What Is AI Visibility? A Measurement Framework for B2B Teams

The important phrase is relevant questions.

Searching your company name tells you whether an AI system can discuss you when explicitly prompted. It does not tell you whether you appear when a buyer asks:

• which providers solve a particular problem;
• which options fit a specific use case;
• how two approaches compare;
• what a buyer should consider before choosing;
• which companies have evidence for an important claim.

A useful AI visibility baseline therefore needs a defined prompt set, defined answer surfaces, clear counting rules and repeated observations.

This differs from conventional search visibility. Traditional search usually presents a ranked set of results for the user to inspect. Search-enabled AI products can instead retrieve information and synthesise it into a generated response. For example, ChatGPT Search can display inline citations, while Anthropic documents cited web-search responses and Google documents grounded responses with citations.

The practical consequence is that visibility can occur inside the answer itself. A company may be named, compared, recommended or associated with cited evidence before the buyer visits its website.

These concepts are related, but they should not be collapsed into one score.

A company can have high visibility and poor representation. It may appear frequently while being described with an outdated category, the wrong audience or an incomplete capability set.

It can also have low visibility but accurate representation when it does appear.

For Kojable, visibility monitoring is therefore one part of a wider AI answer alignment process:

Monitor → Diagnose → Improve → Verify

Monitor establishes what is happening. Diagnose identifies the meaningful representation gap and the evidence associated with it. Improve determines what should change, where and how. Verify retests comparable questions to see what changed.

Visibility is an input to that process, not the final outcome.

A useful measurement system should distinguish at least the following:

The definitions are not administrative details. They determine what a number means.

If one dashboard defines visibility as brand mention frequency and another combines mentions, sentiment, citation presence and answer position, their scores are not directly interchangeable.

The same applies internally. Every metric should have a numerator, denominator, analytical unit, eligible population, exclusions, time period and platform or surface definition where relevant.

Without those definitions, a score can look precise while answering an unclear question.

Our earlier research analysed a large retained set of AI-generated answers collected through repeated monitoring across major AI systems.

The first analysis grouped the observations into a state map based on whether:

• a citation was captured;
• the monitored company appeared in the answer;
• the monitored company appeared in captured citation material.

That first analysis produced an important practical finding: answer mentions and citation-layer representation are related, but they are not the same measurement state.

Further analysis made the interpretation more precise.

How our research changed as the analysis matured

The first public analysis used the complete retained monitoring snapshot and described the relationship between company mentions and the citation layer at a high level.

We later returned to the raw response data, citation records, analysis code and collection pipeline.

That audit changed the analysis in several ways.

We tightened the terminology.

Earlier language could make a company term appearing in citation material sound stronger than the data justified.

We now distinguish between:

• citation material being present;
• a company term appearing in that citation material;
• an actual company-owned domain being cited.

Those are different observations.

We separated non-comparable citation capture.

The retained dataset included a platform whose citation information was captured differently from the other comparable systems.

That difference mattered.

A single blended cross-platform percentage could obscure a measurement-method difference rather than reveal a genuine platform difference.

For citation-layer comparisons, we therefore moved to the population where the capture method was sufficiently comparable.

We changed the question, not the underlying historical observations.

Our earlier analysis asked a broad question across the full monitoring snapshot.

The refined analysis asks a narrower one:

Among answers that mentioned the monitored company within the comparable citation population, how often was the company also represented in the captured citation material?

The refined analysis showed that company mentions and citation-layer representation overlapped strongly in that bounded dataset.

But the important publication lesson is not the precise percentage.

It is that the result depends on the analytical population and the metric definition.

A figure can change substantially when the denominator changes, even when the underlying retained observations have not changed.

That is why an AI visibility percentage should never be published without explaining what population it describes.

Readers who want the evidence parent can review Kojable's underlying research and methodology. This article focuses on what that research changes about measurement practice.

The initial analysis compressed an extended monitoring period into overall totals.

The retained dataset also allowed us to examine the observations over time.

That analysis showed that aggregate visibility and citation measures can hide meaningful variation between periods.

In other words, the overall average was not the whole story.

The number of retained observations differed across periods, and citation participation did not remain constant throughout the monitoring window. A result that looked stable in aggregate could therefore look materially different when viewed as a time series.

That changed the practical conclusion again.

The useful question is not simply:

“What is our AI visibility score?”

It is:

“What did we observe across this prompt set, these answer surfaces, these runs and this period, and how stable was that observation?”

The evolution of the research therefore moved through three stages:

1. Separate mentions from citation-layer representation.
2. Separate comparable measurement populations from non-comparable capture methods.
3. Examine repeated observations through time instead of relying only on an aggregate average.

Each step narrowed what we were justified in claiming.

That is a feature of evidence-led measurement, not a weakness. Better analysis should sometimes make the conclusion more qualified.

Generative answers are variable. The same underlying question can produce different wording, source sets and company inclusion across repeated observations.

Recent measurement research reaches a similar conclusion.

The 2026 paper Quantifying Uncertainty in AI Visibility examines variability in AI-search visibility measurement and cautions against treating one-off observations as stable estimates.

Another 2026 study, Don't Measure Once: Measuring Visibility in AI Search, similarly argues for repeated measurement rather than assuming that a single generated answer represents a fixed result.

For teams building an operating baseline, that means the measurement design should survive repetition.

If an answer changes tomorrow, you need enough retained context to decide whether you are looking at:

• normal run-to-run variation;
• a recurring pattern;
• a platform-specific difference;
• a genuine change in the public information environment;
• or insufficient evidence to make the distinction.

A screenshot cannot answer those questions.

A useful B2B prompt panel normally spans several stages.

Category discovery:
What options solve this problem?

Use-case fit:
Which providers support this requirement for this type of company?

Comparison:
How do the main approaches or providers differ?

Validation:
Does this company support a particular capability or claim?

Recommendation:
Which option is appropriate under a defined set of constraints?

This matters because B2B buyers are already using generative AI during purchasing research. Gartner's recent buyer research found meaningful use of GenAI for vendor and product research, while also showing that many buyers still want human validation of AI-generated information.

The implication is not that AI has replaced the company website, salesperson, analyst or peer recommendation.

It is that AI has become another research surface that can affect the frame a buyer carries into those later interactions.

A branded prompt alone does not measure that.

Several situations that look similar in a dashboard can require very different responses.

Absence

The company does not appear in relevant buyer questions.

Investigate whether the issue is limited to a particular intent, platform or audience. Check whether the public evidence clearly associates the company with the question being asked.

Presence with weak or inaccurate representation

The company appears, but the description is stale, generic or inconsistent with verified positioning.

That is a representation problem, not merely a visibility problem.

Presence without the evidence you expected

The company appears, but relevant citations or proof are absent.

Inspect the source pattern, but do not assume the missing citation proves a specific page was ignored or rejected.

Competitor-led framing

The company appears, but the answer uses criteria, category language or proof that make another provider easier to understand.

The diagnosis should begin with what is observable: recurring framing, missing company evidence, outdated public information or a weakly defined use case.

From there, potential drivers remain hypotheses until stronger evidence supports them.

A recurring source may be associated with a repeated answer pattern. That does not prove the source caused the answer.

The appropriate action might be:

Correct an owned fact.
Update an outdated product, category, audience or capability description on the page that owns that fact.

Add missing proof.
If an important claim lacks defensible evidence, strengthen the case study, research, documentation or other source that should substantiate it.

Clarify an existing page.
Sometimes the correct evidence exists but is difficult to identify because the page is vague, internally inconsistent or mixes several decisions.

Address a realistic third-party gap.
Where an external directory, partner description or another editable source contains stale information, update it through the legitimate owner or publishing process.

Create new content only when the decision genuinely lacks a page.
A new comparison, use-case or evidence page can be justified when no existing asset adequately answers the buyer question.

Change the measurement design.
If the apparent problem changes when prompts, eligibility or capture rules are corrected, the first fix may be measurement rather than content.

This is why AI visibility is useful as an input rather than an endpoint.

The operating sequence is:

Monitor the answer → Diagnose the gap → Improve the justified evidence or information → Verify with comparable tests.

A practical verification design should preserve:

• the core prompt set;
• the platform or answer-surface definitions;
• geography and language where material;
• run count;
• metric definitions;
• numerator and denominator rules;
• known environment or methodology changes.

Then compare the new observation with both the original baseline and the previous checkpoint.

Other things may also have changed: answer variability, source availability, prompt interpretation, product configuration or the wider information environment.

Stronger causal language needs a design capable of distinguishing the intervention from those alternatives.

For most operating teams, the practical standard is:

Record what changed, identify whether the movement recurs, decide what the new evidence justifies, and retest again.

The stakes are higher when:

• the product or service is difficult to explain in one sentence;
• positioning has recently changed;
• category definitions are contested;
• buying involves comparison and validation;
• important capabilities require proof;
• competitors are easier to describe than you are;
• different AI systems give materially different representations.

It matters less as a vanity metric.

A high mention rate is not automatically a business result. A citation is not automatically endorsement. AI referral traffic is not the same metric as answer visibility. And a change in an AI answer does not prove a buyer changed their decision.

The useful objective is more disciplined:

Understand whether the company appears in the buyer questions that matter, assess whether the representation is accurate and evidenced, then build a repeatable process for improving and verifying the gaps that deserve action.

That is the role of AI visibility inside a wider AI answer alignment programme.