How We Read Share of Voice for a Client Without Fooling Ourselves

Share of voice is the metric clients grasp fastest. "Out of every ten times an AI engine names a brand in this category, we are one of them this many times." It looks like a pie chart and feels like a market share. That is exactly why it needs care.

We use it in every client program. We also spend a good part of the first month teaching clients what it does not say.

What the number actually measures

Share of voice in AI answers is a ratio. The numerator is how often your brand is named across a set of responses. The denominator is how often all tracked brands are named across the same responses. Change the set of prompts, the set of brands, or the set of engines, and the ratio changes with no change in the real world.

So the first discipline is naming the denominator. When we send a chart we state the prompts (how many, which types), the engines, the country and language, and the competitors included. A share of voice figure without those four is a decoration.

Where it misleads

A small competitor set flatters you. If you track you plus two rivals, your share is high by construction. Add the five other names buyers actually consider and it drops. Neither number is wrong. They answer different questions.

Prompt mix matters more than prompt count. A panel heavy on brand-name prompts will show a high share because the brand is in the question. We keep brand prompts in their own group and report them separately from organic category prompts. The organic group is where growth is earned.

Engines differ. A brand can lead on Perplexity and trail on ChatGPT. Averaging the two hides the interesting fact. We show the split by engine first and the blend second.

Mentions are not endorsements. Being named as a cautionary example counts the same as being recommended unless you read sentiment. That is why we always pair share of voice with sentiment and position in the answer.

The three companions we always chart beside it

Share of voice on its own is a vanity figure. Beside these, it becomes a diagnostic.

• Position in answer. First-named brands tend to capture the reader's attention. A brand that is present but always fifth has a different problem from one that is absent.
• Sentiment. Promptwatch scores each response and gives a weekly summary. We read the lowest-scored responses by hand, because they usually point to a specific page or review that is shaping the tone.
• Citation rate. Are the answers linking to us, or naming us from memory? A brand named without a citation depends on training data, which updates slowly. A brand cited from a live page can move faster.

How we set up the measurement

In practice we do this inside Promptwatch, which tracks Visibility Score, Share of Voice, Sentiment, and Position in Answer from daily prompt panels. A few settings make the numbers cleaner.

We set brand aliases so that "Acme", "Acme Inc", and the product name count as one entity. We write the Brand Book analysis instructions so the system knows what counts as a mention, for example whether a mention of a parent company counts. We merge duplicate competitors and let the auto-discovery list suggest names we forgot. Several times it has surfaced a rival the client did not realise buyers were comparing them with.

For presentation we lean on the heatmap, which compares any two dimensions. Brands by model is the one clients get immediately: rows of competitors, columns of engines, colour for visibility. It shows in five seconds where the weak spots are.

How we report movement

We never report a one-day change. Responses vary run to run, and a daily wobble is noise. We report the change across a period and, where useful, use the A/B period comparison to show before and after a specific piece of work.

When a figure moves we tie it to something. Possible causes, in the order we check them:

1. Did the panel change? (It should not have mid-month.)
2. Did a competitor publish or get cited by a major source?
3. Did the client ship or edit anything?
4. Did crawler access change?
5. Did the engine change its behaviour?

The fifth explanation is the one clients reach for first and the one we rule in last. It is possible, but it is unfalsifiable if you reach for it too early.

What we promise

We do not promise a share of voice number. We promise a defined panel, a fixed method, and a list of actions tied to what the data shows. If a client asks for a target, we agree on a direction and a time frame and write down that nobody controls the engines. Illustrative targets in a planning meeting are fine, as long as everyone knows they are illustrative.

A worked example, clearly invented

Imagine a panel where Brand A is named in a small fraction of organic category answers, and a competitor appears in many more. The instinct is to write more content. The better first move is to look at which sources the engines cite for those answers. If most of them are third-party comparison pages and a particular community thread, the fastest improvement may be earning a place on those pages, not publishing another blog post on your own site. The citation and offsite mention reports are where that shows up.

Our recommendation

Treat share of voice as the headline and everything under it as the explanation. Run it on a platform that gives you the explanation in the same place. We use Promptwatch for this because the dashboards, heatmap, citation analytics, and crawler data share one project. If you want to see how it looks on your own brand, start at promptwatch.com and run the free report first.