What share of answer measures, and what it doesn’t

Share of answer is the percentage of tracked buyer prompts where an AI engine names your brand. It helps answer a practical question: when someone asks about the problem you solve, are you part of the conversation?

It is also an easy number to make look better than it is. Change the questions, mix branded and unbranded prompts, or count a passing mention as a recommendation, and the headline can improve without the buyer’s experience changing very much.

We put the number near the top of the Friday report. We put the definition and prompt set close behind it. A metric is useful when another person can understand what produced it and what decision to make next.

Start with a clear counting rule

In our basic share-of-answer measure, a response counts when it names the brand. We record a recommendation, a citation and an incorrect fact separately. That keeps one number from trying to describe several different things.

Suppose you run 40 prompts against one engine and your brand appears in 12 responses. The share of answer for that run is 30%. It does not matter whether the brand appeared once or three times inside a response: that response still counts once.

Now imagine two of those mentions contain wrong pricing and three describe you as a poor fit. The share remains 30%, but the interpretation changes. This is why the supporting labels matter as much as the arithmetic.

Keep the prompt set comparable

We group prompts by the decision the buyer is making: discovering the category, building a shortlist, checking facts or describing a problem without naming a category. Branded questions are reported separately from questions that do not already mention the company.

A prompt such as “What does Northbeam cost?” tests whether an engine can explain a named product. “Which CRM works for a small sales team?” tests whether Northbeam enters the shortlist at all. Combining them without explanation makes the final percentage difficult to interpret.

Keep the wording, geography and testing conditions documented. Use fresh sessions where practical, and record the engine and test date. If the prompt set changes, retain a comparable group or mark a new baseline rather than presenting the two periods as identical.

Look at repeated observations

One answer is a sample, not a permanent position. We run the agreed prompts repeatedly and use a weekly view to reduce the influence of an unusual response. We retain the underlying answers so a change can be investigated rather than accepted on the strength of a chart.

Show the denominator alongside the percentage. A movement across a small prompt set deserves a different conversation from a similar movement across a large, stable set. If some runs fail, disclose the missing observations instead of treating them as clean negative answers.

Engine-level results should remain visible too. A blended average can hide a strong improvement on one engine and no meaningful change on another. The average is a summary; it should not replace the detail needed to plan the work.

Three things it can’t tell you

1. Whether the answer helped you. Being named beside an inaccurate limitation still counts as a mention. Review sentiment, factual accuracy and fit separately.

2. How many buyers saw it. Your tracked prompt set is a research sample. It is not a measurement of the total audience asking those questions.

3. How much revenue it created. A mention may influence a buyer without producing a direct click. Pair visibility with the referral and conversion data you can actually observe.

Avoid translating a visibility increase directly into pipeline. The relationship may be commercially useful, but it needs its own evidence. A report becomes more credible when it distinguishes what was observed from what is still an interpretation.

Turn the number into a work list

Read the answers behind the movement. If incorrect pricing is common, inspect the current page and the cited sources. If the facts are right but the brand is absent from category recommendations, the next task is different: review which alternatives and sources those answers include.

Our accompanying report shows whether the brand was named first, whether it received a link, and which wrong facts remained open. Those details explain why a flat headline number can still contain useful progress, or why a rising number can hide a problem.

The right weekly question is not just “Did share of answer go up?” It is “What changed in the answers, what did we ship, and what should we investigate next?” For a starting sample, use our 40 prompts to track first.