How to Measure Share of Voice in AI-Generated Answers

AI share of voice shows if a language model brings up, references, or suggests a brand while responding to an actual question, and how that performance compares to other brands. Here’s how this article will walk you through constructing the metric, from designing the prompts to logging the results, crunching the numbers, and figuring out what those numbers actually mean. Many teams begin on the wrong footing, which ultimately undermines their entire effort. Further details on this will be discussed later.

The way people search has changed. More and more research, particularly by B2B buyers evaluating vendors or tackling technical questions, is now happening in AI tools like ChatGPT, Claude, Gemini, and Perplexity, rather than on Google. People coming from AI tools are more likely to become customers compared to those who find the site through regular search engines. Although AI-driven traffic is relatively small for most websites, its value is higher because it tends to occur at a later stage in the decision-making process. Google’s AI-generated summaries now appear on nearly every search, and users are clicking through to the traditional results far less often than before. People are focusing more on the answer directly, even before they encounter a blue link.

That shift changes what "invisible" means for a brand. Not showing up in AI-generated responses is becoming as damaging for a brand as vanishing from Google’s first page, yet there’s no tool like Search Console to alert you when it happens. Nobody gets an email saying "your citation rate dropped 12 points this month." Generative referral traffic has grown faster year over year than almost any other channel, and the brands tracking their standing in these answers systematically are pulling ahead of the ones who just type their own name into ChatGPT and eyeball the result.

What AI share of voice actually measures and how to define it precisely

Start with a collection of prompts. Run these prompts through the models. It's the proportion of responses that mention a specific brand, either by citing or recommending it, compared to the total number of times that brand is mentioned across all responses to the set of prompts.

To put it simply, the formula is your brand's mentions divided by the total mentions of all brands being tracked, then multiplied by 100 to get the percentage. That's the percentage.

It captures two separate things at once. First, it checks if the brand appears in the results at all, a simple yes-or-no test of visibility. Second, whether it shows up more than the brands it competes with, which is the competitive question. A brand might do well in terms of visibility but still fall behind its rivals if, for example, three competitors are mentioned twice as frequently.

The calculation relies on four distinct components, each with its own role, and swapping them wouldn’t work.

• How often is the brand name mentioned: is it present in the text of the response?

• Citation: does it include a link from the brand's own website as a reference?

• Prominence refers to whether the mention appears in a prominent spot, like the opening line, or is tucked away in a less noticeable place, such as a footnote.

• Sentiment: what does the model say about the brand when it mentions it?

Thing is wrong. A brand's website might appear in a list of sources at the end of an AI's answer, even if the answer itself doesn't mention the brand or influence the user's choice. If you don’t track those two things apart, the metric will just blur the real picture.

To be meaningful, analysis requires two essential references: a clear list of competing products or services and a specific set of questions or topics used to evaluate them. The phrase "Our AI SOV is [some figure]" doesn't convey any real information until it specifies what the 22% refers to and who it is being compared to. In most industries, AI responses rarely mention brands, except for a handful of top players who dominate the conversation. That difference represents a significant opportunity, and it's quite large.

How AI SOV differs from traditional share of voice and why those differences change the measurement approach

Traditional share of voice tracks mentions of a brand in news articles, broadcast media, and social media posts. AI SOV is a bit odd, as it tracks whether an AI, when creating an answer from the ground up, chooses to include your brand. This process involves retrieval and reasoning, not just counting media mentions, and mistakenly treating it as a media count is where most measurement plans fundamentally fail.

Achieving a high rank on Google doesn't have much value in this context. Research looking at large sets of AI prompts has found the overlap between what AI models cite and what ranks on page one organically is small, well under a fifth of citations in most cases, and lower still on ChatGPT specifically. Ranking on Google's first page and being cited by an AI model serve distinct purposes for separate systems. Those who equate a good SEO performance with strong AI visibility are actually measuring the wrong metric.

Different platforms treat a “mention” as having different levels of influence. Perplexity usually lists more sources for its answers, but each source has a smaller impact on the final result. Even though ChatGPT references fewer sources overall, the ones it does include carry significantly more weight in shaping its replies. This means that tracking citations in a spreadsheet, for example, doesn't necessarily capture their equal importance, since citations on different platforms have different weights.

The range of sources is broader than anything Google has ever gathered for its rankings. From forums and review sites to Reddit threads and indexed community Slack exports, all these sources contribute to how a model represents a brand. AI Share of Voice captures a brand's presence across the whole internet, not limited to just its own website. No ready-made dashboard exists for tracking any of this yet. For the time being, all of this must be assembled manually.

How each major AI platform decides which brands to surface

These four distinct AI systems, despite their unique approaches, all prioritize brands that appear reliably in various third-party contexts rather than relying solely on their own promotional materials.

ChatGPT, Claude, Gemini, and Perplexity combine their training data with real-time information, sourcing content from Bing, Google, and Brave, while using safety measures to determine what information is displayed. The methods they use to gather information vary significantly beneath this common framework.

Gemini directly accesses the Google ecosystem, including Shopping Graph data and Google's search index, which gives it the broadest reach among the four platforms for detecting brand signals. Using schema markup significantly boosts the chances of citations in Gemini's responses, more so than on the other three platforms.

ChatGPT routinely pulls information from Wikipedia and independent review sites when generating responses. Brands that build a solid, trustworthy reputation on review sites gain a significant advantage.

Claude favors editorial content that's been earned through traditional reporting, typically from reputable outlets with a proven history of in-depth journalism. If a brand has been featured in trade publications or major media outlets, it gains a built-in edge in how Claude evaluates sources.

Many people often miss this detail: Gemini processes JavaScript. Unlike Gemini, Claude and Perplexity mostly skip JavaScript, relying instead on static HTML to interpret web content. If a brand has a website that relies heavily on JavaScript and its content only appears after the client-side loads, it might as well be invisible to two out of the four main AI models, regardless of how valuable the content is.

The concentration of sources is extremely high. A few websites are cited far more often by these AI models than would be expected, with this concentration being greater than what's seen in traditional search results. That means the ecosystem a brand shows up in, not just the brand's own site, carries real weight. When it comes to measurement, the key takeaway is that brands should track their share of voice separately for each platform, rather than combining them into a single, overall number. A brand might improve its standing on Gemini while dropping on Claude within the same month, due to the unique sourcing practices of each platform.

Building the prompt set that the entire measurement depends on

The biggest mistake is creating a prompt set using only brand-specific questions. "Tell me about [Brand]." Those prompts prove almost nothing, because a model asked directly about a brand will usually say something about it. True AI visibility shows up when a brand appears in responses to category-based questions, like “best project management tools”, not because the user named it, but because the AI chose it as a relevant answer. If you go straight to prompts about specific brands, you're just measuring how well the AI praises them, not how visible they are overall.

A prompt library must include four distinct types of questions to function properly.

• Prompts that compare products, contrasting the brand's category with specific competitors.

• These prompts link a specific problem to a category of solutions.

• Prompts that assess a company’s deep knowledge of its field, measuring how well it demonstrates thought leadership and commands authority within its industry category

• Prompts tailored to particular features, addressing both their abilities and real-world applications.

In each group, create a series of increasingly specific terms. Begin with a wide category like “project management tools”, then narrow it to a more focused group such as “project management tools for agencies”, and finally zoom in further to “project management tools for remote agency teams under fifty people”. The more detailed the category becomes, the more a brand’s prominence can shift dramatically, what’s barely noticeable at a general level might suddenly lead the pack in a hyper-specific niche, and vice versa.

Aiming for twenty to thirty varied prompts is a good rule of thumb for building a functional library that covers both initial awareness and final decision-making stages. New teams can begin with ten to twenty prompts and add more as their resources grow. It's not necessary for the complete set to be ready from the very first day.

The most valuable prompts are those where a brand is expected to be relevant but isn't mentioned in the question itself. Those measure whether the brand shows up naturally in responses, without being mentioned in the prompt, unlike queries that simply echo back what the user already wrote.

For each prompt, explain why you chose it: what question about a competitor it represents and which stage of the customer journey it targets. Documenting your logic now prevents you from later drowning in a six-month pile of raw data that no one can actually interpret.

Running queries consistently across models and recording results in a format that yields a trend

Four different platforms are monitored individually in separate columns. The four platforms tracked are ChatGPT, Claude, Gemini, and Perplexity. Combine them into a single figure should be avoided; the way they cite sources varies so much that averaging would obscure the specific trend that's important to observe.

Most teams underestimate how critical timing is when they start this process. If your brand operates in a competitive space, run these checks every week or every two weeks. If a brand wants to see a trend over time rather than a potentially misleading snapshot, it should check at least once a month.

The way AI answers can vary from one session to the next is mostly because of tweaks like adjusting the model’s randomness, changes in the search data it pulls from, or how much text it can process at once, none of which reflect how well a brand is actually perceived. To get a reliable reading, it's better to run the same prompt multiple times during a session, then note the most common response, rather than relying on the result of just one try.

A tracking sheet should include at least the following information: the text of the query, the platform used, the date, whether the brand was mentioned, the location of the mention (beginning, middle, or end), the competitors that were referenced, the sources that were cited, and a sentiment score.

For small prompt sets, up to a few dozen variations, manual tracking is manageable. Beyond a certain number, manually querying becomes impractical, and using an API for automation is the only feasible solution. Tools that connect via API can take care of this whole process, sending prompts, running them across different AI models, and compiling the outputs, so your team doesn’t have to write that system from scratch.

Here's another useful data point to track: referral traffic from AI websites like chat.openai.com, claude.ai, gemini.google.com, and perplexity.ai, as seen in GA4. By creating a custom channel grouping that isolates this traffic, you get another, separate indicator that moves in tandem with the manual prompt data.

Calculating AI SOV from raw mention data, and weighting it for prominence and sentiment

Begin with the basics. For every prompt you run, tally all the times each brand is mentioned in the responses, and then divide the number of mentions for the specific brand by that overall total. This is the raw share of voice based on mentions. It can be consistently repeated and compared across different periods, so it's a solid beginning. However, it doesn't go any further for a specific reason.

Counting mentions by themselves overlooks a key point: a brand that’s mentioned just once, briefly and near the end of a response, has far less impact than one positioned as the top choice right at the start.

This method of weighting by prominence resolves that issue. A simple three-tier system works well enough: primary recommendation, secondary mention, incidental reference. Assign different values to each tier, and the final figure will show real impact instead of merely counting words.

Sentiment should be tracked independently, without combining it with other metrics. Negative or cautionary mentions have a distinct impact, actively causing harm, rather than being neutral. If a brand appears in a response that includes a caution, it could actually be more detrimental than if it didn't appear at all. Assign each mention a sentiment label, positive, neutral, or negative, and record it separately from the prominence score instead of combining the two.

Combine the two, weighting each mention by both its prominence tier and its sentiment, before dividing into the total. This approach yields a quality-adjusted SOV that accurately gauges a model's brand representation, taking into account not just the frequency of mentions, but also the significance and tone of those mentions.

It's important to keep the original pieces of data showing, even after you've created the overall score. Raw mention counts, prominence ratings, and sentiment analysis should all be monitored individually. If the combined score falls, it might reflect fewer mentions, poorer placement, more negative framing, or a combination of these factors, but without breaking down the individual components, it’s impossible to pinpoint the exact cause. Also, it's important to maintain the scores for each platform individually; combining them into a single number for all platforms obscures the specific platform-related changes that require a tailored approach.

Benchmarking your SOV against competitors to make the number strategically meaningful

Test your brand and two to four direct competitors using the same set of prompts, on the same schedule, and with the same scoring approach. Using the same approach every time is what matters most. Even a single tweak in how you track versus how they do, whether it’s the prompts, timing, or scoring, will throw off the whole comparison.

For each platform and prompt category, create a snapshot of the competitive landscape by listing the share of every tracked brand alongside one another. The total of the shares should add up closely to the overall named mentions in that group, while some mentions will account for brands not included in the tracked list.

What really counts is whether the number is going up or down. Is the brand gaining or losing ground compared to a particular competitor, on a particular platform and within a particular type of prompt? That's the question worth answering every cycle, not "what's our score this week."

Look at the data separately for each category. For instance, a brand could be far ahead in mentions related to solving customer problems, but lag significantly when it comes to being name-checked in product comparisons. This division clearly indicates the areas where future content funding should be focused.

In traditional media, a brand's share of voice often predicts its future market share, rather than reflecting its current position. This reasoning probably applies here as well. Brands that are showing up more often in AI-generated answers right now are likely shaping which options people will think of later, even if the immediate website visits or clicks haven’t spiked yet.

Also, pay attention to unusual changes. You should investigate if a competitor suddenly begins showing up in search prompts they didn't appear in previously, as this indicates a likely change, such as a new public relations effort, a surge in content creation, or the launch of a review site promotion. This calls for deeper analysis rather than just noting it down.

Interpreting trend data and deciding which changes in AI SOV are worth acting on

One week's worth of data doesn't show much at all. The model's performance fluctuates significantly from one session to the next, meaning that a single week's drop or increase might just be random variation. Wait for a trend to persist across several cycles before making decisions or adjusting your content strategy.

Three kinds of signal are worth real attention. If mentions of a particular prompt category consistently decrease, it indicates that the content supporting that category has somehow become less authoritative. A change from positive feelings to neutral or warning tones indicates that something in the wider information landscape, like reviews, forum discussions, or news, is affecting how models discuss the brand. When a rival starts outpacing your brand in the prompts where you were previously dominant, that’s a clear warning sign about your competitive standing.

To confirm the trend, compare the SOV changes with the referral traffic data from GA4 AI. When both SOV and referral traffic decrease at the same time, it indicates a clearer, more reliable signal. When SOV decreases but referral traffic remains stable, it suggests the problem lies in the brand's portrayal rather than its overall visibility.

The importance of recent content is greater than many anticipate. AI systems usually give more weight to newer content than to older pieces that discuss the same topic. When AI's share of voice drops because content is getting old, it usually means you need to publish more often, not that your strategy is flawed; often, simply updating existing material can solve the issue.

Typically, changes in sentiment occur before changes in frequency. Think of sentiment as an early signal, monitor it regularly on its own, rather than waiting to see if it eventually shows up as a decline in the raw number of mentions.

Establish clear thresholds rather than manually checking each data point: a noticeable decline in weekly SOV or any sign of negative sentiment on a key prompt should automatically prompt a review. Lettertrace automatically flags these changes, saving you from the tedious task of comparing spreadsheets by hand each Monday.

What actions the SOV data should drive in content and distribution strategy

The effort put into these measurements is pointless unless they influence what content is released and where it's distributed. A metric that just sits on a dashboard and gets glanced at once every three months isn’t serving its purpose.

When share of voice is low for prompts about problems and their solutions, it suggests a clear approach: create content that directly addresses the issue, states a definite viewpoint, and supports it with data or references that an AI can readily access and utilize. A Princeton study on tuning generative AI engines discovered that embedding verifiable sources, verbatim quotes, concrete data, and properly attributed references in content yields the biggest boost in how often models cite it.

A low share of voice on comparison-related queries likely indicates a missed opportunity elsewhere in the content strategy. The same study discovered that structured comparison tables receive significantly more AI citations compared to the same content presented in paragraph form. If a brand's competitive positioning lives only in paragraph form, that's worth fixing before anything else on the list.

A brand that appears sporadically on the four platforms often has a weak entity signal, which generally stems from a limited source ecosystem. Forum activity, reviews from independent sites, and media coverage in outlets that aren’t owned by the brand, these are the signals that AI models use to confirm a brand exists and is credible enough to mention.

Fixes should be tailored to the specific gaps on each platform. When Claude hardly ever references a brand while ChatGPT consistently does, the difference usually comes down to whether the brand appears in editorial articles, Claude’s main reference point, or on review sites, which ChatGPT pays more attention to. Treating that as one generic "AI visibility problem" wastes effort; the fix looks different depending on which platform is lagging.

Issues with sentiment go beyond what a content strategy can address by itself. When AI responses portray a brand cautiously or negatively, it often stems from genuine issues embedded in the training data, such as outdated negative news, complaints on review sites, or unresolved discussions in forums. Addressing the issue requires work on reputation upstream, not just another blog post. No matter how much new content is produced, it won't cover up the problem.

Handle this entire system just like a seasoned marketing team handles SEO rankings: track it monthly to establish your baseline, set up weekly alerts for any rapid competitive shifts, and conduct deeper strategic reviews every quarter. Companies that incorporate this routine into their regular schedule are the ones that notice a change when there's still enough time to react, rather than discovering three months later that a competitor has subtly captured the market.