How Do You Measure Whether AI Understands Your Business?

Being found is no longer enough. AI has to understand you. But I’ve come to think “understanding” isn’t a single measurement, and treating AI visibility as one score is where most attempts to measure it go wrong.

A business can be recognized by an AI system, cited as a source, and still never get recommended. It can be described accurately in one context and confidently wrong in another. Being mentioned, being cited, being understood, and being recommended are not the same thing. Measurement exists to determine which of those problems you actually have.

Five Questions, Not One Score

• AI visibility isn’t a single ranking. It’s at least five separate, observable dimensions, and a business can perform strongly on one while failing another.
• Recognition, understanding, inclusion, citation, and recommendation influence each other, but none of them guarantees the others.
• A business can be accurately recognized and still never get recommended, if nothing about it stands out from the other businesses an AI system also recognizes.
• Most attempts to “check AI visibility” only test one of these five dimensions, usually recognition, and mistake a pass on that test for the whole picture.
• Measuring this properly means testing each dimension on purpose, not inferring all five from a single prompt.

Why “Are We Visible to AI?” Is the Wrong Question

When a client asks me this, what they usually mean is something narrower: does an AI system know we exist. That’s answerable, and it’s usually the easiest of the five dimensions to satisfy, which is exactly why I think it’s the least useful one to stop at.

A business can exist clearly in a system’s model of the world, correctly named, correctly described, correctly located, and still lose every meaningful opportunity that runs through that system. Existing doesn’t guarantee being considered. Being considered doesn’t guarantee being cited. Being cited doesn’t guarantee being recommended.

I want to be careful about what that progression means, though. These five dimensions influence one another, but they aren’t a funnel, and I don’t want to overstate the relationship. An AI system can cite a business without demonstrating much real understanding of it. It can recommend a business without ever surfacing a visible citation. The relationship runs in more directions than a simple sequence would suggest. What doesn’t change is this: strength in one dimension is never proof of strength in another.

Treating AI visibility as a single yes-or-no, or a single number from one tool, collapses five different diagnostic dimensions into one. In practice, a business that fails can spend months fixing the wrong one.

What We Can Actually Observe

Before going further, I want to be honest about a limitation. When I talk about measuring whether AI understands a business, I’m not suggesting we can look inside a model and inspect what it “knows.” We can’t.

What we can observe is behavior: how consistently a system identifies an entity, what attributes it associates with it, whether it includes that entity in relevant answers, what sources it cites, and when it chooses to recommend it. Repeated systematically, those observations give us a practical picture of how a business is being understood, not a direct readout of it.

The Five Diagnostic Questions

Each of the following is a separate, observable question. None of them can be reliably inferred from the others.

Recognition

Does the system identify the business, the person, or the entity correctly? This is where I always start, not because it’s a hard gate everything else depends on, but because poor recognition usually makes everything downstream harder to diagnose.

Understanding

Recognition confirms a system knows a business exists. Understanding is what it believes about that business once it does: what expertise, geography, services, and relationships it associates with the name. In practice, a business can be correctly recognized and badly understood, accurate by name but wrong by substance, which can be more commercially dangerous than not being recognized at all. Incorrect understanding can actively send a prospective buyer in the wrong direction. Absence just leaves them looking elsewhere.

Inclusion

Recognition and understanding are both about a business in isolation. Inclusion is about whether it shows up when it matters: does the entity enter the consideration set when someone asks a relevant question, not a question about the business by name, but about the problem the business solves. A business invisible to a branded question likely has a recognition problem. A business invisible to a buyer’s actual question has a commercial one.

Citation

What sources does the system surface in support of its answer about the business? Citation gives us a visible part of the evidence layer: which sources a system chooses to associate with an answer, and whether those sources support, weaken, or contradict the business’s own claims. A business with strong self-description and weak citation is standing on an assertion, not a foundation.

One engagement illustrated this particularly clearly. A client’s page remained uncited for weeks despite being live, relevant, and editorially strong. After we rewrote the opening paragraph so it answered the target question directly, its citation behavior changed materially. Within ten weeks, it had become the fourth most-cited page on the site. We can’t prove from the outside exactly why the systems changed their behavior. But the result reinforced something we’d already begun observing: content can contain the right information and still present it poorly for a system trying to identify a useful answer.

Recommendation

Under what circumstances does the system actually put the business forward, not just mention it, but recommend it. Recommendation is inherently competitive because being understood is not enough. What matters is whether the available evidence and relevance are sufficient for one business to be put forward in a particular context rather than the alternatives the system also recognizes.

Dimension	Diagnostic Question	Observable Signal
Recognition	Does AI identify the entity correctly?	Correct name, role, location, company/entity relationships
Understanding	What does AI associate with the entity?	Expertise, services, markets, attributes, credentials
Inclusion	Does the entity enter relevant consideration sets?	Appearance across non-branded buyer/problem questions
Citation	What evidence shapes the answer?	Sources cited, source concentration, conflicting evidence
Recommendation	When does AI put the entity forward?	Recommendation frequency and context relative to alternatives

Why These Five Don’t Move Together

I saw this pattern clearly in a competitive dataset for a luxury real estate market. One of the most credentialed brokers in the field, real national media citations, a real professional history, was still underperforming on citation share relative to what those credentials should have earned.

The most obvious weakness wasn’t the credentials. It was the evidence architecture around them: the broker had two separate bio pages live at once, each manually updated with slightly different “years of experience” language that had drifted out of sync with each other. We can’t see inside the systems well enough to say that the duplicate pages alone caused the weaker citation performance. What we could see was an avoidable contradiction in the evidence those systems had available to work with, and recognition and understanding were both fine the whole time. An AI system asked about that broker directly could describe them accurately. What broke down was citation: nothing pointed cleanly enough at one place to concentrate authority there.

The Same Score Can Hide Completely Different Problems

Picture two businesses that a simplistic AI visibility tool might score identically.

Business A has strong recognition, strong understanding, and strong citation, but weak inclusion and weak recommendation. Business B has weak recognition, inconsistent understanding, moderate inclusion, and scattered citation, but picks up an occasional recommendation anyway.

A single aggregate score could easily land both of them in the same range. But they don’t have the same problem, and they don’t need the same fix. Business A is understood correctly and cited credibly, and still isn’t being recommended in the contexts that matter, which points to a competitive problem. Business B has a shakier foundation across nearly every dimension, and its occasional recommendation is an isolated outcome, not yet a pattern worth trusting.

This is the real argument against single-score tools. Not that measurement is pointless, but that aggregation can obscure diagnosis exactly when the diagnosis is the part that matters.

What Measuring This Actually Looks Like

The purpose of measurement isn’t to produce a grade. It’s to identify the constraint.

• If recognition is weak, publishing more content may be premature.
• If understanding is wrong, more visibility can amplify the wrong interpretation.
• If inclusion is weak, branded success tells you very little.
• If citation is weak, the evidence architecture needs attention.
• If recommendation is weak while the others are strong, you’re dealing with a genuinely competitive problem, not an identity one.

None of these five dimensions are visible from a single prompt, and none of them hold still. A system’s answer to the same question can shift from one day to the next, which is exactly why I don’t put much stock in a single test.

One engagement tracked a luxury brokerage’s competitive share of AI citations weekly against its two closest competitors for about fourteen weeks, across roughly fifty representative buyer and seller questions relevant to that market. It began at roughly 30% of the three-way citation share. Over the tracking period it reached approximately 40% at its peak and subsequently held around 38-39%, even as overall citation activity across the tracked questions cooled with the season. A single snapshot taken at the wrong week would have told a different story.

In practice, a real measurement effort looks less like a score and more like a standing set of questions, asked consistently, across the range of ways a real buyer would actually ask them, some naming the business directly, most not, and tracked over time rather than checked once. The question set matters as much as the tracking cadence. If the prompts don’t represent the decisions and problems real customers bring to AI, the measurement can be precise and still tell you very little.

How This Connects to the Rest of the Series

This framework isn’t separate from what I’ve argued elsewhere in this series. It’s the diagnostic version of it.

Recognition and understanding are what the entity problem is actually about: a business that isn’t clearly, consistently identified struggles on the first two of these dimensions, regardless of how good its content is. Citation is what a website’s validation role actually produces, or fails to, when a visitor or a system goes looking for something to check a claim against.

Inclusion and recommendation are where the commercial stakes actually land. A business that’s recognized, understood, and even cited can still lose every opportunity that depends on those two dimensions, without any way to see it happening from a standard analytics dashboard.

Several later pieces extend this in different directions: the commercial risk of a company never seeing its own inclusion gap, the same problem applied to an individual professional inside a larger organization, what it might mean for when the competitive process for a deal actually begins, and what the Citation dimension specifically looks like over time, not just whether it’s present. A separate piece asks who inside an organization is actually responsible for keeping all of this coordinated, and another looks at the broader structural reason any of this matters: AI increasingly sits between a buyer and the businesses competing for their attention.

A business that can’t answer which of the five it’s failing is optimizing blind, however much content it publishes. The point isn’t to make every number go up. It’s to understand which constraint is preventing the business from being understood, considered, evidenced, or recommended, and work on that problem first. The businesses that get ahead of this aren’t necessarily the most visible today. They’re the ones that know exactly which question they haven’t answered yet.