Measuring and increasing your share of voice in AI answers

What share of voice in AI answers really means

Classic share of voice comes from advertising: it measures what share of a market's total visibility your brand holds, for example across ad placements or search-result real estate. In AI answers, the same idea moves to a new venue. Instead of ad slots or ranking positions, you count how often a language model names your brand when someone asks for solutions, providers or recommendations. The location of visibility changed; the underlying principle didn't.

The difference from classic search is real. Google still shows a list of ten or more results for the user to scan themselves. An AI system typically surfaces only two to five names and presents them as a recommendation, not a menu. Ahrefs' analysis of AI citations found that only about 6-8% of URLs cited by ChatGPT even appear in Google's top 10 for the same query, and roughly 80% don't rank in the top 100 at all — getting cited by an AI system is a different contest from ranking in search, with different winners. Whoever isn't on that shortlist simply doesn't exist for the user in that moment, which is why a share of these few mentions is worth more, and gets fought over harder, than a mid-page search ranking.

Across industries the pattern repeats: whether tax firm, tool manufacturer, software provider or bicycle shop, people now ask AI systems purchase-decision questions before they ever open a search results page. Pew Research found that when an AI summary appears, users click through to a traditional result only about 8% of the time, versus 15% without one — roughly half the rate. Your share of voice tells you whether you show up in these conversations or whether the model keeps recommending only your competitors. It is less a vanity metric than a direct read on visibility at the exact moment purchasing decisions get made.

The measurement basis: a question set instead of guesswork

A robust measurement starts with a fixed question set. Collect the questions your target group realistically asks, phrased the way they'd actually type or say them. Examples across industries: "Which project management software is suitable for small agencies?", "Who offers sustainable packaging for food?", "Which physiotherapy helps with runner's knee?" Fifty to two hundred such questions is a solid base. Keep the set constant, or your next measurement compares apples with oranges.

Ask each question multiple times and across several systems. AI answers aren't deterministic: the same question can produce a slightly different answer from one day to the next. That's why you measure a frequency, not a single hit. Ask a question ten times and your brand appears in four answers, and your mention rate for that question is forty percent. Only this repetition turns a random snapshot into a number you can act on.

Separate your measurement by system, too. ChatGPT without web browsing enabled answers purely from training data, Perplexity and Google AI Overviews pull from live sources at query time. These modes behave differently, and what wins in one won't necessarily win in the other. Run a separate evaluation per system, then combine them afterward. That's how you spot whether you're strong in source-backed answers but nearly invisible in pure model-knowledge answers.

From counting to a metric: how to calculate your share

The actual share of voice comes out of comparison. Define a competitive field, the handful of brands that realistically compete with you for the same recommendations. Then count every mention across your entire question set: yours and every competitor's. Your share of voice is your mentions divided by total mentions in the field. Reach 40 of 250 total mentions and your share is 16 percent.

Track two more numbers alongside it. Presence rate shows in what share of all answers you appear at all, independent of competitors. Position tells you whether you show up as the first recommendation or a footnote at the end. A brand can have a high presence rate and still be poorly positioned. Only share, presence and position together give you an honest read on where you stand in AI answers.

Record your measurements at equal intervals, monthly is a reasonable cadence. A single measurement is a snapshot; the trend is the real information. If your share climbs from 12 to 19 percent over three months after a content push, that's solid evidence the push worked. If it drops even though you changed nothing, that's usually new competitor content or a model update. You only catch either one by measuring regularly, the same way each time.

Why models name you in the first place

Language models name brands from two sources. First, training knowledge: whatever appeared often, consistently and in a credible context on the open web gets anchored in the model. Second, live sources that source-based systems retrieve at query time. For your share of voice, that means working both fronts. Marketing copy on your own site isn't enough on its own; what actually moves the needle is how often, and in what context, other people write about you.

Frequency and consistency here often beat the polish of any single page. Ahrefs' study of roughly 75,000 brands found that how often a brand gets mentioned across the web correlates with AI citation rate at about 0.664, roughly three times stronger than the correlation with backlinks, at about 0.218. If your brand keeps showing up linked to a clear topic in expert articles, industry directories, test reports, forums and comparison lists, the model learns that connection. A plumbing wholesaler named as reliable for spare parts across dozens of tradesperson forums shows up in answers on exactly that topic. A single polished landing page doesn't get you there on its own.

Thematic focus matters. Models link brands to specific topics, not to everything at once. Try to be present for twenty topics simultaneously and you'll be named clearly for none of them. It's more effective to become the obvious answer in one well-defined area than to show up weakly everywhere. So ask yourself first: for which three topics should a model reflexively think of my brand?

Content that AI systems actually cite

Source-based systems favor content they can lift out and drop into an answer with minimal editing. Concretely: clear questions as headings, a direct answer in the first sentence, definitions, numbers with context, and clean paragraphs instead of nested marketing copy. A model assembling an answer reaches for passages that are already almost fully formed. Prepare your core information to be citable and you make yourself an easy source to pull from.

Structure helps the machine parse your page. A sensible heading hierarchy, FAQ blocks, comparison tables and, where it fits, structured data all make content easier to lift cleanly. That said, Google has said explicitly that no special markup or AI-specific schema is required for AI Overviews or AI Mode, so structure earns you clarity, not a shortcut. A software provider who states prices, limits and target audience in a plain table gets cited correctly far more often than one who buries the same information in marketing prose. Concrete, honest detail beats superlatives here, because models work with facts, not adjectives.

• Answer the core question in the first sentence of a section, not three sentences of preamble later.
• Use real questions as subheadings, phrased the way users actually ask them.
• Back up claims with concrete numbers, timeframes and examples instead of marketing adjectives.
• Build comparison tables for prices, features or use cases.
• Keep definitions short, self-contained and understandable out of context.

Building mentions across the web on purpose

The biggest lever sits outside your own website. Since models learn heavily from external mentions, work systematically at appearing in relevant sources: expert articles and interviews, industry directories and comparison portals, and real discussion in professional forums and communities. An outdoor-clothing outfitter gains far more from being named in gear reviews and hiking forums than from another self-description in its own shop.

Consistency decides how much this compounds. Make sure your name, topic area and core message read the same everywhere. If a consultancy shows up sometimes as "process consulting", sometimes as "efficiency coaching" and sometimes as "transformation partner", the signal fragments and a model can't form a clear association. Uniform language across sources sharpens the link between your brand and its topic, and raises your share of mentions over time.

Favor real, verifiable substance over sheer volume. Bought mass mentions with no substantive value get filtered out by sources increasingly and barely anchor in model knowledge. What holds up is original data, studies, practical guides or case examples that other people pick up voluntarily. Give the industry something worth citing and you get cited, and every citation is a building block for your share of voice in future AI answers.

Common mistakes, and an honest look at the limits

A common mistake is treating AI visibility like classic SEO, or worse, like a checklist of new AI-specific files. Publishing an llms.txt file is a popular example: Google's John Mueller has confirmed no Google Search system reads it, and an Ahrefs analysis of roughly 137,000 sites that published one found about 97% saw no measurable referral traffic from it. Keyword density, backlink counts and meta tags only apply here in part; what matters more is clarity, topical authority and how often other sites mention you. A second mistake is the single measurement: ask once and feel good or bad about the result, and you've measured noise, not a trend. Only repeated, standardized measurements tell you anything reliable about your actual share.

Be honest about the limits of the method too. You're measuring a moving target: models get updated, answers fluctuate, and the same question can land differently depending on phrasing. Accuracy itself is imperfect — Columbia Journalism Review tested eight AI search tools on identifying the source of a given article and found more than 60% of answers were wrong across the board, with ChatGPT misidentifying 134 of 200. Claiming an exact percentage to the decimal point would be dishonest. What's useful is a range plus a direction. Treat your share of voice as a compass, not a speedometer, and act on patterns that hold across several measurements, not on any single one.

Finally: share of voice isn't an end in itself. It matters because AI answers increasingly shape purchasing decisions before people ever reach your site. So connect the metric to outcomes, for example inquiries that open with "ChatGPT recommended you". That keeps the measurement honest. Its job is to make you visible in the right places, not to give you a nice number to admire in a report.