grade Bar

The AI visibility market now has shared definitions. On 3 August 2026 the IAB published Measuring Visibility in the AI Era, which names the metrics, sorts them into a hierarchy, and sets a bar for when a number is good enough to spend money against. More than 20 companies sell tools that claim to measure how brands appear inside AI answers, using different query sets, different engine coverage, and different definitions of the events they count. This is the first attempt by a neutral body to make those numbers comparable.

Key takeaways

The document is measurement guidance for a market that grew faster than its vocabulary. Caroline Giegerich, the IAB's VP of AI, framed the gap in the announcement: "Consumers are increasingly discovering and considering brands and products in AI platforms, but measurement frameworks haven't kept pace."

Two details matter more than the metric list.

The first is what the document declines to be. As AdExchanger reported, the IAB deliberately avoided calling this a standard, or even a framework, on the grounds that standardization needs a stable landscape and AI search does not have one yet. It also makes no vendor recommendations. So the correct read is shared vocabulary plus quality criteria, and the correct response to any tool that claims to be "IAB compliant" is a raised eyebrow.

The second is the adoption number underneath it: only 16% of brands systematically track AI visibility today, a figure Marketing Dive traces to McKinsey CMO surveys. That is the actual state of play. Most of the argument about which AI visibility metric is best is happening among the 16%.

The hierarchy is the useful part, because it is ordered by how close each metric sits to a business outcome. The IAB's own definitions, from the release:

Three things stand out.

Presence separates mention from citation, and that separation is load-bearing. Being named and being linked are different events with different fixes, which is why Semrush and Kevin Indig found 61.7% of brand appearances in AI answers were citations with no brand mention. A single "visibility score" that blends the two hides which one you are failing.

Portrayal treats accuracy as a brand safety problem. Hallucination Rate and Factual Inaccuracy Rate are listed as distinct metrics, and the IAB flags this as a dimension unique to AI measurement. There is no equivalent in a rankings report: a top-ranked page cannot get your pricing wrong on your behalf.

Persuasion is the part almost nobody measures. Recommendation Strength is the difference between being mentioned in a hedged list and being the answer's actual pick, and post-citation click-through is the downstream number most engines do not hand you. The IAB says this section bridges to a forthcoming attribution framework, which is a polite way of saying it is not solved.

Publishers get a parallel set. Alongside Citation Rate, ppc.land reports the document defines Citation Decay Rate, Content Utilization Rate, and Attribution Clarity, which are the right questions for anyone whose content is being consumed rather than visited.

This is the part that will change conversations. The IAB defines two tiers, in its own words:

Six dimensions, and a number has to clear all of them. Most AI visibility reporting in circulation today clears two or three.

On volume, AdExchanger reports the guidance puts a floor under the bottom tier: anything under 50 queries is "exploratory", and fewer than 50 prompts "cannot meaningfully characterize a category". Worth pairing that with the obvious caveat, which the guidance also makes in its emphasis on varied prompts: 50 prompts that match real buyer questions are worth more than 500 generic ones. Volume without relevance just buys a stable measurement of the wrong thing.

Five of the six criteria are answered by doing more work. Reproducibility is different, because it is a property of the surface you are measuring, not of your effort.

We ran the experiment. Across 28 prompts tracked daily on ChatGPT and Google AI Mode for 42 days, the set of domains cited for a given prompt turned over by roughly 60–70% from one day to the next. Underneath that churn there was a stable core — on Google AI Mode a small set of domains earned about 56% of citations every day — but the surface layer moves constantly, and how much depends on the question. Broad recommendation prompts churned hardest; narrow branded comparisons were the most stable.

Two consequences for anyone applying the IAB's tiers.

A single run is not directional, it is anecdotal. If day-to-day turnover is 60%, a week-over-week change of 15% in your citation count is well inside the noise floor. In that study we required around 20 day-to-day comparisons before trusting a volatility figure, and the same logic applies to mention and citation rates.

And reproducibility has to be measured, not asserted. The cheapest test is to run your prompt set twice and compare the results — a check any team can run, and one worth asking a vendor to run in front of you. If the two runs disagree by more than the movement you plan to report on, you are reporting noise with a decimal point.

This is also why methodology disclosure is doing real work in the guidance. Providers are asked to disclose platform coverage, query construction, data collection method, and how they detect factual inaccuracy. When methods differ, results diverge wildly: published estimates of the overlap between top-10 rankings and AI citations range from 12% to 76%, mostly because different parsers count different things. Two vendors can both be honest and still hand you numbers that disagree by a factor of six.

Its blind spots are worth naming, partly because the document names some of them itself.

It scores single responses. Real buying research is a conversation, and a brand confidently named in turn one can vanish in turn three once the user adds a budget, a compliance requirement, or an integration they need — a limitation AIVO Journal argues sits at the centre of the framework. Because the metrics score isolated answers, that disappearance can still register as presence. It is the same gap we covered in multi-turn AI search: the last prompt in a conversation carries only about a third of what the user actually said, so prompt sets built from single questions measure the best case.

The denominator is unresolved. When a competitor does not appear, was it pushed out or was the query irrelevant to it? The guidance acknowledges it cannot yet tell those apart, and files the problem as not ready for standardization. Share-of-voice figures inherit that ambiguity, which is a good reason to read them as comparative trends rather than absolute market position.

Observed and sampled data are still different things. Prompt-based measurement is sampling, however rigorous. First-party citation logs like Microsoft Clarity's Citation dashboard — Microsoft contributed to the guidance — report what one platform's retrieval actually did, with no sample and no prompt set to argue about, and cannot tell you what the answer said or who was named beside you. Decision-grade practice uses both.

The guidance is most immediately useful as a set of questions rather than a target.

If you already measure, audit one number against the six criteria. Take the figure you would put in front of a CFO, and write down its distinct prompt count, runs per segment, prompt mix, cadence, reproducibility test, and engine coverage. Most teams find the gap in cadence or reproducibility, and both are fixable without new tooling.

If you are buying, make the disclosure list your RFP. Which engines, how the queries were constructed, how the data is collected, and how factual inaccuracy is detected. A vendor who answers all four clearly is differentiating on rigor, which is what the IAB says it wants the market to do. Our comparison of the best AI visibility tools and the AI visibility software hub are useful starting points, and the method behind the metrics is in our guides to prompt tracking and tracking your brand in AI search.

If you are not measuring at all, you are in the 84%, and the 4 P's are a reasonable order of operations: get presence right, then prominence, then portrayal, then worry about persuasion.

One structural note. Reproducibility is much easier to demonstrate when the method is inspectable, which is the case for auditing an open method against a black box. Elmo is open source and self-hosted, so the prompt set, the engines queried, and the way every mention and citation is counted are all visible in code and re-runnable on demand — the disclosure the guidance asks for, by construction.

The document will not settle what a good number looks like; the landscape is moving too fast for that, which is why the IAB declined to call it a standard. What it does is make the vagueness visible. A "visibility score" with no stated prompt count, cadence, or engine list is now conspicuously missing something, and that is a real improvement on where the category was last week.