← Research library

Measuring share of answer: tracking brand presence in AI responses

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Search visibility used to be measured with rankings: where your page appeared for a given keyword. AI-generated answers have no stable rankings, and the same question can produce different answers depending on wording, the user, the time, and the system.

Measuring presence in those answers requires a different approach. We use a measure we call share of answer: across a defined set of questions and repeated samples, how often your brand is mentioned or cited, and how it is described.

Why a single check misleads

Generated answers are sampled, not looked up. Ask the same question twice and you may get two different lists of brands, in a different order, with different sources cited. Ask it next month and the system itself may have changed.

That means one screenshot proves little. A brand that appears in today's answer may be absent tomorrow, and the reverse. A useful measure has to treat each answer as one draw from a range of possible answers and estimate how often a given outcome occurs.

We think of share of answer as closer to a survey than to a ranking report. You choose the questions carefully, sample them repeatedly, and report rates with a sense of how much they could move by chance.

Building the question set

Start with the questions that matter to your buyers, not just high-volume keywords. Good sources include sales calls, support tickets, and community discussions. We describe how to collect and group them in From Keywords to Questions.

Organize the questions by type, so you can see where you are strong and where you are absent. We use four types.

  • Category questions: "What tools help with marketing reporting?"
  • Problem questions: "How do I reduce time spent on weekly reports?"
  • Comparison questions: "How does A compare with B?"
  • Brand questions: "What is [your brand] good for?"

Thirty to 50 questions is a practical starting size. Fewer than that and one odd answer can swing your results. Many more and the review work becomes hard to sustain.

Freeze the wording once you start. Changing a question mid-series breaks the trend line, so add new questions as new entries and retire old ones deliberately.

Sampling

Because answers vary, ask each question several times, in each system you care about, and at regular intervals. We suggest three samples per question per system each month as a starting point. Use a fresh session for each sample where you can, so earlier answers do not shape later ones.

For each answer, record the same fields in the same format. Consistency in recording matters as much as the sampling itself.

  • Mentioned: whether your brand was named.
  • Cited: whether your content was linked or credited as a source.
  • Competitors: which other brands were named.
  • Description: how your brand was described, in the answer's own words.
  • Accuracy: whether that description was correct and current.

Keep the full answer text as well. When a number moves, you will want to read what actually changed.

Calculating share of answer

For each question type and each system, calculate four rates. Report them separately rather than blending them into one score, because each points to different work.

  • Mention rate: the share of responses that name your brand.
  • Citation rate: the share of responses that cite your content.
  • Competitive share: your mentions divided by all brand mentions in the set.
  • Accuracy rate: the share of your mentions that describe you correctly.

Consider a hypothetical team with 40 questions, sampled three times each in two systems, for 240 answers a month. If its brand appears in 72 of them, the mention rate is 30%. If 60 of those 72 mentions are accurate, the accuracy rate is about 83%.

Now consider how much that 30% could move by chance. With 240 answers, a shift of a few points from one month to the next is within ordinary variation. A sustained move across several months, or a large jump in one question type, is more likely to mean something.

The true uncertainty is a little wider than simple arithmetic suggests. Three samples of the same question are not fully independent, since they share wording and often share sources. If you want a firmer read, add more distinct questions before you add more repeats of the same ones.

Interpreting results carefully

Share of answer is a noisy measure, and it is easy to read too much into it. Three cautions help.

Variation is normal. Small month-to-month changes may be noise. Look at a rolling three-month view before drawing conclusions.

Systems change. Model and product updates can shift results suddenly, independent of anything you did. If every brand in your set moves at once, suspect the system before your content.

Accuracy matters as much as presence. Being mentioned with outdated or wrong information is a problem to fix. The fix usually runs through clearer content on your own site and consistent third-party information, as we discuss in Citations, Not Rankings.

Where agents help

Running dozens of questions across several systems every month is tedious, and tedium leads to skipped months. An agent can run the sampling on a schedule, record results in a fixed format, and flag notable changes and inaccurate descriptions for review.

Judging accuracy is the step that needs the most care. An agent can compare each description against an approved fact sheet and flag mismatches, but a person should confirm each flag before anyone acts on it. Treat the fact sheet as a test set, and update it when products and pricing change, or it will drift the way any test does. We describe that problem in Evaluation Drift.

People decide what the results mean and what to do about them. The agent's job is to make the sample reliable and the changes easy to see.

Reporting it

Share of answer works best as one line in a broader report, next to traffic, branded search, and pipeline. We suggest placing it with the other measures you review on a regular rhythm, as described in Leading Indicators for Agent Programs.

The question we are studying next is how to weight questions by value. A mention in a decision question likely matters more than one in a broad category question, and we are testing simple weighting schemes that teams can apply without a statistician.

How to apply this: build a set of 30 buyer questions, sample each three times a month in your most important AI systems, and record mention, citation, competitor, and accuracy data. Review a rolling three-month trend, not individual results.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.