Share of Voice in AI Engines: A Practical Measurement Model

Share of voice used to mean a simple count: how often your domain appeared in the top results for the queries that mattered. AI engines broke that formula, because there is no list of ten links to count. The answer is a generated paragraph, and your share of voice is a question of whether the assistant names you, cites you, or recommends you, and how often. Measuring that requires a model, not a ranking report.

Why Share of Voice Has to Be Redefined

In classic search, share of voice is a proportion. You count the queries where your page appears in the top positions, divide by the total queries in your set, and compare the result against competitors. The denominator is stable, the positions are discrete, and the whole thing can be exported from a ranking tool.

AI engines break every part of that. There are no fixed positions to count; an answer may mention three brands, one brand, or none. The same query can produce a different answer on different days, for different users, or in different phrasings. And a brand can be mentioned without a link, recommended without any mention of a competitor, or cited as a source without being named in the visible text. None of those outcomes maps cleanly onto the old ranking arithmetic.

The practical answer is to stop trying to force AI answers into a search-shaped metric and instead define share of voice as something new: the proportion of your tracked questions in which your brand is meaningfully present in the answer, weighted by the strength of that presence and compared against the brands you compete with.

The three levels of presence

Before you can count anything, you need to distinguish three levels of presence, because they are worth very different amounts.

A mention is the weakest level: your brand name appears somewhere in the answer, but the assistant is not relying on you as a source. A citation is stronger: the assistant names you and links to your content as the basis for part of its answer. A recommendation is the strongest: the assistant suggests you as a specific choice or authority for the reader's need. Most measurement models collapse these into one count and lose the signal that matters.

Define the Question Set Before Counting Anything

Share of voice is only meaningful relative to a fixed question set. If you change the questions every month, you are measuring noise. Build the set once, deliberately, and hold it stable.

A good question set has three qualities. It is relevant, meaning the questions are ones your actual buyers ask in natural language rather than keyword strings. It is bounded, meaning you can afford to run every question against every engine you track on a regular cadence. And it is shared, meaning the same set is used for you and for every competitor you benchmark against, so the comparison is apples to apples.

Start with somewhere between a few dozen and a few hundred questions. Too few, and a single flaky answer distorts the trend; too many, and the manual review that makes the data trustworthy becomes unsustainable.

Who is in your competitive set

Decide up front which competitors you are measuring against, and include yourself in the same list. Share of voice only has meaning as a relative number, so the competitive set defines the denominator. A workable set is small: the three to five brands a buyer would realistically weigh against you, not every company that ranks for the topic.

A Practical Measurement Model

Here is a repeatable model that produces a share-of-voice number you can defend and trend over time. It has six steps.

Fix the question set and the competitive set, and write both down so the measurement is reproducible from month to month.
Run each question against each engine you track, using a consistent method, and capture the full answer text verbatim.
For each answer, record which brands from your competitive set appear, and at what level: mention, citation, or recommendation.
Weight the levels. A simple scheme is one point for a mention, three for a citation, and five for a recommendation, so that a recommendation counts for more than a passing name-check.
Sum the weighted points per brand across all questions and all engines, then divide each brand's total by the sum of all brands' totals to get a share-of-voice percentage.
Record the raw counts alongside the percentage, because the percentage alone hides whether the whole market moved or one brand did.

This model gives you a single headline number plus the underlying detail you need when someone asks why the number changed.

Keep the raw answers, not just the scores

Store the verbatim answer text next to every score. When a brand's share jumps or collapses, the first question everyone asks is "show me an example," and the answer text is the evidence. Scores without the underlying text cannot be audited, and an un-auditable metric is a metric that will be quietly ignored.

Weighting Without Over-Engineering

The weighting scheme matters less than consistency. Whether you use one, three, and five points or a different set of numbers, the value comes from applying the same weights every month so the trend is comparable.

Resist the temptation to build a more elaborate scoring model before the basic one is running. A single well-maintained metric you trust is worth more than a sophisticated score that nobody can explain. If you later need to distinguish, for example, a recommendation from a qualified recommendation, add that as a separate column rather than silently reweighting the whole model.

Handle "no answer" honestly

Some questions will return no meaningful answer, or an answer that refuses to give a recommendation. Do not drop those questions from the set, because that would quietly change the denominator. Record them as "no presence" for every brand and keep them in the total. A question where no brand appears is a real signal about the market, not an error to be discarded.

Turn the Score into a Trend

A single share-of-voice number is a snapshot; the value is in the line. Track the percentage, the raw weighted points, and the level mix (mentions versus citations versus recommendations) every month, and watch how they move together.

Three movements are worth reading. If your share rises on the back of citations and recommendations rather than mentions, the gain is durable. If a competitor's share falls while yours holds, look for what they stopped doing before celebrating. And if the total number of brands present per answer is shrinking, the field is consolidating, which changes the meaning of every share number at once.

Use the metric to direct content, not just report it

The whole point of measuring share of voice is to change it. When the model shows a competitor being recommended for questions where you are only mentioned or absent, that is a specific, actionable gap. Write or rewrite content that directly answers those questions with the clarity and original detail the assistant appears to reward, then watch whether your level of presence moves from mention toward recommendation in the next cycle.

Conclusion

Share of voice in AI engines cannot be read off a ranking report, because the answer is generated, not ranked. It needs a model: a fixed question set, a fixed competitive set, a three-level distinction between mention, citation, and recommendation, and a consistent weighting applied month after month. The result is a number that is smaller and more manual than search reporting, but far more useful, because it measures what an assistant actually says about you rather than where a page happens to rank. Build it once, keep it stable, and let the trend tell you where to write next.

Frequently Asked Questions

What is the difference between a mention and a recommendation in share of voice? A mention is the weakest signal: your brand name appears in the answer, but the assistant is not relying on you as a source or suggesting you. A recommendation is the strongest: the assistant points to you as a specific choice or authority for the reader's need. A citation sits between them, naming you and linking to your content as the basis for part of the answer. Weight them separately, because a recommendation is worth far more than a name-check.

How many questions should I track for AI share of voice? Somewhere between a few dozen and a few hundred, held stable over time. Too few questions and a single flaky answer distorts the trend; too many and the manual review that keeps the data trustworthy becomes unsustainable. The questions must be relevant, bounded, and shared with your competitive set, so the same set is run for you and for every competitor you benchmark against.

How often should I re-measure share of voice across AI engines? Monthly is the most practical cadence, using the same question set and weighting each time so the trend is comparable. AI answers change between runs, so a single cycle is noise; the value is in the line across several months. If a number jumps sharply between cycles, inspect the stored answer text before concluding anything, since assistant behavior can shift for reasons unrelated to your content.