How to measure AI visibility: share of voice, sentiment and citations
Buyers ask AI before they search. Here are the three numbers that show how engines describe and cite you — and why you must read them across engines, not just one.
You can't improve what you can't see
Rankings have tools, dashboards and decades of habit. AI answers have almost none. Yet a growing share of buyers now ask ChatGPT, Gemini or Perplexity before they ever open a search results page — and the answer they get rarely lists ten blue links. It names two or three brands, describes them in a sentence each, and maybe cites a source. If you're not one of those names, the buyer never knows you exist. If you are named but described badly, the mention works against you.
That is the problem AI visibility measurement solves. The hard part is that none of it shows up in your analytics. There's no "AI answer" line in Search Console, no impressions count, no click curve. The only way to know how the engines talk about you is to ask them the questions your buyers ask — repeatedly, across engines, and on a schedule — and score the answers.
The three numbers that matter
AI visibility isn't one metric. Three numbers, read together, tell the story:
- Share of voice — across a fixed set of buying-intent prompts, how often is your brand named versus competitors? This is the closest thing to a "ranking" the AI surface has.
- Sentiment — being mentioned is not enough; *how* you're described (trusted, premium, cheap, outdated, "good for beginners") shapes whether the buyer clicks toward you or past you.
- Citations — which of your pages, if any, the engine links as its source. Citations are where AI visibility loops back into ordinary SEO: the pages that get cited are pages you can deliberately strengthen.
Build the prompt set first
Every number above is only as good as the prompts behind it. A useful prompt set is small, fixed and buyer-shaped — not a keyword list. Group prompts by the job the buyer is doing:
- Category discovery — "best CRM for a 5-person agency", "alternatives to Mailchimp"
- Comparison — "Notion vs Coda for a wiki", "is Ahrefs or Semrush better for a small site"
- Qualification — "is [your brand] good for enterprise", "does [your brand] have a free plan"
- Problem-first — "how do I track keyword rankings without a spreadsheet"
Twenty to forty prompts is plenty to start. Lock the wording and reuse it every cycle — the value is in the trend, and a trend only exists if the question stays the same.
A worked example
Say you sell rank-tracking software and you run a 30-prompt set across three engines. After the first scan, you tally how often each brand is named. Here's a realistic read for a mid-market challenger:
You're named in roughly a third of answers — behind the two category leaders but ahead of the long tail. That's your baseline. Now break the same scan down by engine, because the aggregate hides where the work is:
| Engine | Times named (of 30) | Sentiment | Cited your domain |
|---|---|---|---|
| ChatGPT | 6 | neutral / "newer option" | no |
| Perplexity | 14 | positive | yes (2 pages) |
| Gemini | 11 | mixed | yes (1 page) |
The aggregate "34%" was misleading. You're actually doing well on Perplexity, where live-web citation rewards good pages, and badly on ChatGPT, where model-memory still treats you as the new kid. Those two gaps need completely different fixes — and you'd never see them from a single blended score.
Measure across engines, not just one
Each engine has its own training data, retrieval method and biases. A brand that's strong in ChatGPT can be invisible in Perplexity, and vice versa, for reasons that have nothing to do with quality:
| Engine | What it reveals | What moves it |
|---|---|---|
| ChatGPT | Mainstream model-memory + how it frames you | Broad, durable web presence; brand mentions across the open web |
| Perplexity | Live-web citations you can win this week | Pages that directly answer the prompt, recently and clearly |
| Gemini | Google-adjacent answers + AI Overview overlap | The same signals that earn Google AI Overviews and featured snippets |
The practical takeaway: ChatGPT is a slow surface you improve by being broadly cited and talked about over months, while Perplexity is a fast surface you can win this quarter by publishing pages that answer the prompt cleanly. Treating them as one number averages those two timelines into something you can't act on.
Measure the same prompts on a schedule. A single snapshot is a guess; a trend line is a strategy. Re-run weekly or monthly and watch the share-of-voice bars move — that's your feedback loop.
Turn the read into moves
A score you don't act on is a vanity metric. Each of the three numbers points at a different lever:
- Low share of voice → you have a presence problem. Win mentions on the open web (reviews, comparison pages, directories, podcasts) so models and retrievers have something to surface.
- Negative or stale sentiment → you have a narrative problem. Find the prompts where you're called "outdated" or "limited", then publish the content and proof that contradicts it — and watch the sentiment flip on the next scan.
- No citations → you have a content problem. Identify the prompts where a competitor's page is cited and yours isn't, then build the page that answers that exact question better.
Tie this back to your classic SEO too. Pages that earn Perplexity and Gemini citations are usually the same pages that earn featured snippets and AI Overviews, so the work compounds across both surfaces. And because buyers who arrive via an AI answer are deep in a decision, a one-point share-of-voice gain on a high-intent prompt is often worth more than ten ranking positions on a vague one.
Where Lume fits
Lume runs your prompt set across Gemini, ChatGPT and Perplexity, scores your share of voice and sentiment, tracks the trend over time, and shows the exact prompts where a competitor is winning the mention or the citation. Pair it with daily rank tracking across 100k+ locations and you can watch both the classic and the AI surfaces in one place. Run a free scan and see how AI describes you today.
