Share of voice, mention rate, recommendation rate: the metrics of AI visibility

When a marketer asks "how visible are we in ChatGPT?", they're asking a question that sounds like "what do we rank for?" but is statistically closer to "what's our conversion rate?". There is no rank to look up. There is a distribution of possible answers, and your job is to estimate — honestly — how often your brand shows up in it, and how favorably.

That framing matters, because it dictates everything downstream: what you count, how many samples you need, and which comparisons are legitimate. This post pins down the three core metrics of AI visibility, works through a small example with real numbers, and then covers the reporting traps that make most "AI visibility dashboards" quietly meaningless.

There is no rank, only a distribution

A search results page is a fixed artifact: query it twice, you get the same ten links. A generative answer is a draw. The same prompt, sent to the same engine twice in the same hour, can name different brands — because decoding is sampled, because retrieval pulls different sources, and because system prompts and model versions shift under your feet.

So AI visibility metrics are defined over a sampling frame: a fixed suite of buyer-intent prompts, each run multiple times, per engine, per time window. Change any element of that frame and you've changed the metric. Everything below assumes one: say, 25 prompts × 8 runs on a single engine in one week — 200 sampled answers.

The stakes of getting this right keep rising. Gartner projected traditional search volume falling roughly 25% by 2026 as buyers shift to assistants,² and Pew found users click a conventional result on only ~8% of visits when an AI summary is present, versus ~15% without one.³ The answer increasingly is the shortlist — so measuring your presence in it is the new rank tracking.

Mention rate: are you in the answer at all?

Definition: the share of sampled answers in which your brand is named at least once.

Mention rate is your awareness metric — the AI-era analogue of "do we appear on page one?". It's deliberately generous: a mention counts whether you're recommended, listed as one option among six, or named in passing. That generosity is a feature. It gives you a stable, high-sensitivity signal for "does this engine know we exist in this category?", which is the first question every program has to answer.

Two scope decisions to make explicit before you count:

• Entity resolution. "HubSpot", "Hubspot CRM" and "hubspot.com" are one brand. Decide what counts as a match (brand, product lines, domain) and apply it identically to competitors.
• Unit of counting. Mention rate counts answers, not mentions. An answer naming you three times counts once. (Total mention counts come back for share of voice, below.)

Where it fails alone: mention rate can't distinguish "the market leader" from "avoid this one — support is poor". A rising mention rate driven by unfavorable comparisons is not a win. It also says nothing about competitors: a 40% mention rate is dominant in a category where the leader has 45%, and marginal where the leader has 90%.

Recommendation rate: are you the pick?

Definition: the share of sampled answers in which your brand is named positively — recommended, endorsed, or listed first in a ranked recommendation.

This is the metric that correlates with pipeline. Assistants tend to produce short, opinionated shortlists — often two to four names with a soft ranking — and buyers act on the top of that list disproportionately, just as they did with blue links. Being brand #1 of 3 in the answer is a different commercial outcome from being brand #6 of 6.

Recommendation rate requires a judgment call per answer: was this a recommendation or a mere mention? Codify it. A workable rubric: count an answer as a recommendation if the brand is (a) explicitly endorsed ("X is the best fit for…"), or (b) listed first where the answer implies ordering. Everything else — neutral listings, negative mentions, "X exists but…" — is a mention only. Whether a human or an LLM judge applies the rubric, write it down and version it, because a rubric change moves the metric more than most marketing campaigns will.

Where it fails alone: recommendation rate is always ≤ mention rate, and on realistic sample sizes it's built on fewer events — so it's noisier, and small movements are usually sampling error. It also hides category dynamics: your recommendation rate can hold steady while a new competitor eats every answer you're not in.

Share of voice: how much of the conversation is yours?

Definition: your brand's total mentions divided by all brand mentions in the category, across the sampled answers.

Share of voice (SOV) is the competitive lens. It answers the question your CMO is actually asking: when the assistant talks about this category, how much of the airtime is ours? Because generative answers typically name only a handful of brands, SOV is close to zero-sum — your gain is, mechanically, someone's loss. That makes it the best single metric for tracking competitive displacement over time.

Where it fails alone: SOV is a ratio, and ratios hide the denominator. If the engine starts naming fewer brands per answer, your SOV can climb while your mention rate — your actual probability of appearing — falls. The reverse happens too: a burst of comparison-style answers naming eight brands each dilutes everyone's SOV without anyone getting less visible. Always read SOV next to mention rate.

The qualifiers: accuracy and sentiment

The three rates tell you whether and how much you appear. Two qualifiers tell you whether the appearances help:

• Accuracy — is what the model says about you true? Wrong pricing, discontinued SKUs, a competitor's feature attributed to you. An inaccurate mention can be worse than absence, because the buyer walks away confidently misinformed.
• Sentiment — of your mentions, what fraction are positive, neutral, negative? This is the bridge between mention rate and recommendation rate: it tells you whether growing visibility is growing advocacy or growing criticism.

A worked example, with real arithmetic

Suppose you run a project-management tool and sample one engine with 25 buyer-intent prompts × 8 runs = 200 answers. Coding the transcripts gives you:

• Your brand named in 62 of 200 answers → mention rate = 62/200 = 31%. With a 95% Wilson interval, that's roughly 25–38% — the honest way to say it.
• Named positively or first in 34 of 200 → recommendation rate = 17% (95% CI ≈ 12–23%). So about half your appearances are endorsements — a healthy conversion of mentions into preference.
• Across all 200 answers, 540 total brand mentions in the category (≈2.7 brands per answer). Your brand accounts for 62 of them → share of voice = 62/540 ≈ 11.5%.

Notice what the combination reveals that no single number could. A 31% mention rate sounds respectable — but an 11.5% SOV says competitors collectively out-mention you nearly 8-to-1, meaning you're usually one name on someone else's shortlist. Meanwhile the 17% recommendation rate says that when you appear, the engine likes you. Diagnosis: a reach problem, not a reputation problem — the fix is more corroborating source material, not better positioning copy. The GEO literature backs the levers: adding citations, quotations and statistics to source content lifted generative visibility by up to ~40% in controlled tests.¹

The definitions, in one table

Four reporting traps that produce confident nonsense

1. Cherry-picked prompts. If the person building the suite also owns the KPI, prompts drift toward phrasings the brand wins ("best affordable PM tool for agencies" when that's your exact tagline). Fix: define the suite from buyer research before measuring, freeze it, and version any change.
2. Single runs. One run per prompt turns a probability into a coin flip you observed once. A 25-prompt suite run once has a resolution of 4 percentage points and enormous variance on top. Repeat runs are not optional; they are the measurement.
3. Engine mixing. Averaging ChatGPT, Gemini and Perplexity into one "AI visibility score" hides everything useful. Engines differ in retrieval, corpora and answer style; a Perplexity citation win and a ChatGPT training-data gap cancel into a flat line. Report per engine; aggregate, if at all, as a clearly-labeled composite.
4. No confidence intervals. "Mention rate rose from 28% to 33%" is not a finding if both numbers carry ±7-point intervals. Every rate in this post is a binomial proportion estimated from a sample; report the interval or invite your team to chase noise. (We dig into exactly this — with the math — in the companion post on confidence intervals.)

A checklist worth taping to the dashboard:

• Prompt suite defined from buyer language, frozen and versioned
• Multiple runs per prompt, per engine, per measurement wave
• Metrics reported per engine, never silently averaged
• Every rate shown with sample size and 95% interval
• Mention rate, recommendation rate and SOV always shown together
• Accuracy and sentiment audited on a sample of transcripts each wave

Frequently asked questions