Share of voice in LLMs, explained properly

Share of voice is one of those phrases everyone uses and almost nobody defines the same way.

In classic SEO it usually meant your slice of clicks or impressions for a set of keywords. In AI search it has quietly turned into something else, and the tools measuring it do not always agree on what "it" is.

I want to fix that here, because I keep seeing decks where "share of voice" is a big confident number with no method behind it.

Quick disclosure before we go further: I build LLM Pulse, an AI visibility platform. So I have skin in this game. I will still tell you where the concept gets abused, including by tools in my own category, and where a competitor might suit you better.

What share of voice actually means in an LLM

Forget the SEO baggage for a second.

Share of voice in an LLM is simple to state: across a defined set of prompts, how often does your brand show up in the answer, relative to your competitors?

That is it. You pick the questions your buyers ask an AI. You run them. You count who gets mentioned. Your share is your mentions divided by the total mentions in that competitive set.

The devil, as always, lives in three words: "prompts", "show up", and "count".

Get any of those wrong and your number is fiction with a decimal point.

The prompt set is the real experiment

Before we argue about answers versus URLs, understand that your share of voice is only as honest as the prompts you chose.

If you sell project management software and your prompt set is stuffed with "best project management tool for remote teams", you will look great. Change it to "cheapest project management tool for freelancers" and your share can collapse, because a different set of brands owns that question.

Neither number is wrong. They answer different questions.

This is why I distrust any share of voice figure presented without its prompt set. A single percentage with no visible questions behind it is a vanity metric wearing a lab coat.

So rule one: define a prompt set that matches real demand, keep it stable so you can track it over time, and be honest about what it does and does not cover.

Answers versus URLs, and why it matters so much

Here is the split that most people miss.

There are two completely different things you can count when an LLM responds:

1. The answer text: did the model name your brand in what it wrote to the user?
2. The citations: did your URL appear in the sources or links the model attached?

These are not the same, and treating them as the same is the single most common measurement error I see.

An LLM can recommend "tools like Notion and Asana" in its answer while citing a G2 listicle and a Reddit thread as sources. In that response Notion and Asana have answer-level visibility. G2 and Reddit have citation-level visibility. The user reads the brand names, not the footnotes.

Now flip it. A model might cite your documentation page as a source but never mention your brand by name in the prose. You got the crawl and the citation. You did not get the recommendation.

If you only count URLs, you are measuring who the model reads. If you count answers, you are measuring who the model recommends.

For most brands, recommendation is the thing that moves revenue. A buyer asking ChatGPT "what should I use for X" acts on the names in the sentence, not on the blue links underneath that most of them never click.

Why counting answers is the harder, better metric

Counting URLs is easy. They are structured, they are right there in the response payload, any scraper can grab them.

Counting brand mentions in free text is harder. You have to handle the messy reality of language:

• Aliases and abbreviations. Is "GA4", "Google Analytics" and "Analytics" the same brand? Usually yes.
• False positives. "Apple" the company versus "apple" the fruit. "Notion" the app versus "the notion that".
• Negative context. Getting named as "avoid tools like X" is not the same as being recommended, and a naive counter scores both as a mention.

This is where a lot of cheap AI visibility counting falls apart. Matching a domain string is trivial. Understanding that "I would steer clear of BrandX" is a bad mention requires actually reading the sentence.

So when a tool tells you your share of voice, ask it a blunt question: are you counting my name in the answer, my URL in the citations, or some blend of both? All three are legitimate. Pretending they are interchangeable is not.

The pitfalls that quietly wreck the number

A few traps I have watched people fall into, including sharp people.

One model is not the market. ChatGPT, Perplexity and Gemini surface different brands for the same question. I have seen a brand sit at 40% share in one model and near zero in another. If your tool tracks a single engine and calls it "AI share of voice", you are getting one camera angle and being told it is the whole stadium.

Non-determinism. Ask the same model the same thing twice and you can get different brands. That is not a bug in the tool, it is how these systems work. The fix is volume and repetition: run each prompt multiple times, average across runs, track the trend rather than obsessing over a single day. A share of voice from one run per prompt is noise.

Position and prominence. Being mentioned first in a recommendation is worth more than being the fifth name in a hedge. A pure mention count treats both equally. It is a fine starting metric, but do not confuse "mentioned" with "recommended clearly".

Sentiment blindness. As I said, being named in a warning is a mention your counter will happily add to your total. Share of voice without sentiment can flatter a brand that AI is actually describing badly.

The wrong competitive set. Share is relative. If you benchmark against three tiny rivals and ignore the category leader, your share looks fantastic and means nothing. Pick the competitors your buyers actually compare you against.

How to benchmark against competitors without fooling yourself

A benchmark that is worth acting on has a few properties.

Same prompts for everyone. You and your competitors get scored on the identical question set, or the comparison is rigged.

Enough prompts to be stable. Fifty well chosen prompts across your real buying questions beat five hundred padded ones. But five prompts is not a benchmark, it is an anecdote.

Multiple models, reported separately and combined. You want to see the blended picture and the per model breakdown, because your weak spots hide in the average.

Tracked over time. A share of voice is a snapshot. The value is in the slope. Did your number move after you published that comparison page, earned that Reddit thread, got added to that listicle everyone cites? That is the question worth answering.

How LLM Pulse measures it, concretely

Since this is what I work on daily, here is how we do it, and I will keep it factual.

You define your prompt set. LLM Pulse runs those prompts across five core models on every plan: ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews. You can add Claude, Copilot, Grok, DeepSeek and Meta AI if you want wider coverage. That matters because share of voice on one model is, as I keep saying, one camera angle.

We count brand mentions in the answer, not just URLs in the citations, and we track citations separately so you can see both. Recommendation and readership are different, so we do not blend them into one mushy number.

Sentiment sits on top, so a mention in a negative context does not get quietly scored as a win. You can see the exact words the AI uses to describe you.

A concrete example of why the split matters. A tools brand I looked at was thrilled with its citation numbers: its docs were being crawled and cited constantly. But its answer-level share of voice was low, because the models were citing its pages while recommending two competitors by name. The URL count said "we are winning". The answer count said "we are the reference material and someone else is the recommendation". Only counting answers surfaces that gap. Counting URLs alone hides it.

Pricing is public and starts at 49 € per month, which in this category is still not the norm. Here is the shape of it, as of July 2026 (prices change, check the site before you buy):

Unlimited seats on every plan, no per seat fee, and there is a 14 day trial. There is also a hosted MCP server with 45+ tools over OAuth, included even on the trial, so you can pull your share of voice straight into Claude or ChatGPT and interrogate it.

Where a competitor might be the better call

I promised to be straight, so here it is.

If you are a Fortune 500 brand that needs the deepest possible data, prompt volume estimates, agent analytics, SOC 2, SSO and on-prem, Profound is built for exactly that. It is demo led with no public pricing, and for enterprises that depth earns its keep.

If you want the absolute cheapest honest entry point and you like seeing actual screenshots of AI outputs, Otterly.ai starts at $29 a month. If you are an agency or a team that wants strong per country tracking, Peec AI is a solid pick, just note that extra models are paid add-ons there.

And if you already pay for Ahrefs or Semrush, their AI visibility features are shallower than the specialists but they are already in your subscription. For a lot of people that is a reasonable place to start before spending more.

I think LLM Pulse hits the best balance of model coverage, honest answer level measurement and public pricing for most SEOs, agencies and founders. But "best for most" is not "best for everyone", and I would rather you pick the right tool than the flattering one.

The one thing to take away

Share of voice in LLMs is not a magic number. It is an experiment with a prompt set, a counting method and a competitor list, and every one of those is a choice someone made.

So next time a dashboard hands you a confident percentage, ask the only question that matters: are you counting who the model recommends, or just who it reads? The gap between those two answers is usually where the real work is.