Share of Voice vs Share of Model vs Share of Retrieval

A marketer asks "what's our AI visibility?" and gets handed a percentage. The trouble is that three different tools will hand back three different percentages, all labeled something like "visibility," and none of them will agree. One counts how often you're named in answers. One counts your slice of the total brand mentions in a category. One counts whether you even reached the model in the first place. They are not interchangeable, and picking the wrong one to steer by sends you optimizing the wrong thing for months.

Here is what each of the three actually measures, how they nest inside one another, and which one to watch depending on the problem you're trying to solve.

Share of Voice: the metric you inherited

Share of Voice is the oldest of the three, borrowed straight from traditional marketing. In its classic form it measured your slice of paid advertising or press coverage in a category. Applied to AI search, it usually means this: across a set of answers, what fraction of the total brand mentions are yours?

If an assistant answers fifty category questions and names brands a hundred times in total, and eleven of those mentions are you, your Share of Voice is 11 percent. It's a relative number by design. It moves when you gain mentions and also when a competitor loses them, which is useful for competitive framing and confusing for diagnosis.

Share of Voice answers a real question: in the conversations that matter, how loud are you relative to the field? The catch is that it's an aggregate. It tells you the score of the game without telling you which play failed. A Share of Voice that drops five points could mean your content fell out of retrieval, or that one influential listicle dropped you, or that a competitor published something that now gets cited in your place. The number can't distinguish these, and the fixes for them have nothing in common.

Share of Model: the AI-native reframe

Share of Model, sometimes written Share of Model or "share of AI," is the term that grew up specifically for generative engines. The distinction people draw is that Share of Voice is often measured across a whole channel or category, while Share of Model narrows to a single question: for a given prompt, does this model recommend you, and how prominently?

In practice most tools compute Share of Model close to how they compute Share of Voice, as a mention-share across a prompt set, but with two refinements that matter for AI specifically.

The first is prominence weighting. Being the first brand named in an answer is worth more than being fourth, and far more than a passing mention in a caveat. A good Share of Model score weights position, not just presence, because the model's ordering carries real signal about which recommendation a user acts on.

The second is per-model separation. ChatGPT, Claude, Gemini, and Perplexity do not agree with each other. They pull from different indexes, rerank differently, and were trained on different data. A single blended "Share of Model" across all four hides the fact that you might dominate Perplexity and be invisible in Gemini, which are different problems with different causes. Any Share of Model worth tracking is reported per model, not averaged into one comforting number.

Share of Model is a genuine improvement on Share of Voice for this medium. But it shares the deeper limitation: it measures the outcome, the visible answer, after every upstream stage has already run. When it moves, the cause is somewhere behind it, and you're left reverse-engineering why.

Share of Retrieval: the metric one stage upstream

Both metrics above read the final answer. Share of Retrieval reads the stage before the answer exists.

When an AI assistant answers a question, it rarely searches your exact words. It rewrites the question into several sub-queries, runs them against a search backend, and gets back a candidate set of sources for each. A reranker trims that set to a handful of passages, those passages go into the context window, and only then does the model write an answer and decide whom to name. Retrieval is the gate near the front of that process. Share of Retrieval measures whether you make it through: across the sub-queries your target prompts generate, in what fraction of candidate sets does your content appear, before reranking, before synthesis, before naming?

The reason this is worth measuring separately is that it isolates a cause the other two blend together. Share of Voice and Share of Model can only fall for reasons they cannot name. Share of Retrieval names one of them directly. If you're not in the candidate set, nothing else in the pipeline can save you, and no amount of on-page copywriting changes an answer you were never a candidate for.

Two properties make it the better day-to-day lever. It's causally upstream, so it's a necessary condition for the other two rather than a lagging summary of them. And it's less noisy, because generated answers are sampled and swing from run to run, while whether your URL surfaces in retrieval for a given query is comparatively stable. A metric that moves less for random reasons is a metric you can actually optimize against.

How the three nest

The cleanest way to hold these in your head is as a funnel, each metric measuring a narrower survivor set than the last.

Everything retrievable → in the candidate set → cited in context → named in the answer

Share of Retrieval measures the second stage: did you enter the candidate set. Share of Model and Share of Voice both measure the last stage: did you get named, differing mainly in whether they score one prompt or a whole category and how they weight prominence. Nothing reaches the final stage without clearing the earlier ones, which is the whole point. The output metrics are conditional on the retrieval metric, never the reverse.

This nesting is what makes the combination diagnostic. Watching only the output tells you that something changed. Watching retrieval alongside it tells you where. The ratio between "appeared in retrieval" and "named in the answer" is itself a number worth tracking: call it citation efficiency, the rate at which being in the room converts to being named. High retrieval with low efficiency is a content problem, because you're getting seen and passed over. Low retrieval is a distribution problem, because you're not getting seen at all. One measurement, two very different work streams.

When each metric misleads

Every one of these fails in a specific way if you read it alone.

Share of Voice misleads through aggregation. Because it's a category-wide relative figure, it can hold steady while your position collapses on the prompts that actually drive revenue, as long as you gain ground on low-value ones. It can also swing purely because of a competitor's move, handing you a "win" you did nothing to earn and can't repeat. Read it for competitive context, not for cause.

Share of Model misleads through sampling and blending. A single query, run once, is one draw from a distribution, so a screenshot of "we're number two in ChatGPT" is weak evidence on its own. And a Share of Model averaged across engines can look healthy while hiding total invisibility on one of them. Read it per model, and over many samples, or it will lie to you in both directions.

Share of Retrieval misleads if you treat it as the finish line. Entering the candidate set is necessary, not sufficient. You can have strong retrieval and still lose every answer because your passages are generic and the reranker prefers a sharper source. Retrieval tells you the door is open. It does not tell you that you walked through it, so you track it next to an output metric rather than instead of one.

Which to track, and when

The practical answer is that you track all three, but you reach for different ones depending on the question in front of you.

When the question is competitive, "are we winning our category in AI answers," Share of Voice is the right frame, because relative standing is the thing you actually care about there. When the question is specific, "does ChatGPT recommend us for this exact buyer prompt, and how prominently," Share of Model per engine is the sharper instrument. And when the question is causal, "why did our visibility move, and what do we change," Share of Retrieval is the one that points at an answer, because it sits close enough to the cause to name it.

The mistake to avoid is steering by a single blended "AI visibility" score, whichever of the three it happens to be built from. That number will rise and fall for reasons it cannot explain, and you'll find yourself shipping content changes in response to movement that had nothing to do with content. Track the outcome so you know the score. Track retrieval so you know why. The gap between the two is where the actual work lives.

A note on measurement

You can compute a first version of all three yourself from the same raw data: a fixed prompt set run repeatedly against each engine's API, with two things logged per run, the sources the engine retrieved and the brands it named. Mention-share of the named brands gives you Share of Voice and, filtered to one prompt and weighted by position, Share of Model. Appearance-rate in the retrieved sources gives you Share of Retrieval. The engineering that turns that into a trend line across models, regions, and time, rather than a spreadsheet you re-run by hand, is what a dedicated tool like shareof.ai↗ exists to handle. The concepts, though, are yours to reason about, and reasoning about them correctly is most of the battle.

Primary sources and evidence

• Web search in the OpenAI API↗ — current search-tool behavior and citations.
• OpenAI web crawlers↗ — official user-agent controls for search and training.

Continue through the system

• The AI Visibility Funnel: A Diagnostic for Where You're Losing Citations
• Map Your Citation Surface: Every Page Where AI Can Find You
• How AI Models Actually Choose Which Brands to Cite

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.