How to Measure AI visibility

A practical guide to tracking brand presence, share of voice, and citations across ChatGPT, Claude, Perplexity, and Google AI Mode without drowning in vanity metrics.

Updated on: 2026-05-21

The first time I ran an AI visibility audit for a B2B SaaS client, the dashboard told us we had a "76% visibility score." Sounded great. Then I actually read the prompt outputs. We were getting mentioned in answers about "best alternatives to [the actual market leader]" but never in the head-term prompts like "best [category] software for mid-market teams." The single score had averaged a strong result on weak prompts with a weak result on the prompts that actually drove pipeline.

That experience is most of what I want to talk about here. Measuring AI visibility is not hard because the metrics are exotic. It is hard because most marketing managers inherit a number, treat it like a rank tracker reading, and miss the texture underneath.

Strip away the marketing copy and AI visibility is three things stacked on top of each other:

Every credible framework I have seen, including the one Semrush uses in their AI search visibility reporting guidance, maps to some version of these three. If your reporting only answers one of them, you are flying blind on the other two.

One thing worth being honest about: every "AI visibility score" on the market is modeled, not measured. The team at Franco put it bluntly, and they are right. No tool has a feed of what real users are typing into ChatGPT. What tools do is run synthetic prompts on a schedule against LLM APIs and parse the outputs. That is useful and directionally accurate, but it is not panel data. Treat the numbers like a barometer, not a thermometer.

Here is the working set I use with clients. You do not need all of them on a weekly dashboard, but you should have a defensible answer for each.

Raw count of brand mentions and URL citations across AI answers over a defined prompt set, broken out by platform (ChatGPT, Claude, Perplexity).

Why it matters: this is your baseline. Before you can correlate AI visibility to branded search or conversions, you need a stable count to trend against. Brainlabs makes the same point and recommends establishing this before correlating to downstream metrics.

What to watch for: a rising mention count on a shrinking prompt set is not progress. Lock your prompt universe before you start trending.

Your citations divided by total citations across the same prompt set, expressed as a percentage. Tracked overall, by topic cluster, by platform, and ideally by intent (informational vs. commercial).

The formula is straightforward: (Your Citations ÷ Total Citations) × 100.

This is the metric I argue about most with marketing managers. Share of voice forces you to define your competitive set honestly. If you include three brands that nobody in your category seriously evaluates, your share looks artificially high. If you leave out a real competitor, you are lying to yourself.

For each prompt in your tracking set, which brands appear, in what order, and with what framing?

This is where the work actually happens. A 40% share of voice across 50 prompts is interesting. Knowing that you own 80% of share on "best tools for X" but 0% on "alternatives to [competitor]" is actionable. The first you defend. The second you build content for.

This is also where platforms like SEOforGPT earn their keep, because manually re-running 50 prompts a week across four AI assistants is not a job, it is a punishment. Tracking 25 to 100 prompts on a regular cadence is the minimum viable setup for most B2B brands I work with.

Google-specific, and increasingly the metric your CFO will ask about because it most closely resembles the SEO numbers they already understand.

Track: what percentage of your target keywords trigger an AI Overview, and within those, what percentage cite your domain. SE Ranking's AI visibility tracker and Semrush both surface this, and the methodology is similar across tools.

Not all mentions are equal. A definitive recommendation in the first paragraph of a ChatGPT answer is worth more than a footnote-style mention three competitors deep. Some teams score each mention (e.g., 3 for primary recommendation, 2 for supporting mention with link, 1 for passing reference). It is a little subjective. It is also more useful than a flat count.

How is your brand described? Expert and specific, or generic and substitutable? Associated with which pain points or use cases?

This one is qualitative and easy to skip. Do not skip it. I have seen brands with strong mention volume get systematically framed as "the budget option" or "best for beginners" when their actual positioning is enterprise. That framing problem will cap your pipeline regardless of how high your share of voice climbs.

Real sessions arriving from chat.openai.com, perplexity.ai, claude.ai, and similar sources, visible in GA4 as referrals. Compare engagement and conversion rate to organic search and direct.

This is the closest thing to bottom-of-funnel proof you currently get. Airops treats it as the clearest signal that AI systems are actually sending users to your site, not just talking about you. The catch: volumes are still small for most brands, and attribution is messier than classical search because users often paste, summarize, or screenshot rather than click through.

Branded search volume. Direct traffic. Returning users. Demo requests. Pipeline. Revenue.

You will not get clean causation here. You can get useful correlation. Overlay AI mention volume and share of voice against branded search trends over six to twelve weeks. If they move together, you have a leading indicator worth reporting. If they do not, you have a content quality problem upstream.

Here is the rough monthly view I recommend to marketing managers who are presenting to leadership for the first time:

Three layers. Each one answers a different question. Foundation says "are we present?" Diagnostic says "where do we win and lose, and how?" Outcome says "is this driving the business?"

If you are starting from zero, do not buy a tool on day one. Do this instead:

This is where I would point a marketing manager at SEOforGPT honestly. The platform handles the automated prompt re-running, share of voice tracking across ChatGPT, Claude, and Perplexity, and the part most tools skip: identifying which content gaps are causing you to lose specific prompts, then generating and auto-publishing structured content designed to be cited. For a marketing team without a dedicated AI visibility analyst, that compresses a multi-tool workflow into something one person can actually manage. The free Bootstrap tier is enough to run a first audit. The $99 Launch plan gets you 25 tracked prompts and weekly testing, which is roughly the minimum cadence I would recommend for a B2B brand serious about this.

The other thing worth knowing: agencies using the platform's white-label workspaces are reportedly selling AI visibility retainers at $2,000 to $5,000 per client per month. That gives you a rough market benchmark for what this work is worth when packaged well, which matters if you are an in-house manager trying to justify internal investment.

A few patterns I keep seeing:

Treating AI visibility as an SEO subtask. It is related but not the same. SEO optimizes for ranked results. AI visibility optimizes for being cited inside generative answers. The content patterns that win citations (clear definitions, structured comparisons, original data, named expert authorship) overlap with strong SEO content but are not identical. Searchengineland's overview covers the distinction well.

Obsessing over one platform. ChatGPT gets the headlines. Perplexity sends a surprising amount of high-intent traffic for its size. Google AI Mode is going to matter most for brands with strong existing SEO. Claude is underrated for B2B technical buyers. Track all four. Weight them by where your buyers actually are.

Reporting visibility without tying it to revenue. A 12% share of voice gain is a slide. A 12% share of voice gain correlated with a 9% lift in branded search and three closed deals attributed to AI referral is a story. Build toward the story.

Confusing more content with better content. AI assistants disproportionately cite pages with strong E-E-A-T signals: named experts, original data, methodology disclosed, depth on the actual topic. Publishing 30 thin posts a month will not move your citation share. Publishing 5 substantive ones with clear authorship usually will.

How often should we measure AI visibility? Weekly is the sensible cadence for tracking. Monthly is the right cadence for reporting to leadership. Daily is overkill and will make you chase noise. AI answers vary run-to-run, so any tool that promises "real-time" precision is selling you something the underlying systems cannot deliver.

Do we need a separate budget line for AI visibility, or is it part of SEO? Honestly, both answers are defensible. What matters more is having a named owner. The teams that get traction here have one person whose job description explicitly includes AI visibility metrics. The teams that struggle have it as the seventh priority of someone running paid and SEO and content.

Is there a "good" share of voice number? Not really, and anyone who quotes you a benchmark without knowing your category is guessing. In a fragmented category, 15% might be category-leading. In a category with two dominant brands, 15% might be a distant third. Track your own trend and your gap to the leader.

What if our brand barely appears at all right now? Then your first job is not optimization, it is presence. Audit which prompts in your category trigger answers at all, identify which sources those answers cite, and figure out why those sources are trusted. Usually it is some combination of domain authority, structured content, original data, and being mentioned by other sources the AI systems already trust. Build toward that.

Can we do this without a dedicated tool? For the first 30 days, yes. Past that, no. The volume of prompts you need to track across four platforms on a weekly basis is not something a human will sustain, and the manual approach loses to automation on cost within a month.