Share of voice in AI answers, measured honestly

Two vendors demoed AI-visibility tools to the same team last week. The first showed the brand as nearly invisible in AI answers. The second showed it leading the category. Neither had lied, and the answers had not moved an inch.

Share of voice, your slice of the AI answer space measured against the competitors you track, is the number these tools most love to sell and the easiest to inflate. Change the prompt list, blend branded questions into the denominator, or count a passing mention as a win, and the same brand swings from invisible to leader.

The honest version is the core craft of answer engine optimization (AEO), the work of getting your pages cited in AI answers. It comes down to a few orderings, held even when they make your number smaller.

Citations rank above mentions, always

A citation and a mention are different events, and the gap between them is not subtle.

A citation is the engine showing its work: your page in the source list, doing the answering. A mention is your name in the answer with no page behind it, on grounds the engine never disclosed.

The difference shows up in what buyers do. When somebody asks an assistant for the best options in your category, the cited pages write the shortlist and the mentioned brands merely appear on it.

So the ordering is fixed. Citations above mentions, every time. Quattr reports both in that order, and when the two roll up into a single share-of-voice number, a citation carries more weight than a mention by design, because being named as a source is the harder of the two to earn.

Customers see the exact formula in the product. What never ships is a score that hides whether you were cited at all.

Your prompt basket decides what share means

Share of voice is a share of something, and that something is your prompt basket.

It is the set of questions replayed in the engines to measure you, and it decides more about your number than the engines do. The named anti-pattern is the hand-picked list: prompts chosen by the team being measured, drifting toward the brand's own name, producing you-typed-our-name theater.

Quattr generates baskets from first-party demand instead. Your search queries, your logs, your AI referrals. The denominator becomes the demand you actually compete for rather than the demand you would enjoy winning.

Branded prompts get the same treatment. Non-branded is the default lens, with branded reported separately and never silently blended. An audience that already typed your name is loyalty, not voice.

See also: the basket recipe, and why hand-picked lists lie →

Sentiment is a trend and a gap

There is no threshold at which AI sentiment becomes good. The honest reading is direction and distance. Is positive sentiment rising against your own trailing window, and where do you sit against competitors on the same prompts?

A sentiment percentage on its own is decoration. The same percentage as a trend line, with a competitor gap beside it, is a decision input.

Engines deserve the same treatment as metrics. ChatGPT, Perplexity, Gemini and AI Mode, which is Google's conversational search surface, all cite differently, and an aggregate can hold flat while you gain one and lose another. Read the per-engine cuts first and the blended view second. That ordering catches the trades a single line hides.

Separate the market's growth from yours

AI answer surfaces are still rolling out. Citation counts move because adoption moved, not necessarily because you did anything at all.

Every share-of-voice trend gets decomposed into market-size effect and share effect before anybody takes credit. If citations doubled while your share held flat, the market grew and you stood still inside it, and the deck should say so rather than calling the doubling a win.

Quattr applies the same decomposition on the search side, where the headline is CTR-modeled search market share of tracked demand.

In both worlds a rising tide is not your rising share, and that difference is the entire question a leadership team is asking when it asks how you are doing.

The question travels beyond your own reporting. Put it to any vendor demo. If the chart cannot say how much of the rise was adoption, it is not yet a measurement.

The workflow that does this: Two-quarter share trajectory →

The questions an honest number survives

Google's AI Overviews, the AI-written blocks on top of a classic results page, are not answer-engine citations. One is a block on the results page, the other is an assistant crediting its sources. The tools keep them apart: one family of analysis reads the answer-engine side, while AI Overview inclusion sits with the on-results-page competitive metrics.

Which brings us to the test. An honest readout survives the follow-up questions, and there are only four of them. When the orderings held, all four answers are already on the card.

• Which surface is this, results page or answer engine?
• Whose prompts is it measured on, and who chose them?
• Cited, or merely mentioned?
• Your motion, or the market's?

See also: two surfaces, two metrics, one common conflation →