Measurement · 27 August 2026

By Eddie Eastwood, Founder

AI Visibility Analytics: What You Can Actually Measure (and What You Can't)

What is AI visibility analytics, and why doesn't it work like GA4?

AI visibility analytics is the measurement layer on top of AI visibility itself: whether, how often, and in what form AI platforms mention a business, measured through the closest available proxies rather than a direct feed from the platforms themselves. It gets treated as a GA4 problem because the vocabulary is borrowed wholesale: "channel," "conversion," "session," "attribution." GA4 was built around a visitor's session; Search Console around a URL with a checkable position in a list. An AI answer offers neither of those anchors. It is a single generated response to a single prompt, produced by a system that can legitimately return something different the next time you ask, built from a source mix that shifts as the underlying web changes. Importing GA4's assumptions onto that object breaks the reporting before a single number gets pulled.

What you can actually measure today

Five things are repeatably measurable right now, each with real limits worth stating alongside the number rather than after it.

Mention share from a fixed, repeated prompt set. Build a list of real buyer questions, run it across the platforms your customers actually use, and record whether you're named. How to Check If AI Recommends Your Business covers building that prompt list and running the test; this page picks up once you have the raw counts. The number is real. Its meaning is bounded by the prompt list and platform set you chose, which is why How to Measure Your Brand's Share of Voice in AI Answers treats the resulting percentage as a statement about your method, not a market-wide fact.

The mix of sources an engine cites for your category. Which domains and page types a platform pulls from when it answers a question in your space is directly observable by running the test and reading the citations. It tells you where to build presence, not how often you're winning against a named competitor.

Whether you're present when you test a real question. This is the mention-share test's simplest form: yes or no, for this prompt, on this platform, today. Useful as a spot check, unreliable as a trend on its own, for reasons the next section covers in full.

Your site's technical readiness. Whether your pages render without JavaScript, carry structured data, avoid an accidental crawler block, and answer real questions directly is fully measurable with a technical audit, and it's a precondition for being cited at all. QBiz's own AI Visibility Gap US study audited 191 US local service business websites against ten binary readiness signals. Of the 179 sites that returned a successful fetch, 46.9% scored Ready, 45.3% Partially Ready, and 7.8% Not Ready.^[1] That tells you how prepared a page is to be read and cited. A readiness score and a citation rate are two different layers, and a business can be fully Ready on one and still invisible on the other.

The visible, labelled slice of AI-referred traffic. GA4's AI Assistants channel and Search Console's generative-AI performance reports both exist and both work, within limits set out in full in How to Measure AI-Search ROI. The short version: what you see through either tool is a floor, not a count, because AI-driven visits that arrive with no referrer attached get filed as ordinary Direct traffic.

What no tool can measure yet

Each item below follows from how the engines are built, so better tooling will not close it.

A confirmed line from a particular AI answer or mention to a particular sale. An AI-referred visit can sometimes be tied to a sale when it carries a usable referrer or tagged link. What no tool can currently prove reliably is that a particular AI answer or mention caused that sale: platforms do not provide citation-level exposure logs, and many AI visits lose their referrer and appear as Direct. What you can build, using the correlation method in Measuring AI-Search ROI, is a correlation: mentions rising alongside a rise in engaged, deep-page Direct traffic. That is evidence, not attribution, and the difference matters when you're reporting the number upward.

The actual wording and framing an engine used for you, across its full range of possible answers. A single screenshot of one AI answer describes one output, generated at one moment, from one version of the model reading one snapshot of the web. It is not a stable description of how you're characterised. What Does AI Visibility Mean for a Brand? covers the narrative and framing question in full; the analytics point here is narrower: there is currently no tool that captures the full distribution of ways an engine might describe you, only individual samples from it.

A model's actual internal reasoning for a citation decision. No platform exposes why it selected one source over another for a given answer. Vendors describe general tendencies (which source types tend to get cited, which content structures tend to help), but the specific reasoning behind any one generated answer is not observable by the business it named, or by any third-party tool measuring it from outside.

A single visibility score that means the same thing everywhere. Tools present these numbers with enough confidence that a score reads like a fact about the brand rather than the output of a method. Every tool that reports an "AI visibility score" builds it from its own choices: which prompts, which platforms, how often it samples, how it weights frequency against position against sentiment. None of that is standardised across vendors, which is why the same brand can score very differently on two tools in the same week without anything about the brand's actual visibility changing. With no shared measurement standard for a category this new, two defensible methods will land on two different numbers.

Reliable consistency from a single observation. This deserves its own weight because it undermines almost every other number if ignored. GEOforge, a generative-search tooling vendor, published a study that ran the same buyer-intent prompt through ChatGPT roughly 49 times each, across 569 real prompts and 21 brands, for 65,478 total answers analysed.^[2] Across that set, 45% of prompts never named the brand in any run, 12% named it in every run, and the remaining 43% named it in some runs and not others.^[2] Put differently: for most prompts that produced a mention at all, whether you saw it depended on which run you happened to catch, not on a stable fact about your visibility. A single check, however carefully run, samples one point from a variable distribution. It is closer to one roll than to a measurement.

Underneath that instability sits a mechanical reason, not a flaw someone will patch. The underlying models are probabilistic: each answer is sampled, not retrieved from a fixed table, so built-in run-to-run variation in which entities get named is by design rather than by accident, a point covered in more depth in Your Business Showed Up in ChatGPT Last Week and Not This Week. Commercial visibility platforms work around this by drawing on large real-user prompt panels instead of single test runs. Profound states that its data comes from double opt-in consumer panels supplying tens of millions of real prompts a month, with no synthetic or simulated data used, and with probabilistic modelling applied to correct for demographic and geographic bias in the panel.^[3] That is an improvement on a single test run. It is still a modelled estimate built from a sample, not a census of what every AI platform said to every user, and reading it as anything more precise than that overstates what even a well-built panel can tell you.

Why this doesn't behave like GA4

Five structural differences explain why importing a web-analytics mindset onto AI visibility produces numbers that look precise and aren't.

No session model. GA4 tracks a session: a visitor arrives, does things, leaves, and the whole path is one connected record. An AI citation is a binary outcome, named or not named, inside one generated answer to one prompt. There is no equivalent unit to a "session duration" or a "pages per session," because there is no multi-step visit to measure inside the answer itself; the closest thing is what happens after someone clicks through, which is ordinary GA4 territory again, with its own referrer problem layered on top.

No reliable referrer. An outbound click from an AI assistant often carries no referrer, so the visit arrives unlabelled and the analytics record cannot say what sent it. Measuring AI-Search ROI covers what that does to attribution. GA4's dedicated AI Assistants channel and Search Console's generative-AI reports both narrow the gap; neither closes it.

No cookie-based journey. GA4's funnel and attribution models assume you can follow one identified visitor across a sequence of touchpoints. An AI platform doesn't expose a comparable journey for the businesses it mentions: there is no dashboard showing that a given user saw your business named on Monday, researched a competitor on Wednesday, and converted on Friday. What a business can observe is its own side of the traffic, once someone clicks through, and even that arrives with the referrer problem above.

Sampling and non-determinism. A Google ranking is close to deterministic: the same query returns close to the same list, checkably, day to day. An AI answer is sampled from a probability distribution and rebuilt from a live, changing set of sources each time, which is why the same prompt can return a different answer on back-to-back requests, not only week to week. The mechanics and the evidence for how much this moves are covered in Your Business Showed Up in ChatGPT Last Week and Not This Week; the analytics consequence here is simpler to state: a metric built on one observation is measuring noise as much as signal.

Version drift. A web page's ranking factors change slowly and are broadly documented. The model producing an AI answer can change its behaviour with a silent update the vendor doesn't publicise in a form that reaches individual businesses. A measurement taken in one month is a description of that month's model, reading that month's web, and there is no guarantee the same test run six months later is measuring a stable version of the same thing, in the way that "position 4 on Google" was a stable statement for as long as ranking factors changed slowly.

How do you build a measurement routine around these limits?

None of the above is a reason to give up on measuring this. It's a reason to build a routine that doesn't claim more precision than the underlying system can support.

Run every prompt more than once, and record the spread, not a single result. Given that a single test can land anywhere inside a variable set of answers, the useful unit isn't "did we appear" but "in how many of N runs did we appear." A prompt that names you in nine of ten runs is a different, more durable finding than one that named you once and never again on a retest.

Fix the prompt list, the platform set and the cadence, and don't change them quietly. Every one of the measurable items above only means something relative to the method that produced it. Changing your competitor set, adding a platform, or rewording a prompt resets the baseline, whether or not anyone intended it to. Note every change alongside the new number rather than letting a moved baseline read as a moved market.

Separate readiness from citation, and report them as two different layers. A technical audit tells you whether a page can be read and cited. A prompt test tells you whether it has been. Treating a high readiness score as proof of strong citation performance, or a low citation result as proof of a technical problem, conflates two measurements with different failure modes.

Treat every "visibility score" as a method, and ask what the method is before comparing it to anything. If a tool, a report, or a competitor claims a number, the useful question is not "is it accurate" but "what prompts, what platforms, what sampling, over what window." Two numbers with the same label are only comparable if the method behind them matches.

Read GA4 and Search Console as a floor, and pair them with a deliberate deep-page-versus-homepage check on Direct traffic, the practical proxy method set out in full in Measuring AI-Search ROI. It won't turn an unlabelled visit into a labelled one. It gives you a defensible trend line instead of a single misleading total.

Report a range, not a point estimate, whenever the underlying test ran on a small sample. A number presented as "32%" invites a false sense of precision that a variable system cannot support from a small sample. "Named in roughly a third of repeated tests this month" carries the uncertainty the number actually has.

FAQ

What can you actually measure in AI visibility today?

Mention share and citation frequency from a fixed set of prompts run repeatedly, which domains an engine draws on in your topic, whether your business is present when you test a relevant question, your site's technical readiness signals, and the visible slice of AI-referred traffic your analytics can label. Each of these is a real, checkable number. None of them is a complete picture on its own.

What can no AI visibility tool measure yet?

A confirmed link between a particular AI answer or mention and a particular sale, the complete set of ways an engine might describe you rather than the single sample you captured, a model's internal reasoning for why it chose one business over another, and a stable, tool-independent visibility score that would read the same on every platform. These are not measurement gaps waiting on better tooling; they follow from how the engines are built.

Why doesn't AI visibility analytics work like GA4 or Google Search Console?

GA4 was built around a session someone can be tracked through, and Search Console around a URL with a stable, checkable ranking. An AI answer has neither: no session model, an often-missing referrer, no cookie-based path through a funnel, and an answer that can change between two identical requests because the underlying model is probabilistic and the web it reads keeps changing.

Is a single AI visibility check reliable?

No. A published GEOforge study that ran the same prompt roughly 49 times per question found the named brands changed in 43% of cases, with only 12% of prompts producing a consistently reliable mention. One check tells you what happened on one roll; it does not tell you your visibility.

Why do two AI visibility tools show different numbers for the same brand?

Because each tool builds its own prompt set, samples a different mix of platforms, and applies its own scoring logic on top, and none of that method is standardised across vendors. Two tools measuring the same brand in the same week are frequently answering two different questions that happen to share a label.

What does QBiz's own research measure, and what does it not measure?

QBiz's AI Visibility Gap US study audited whether a website is technically ready to be read and cited: rendering, crawlability, structure and answer-oriented content. It does not measure, and states plainly that it does not measure, whether any AI platform actually cites, ranks or recommends the businesses it audited. Readiness and citation are separate measurements, and neither predicts the other.

Technical readiness is the one layer here you can check on your own site today.

Check whether AI platforms can find and cite your business →