How to measure AI visibility.

Several different things, and blending them is where most reports go wrong. When an AI engine answers a buyer, your company can be mentioned, cited as a source, compared with rivals or recommended, and each can happen without the others. An answer can cite your page and recommend your competitor.

The IAB’s framework splits it the same way, into four metric groups: presence (mention rate, citation rate, share of voice), prominence (position), portrayal (sentiment, framing, accuracy) and persuasion (recommendation strength, clicks after a citation). It treats any composite score as a presentation layer, and says the numbers underneath must always be accessible.

So keep a scorecard, not a score. Track six rows, each with its own numerator and denominator: mentions, citations, comparisons, recommendations, traffic and business outcomes. Then a change points to a cause instead of a mood.

The ones buyers actually ask, in their words, frozen in place. Harvest them from sales calls, support tickets, reviews and community threads, then group them by the decision each one settles: problem, category, fit, comparison and implementation. Leave your brand name out of most of them, or you’re measuring recall of a name you supplied.

Keep the constraints buyers mention, like budget, company size or a must-have integration. In real conversations, a detail stated earlier is often missing from the final message, so a bare question tests less than the buyer asked.

Then freeze the set, and version it rather than editing it. About thirty questions is a practical start, but be honest about what that is: the IAB’s framework treats fewer than 50 queries as exploratory, and wants informational, comparison, recommendation and transactional questions all covered before a result counts as decision-grade. Start small to find direction, and grow the set before a budget rides on it.

More than one, and more than one kind. AI answers change from run to run, so a single screenshot is an anecdote. In a 2026 variance-components study of AI brand answers, one answer from one model, in one language and one phrasing, ranked brands with a reliability of about 0.01: next to no signal.

The same study shows where precision comes from. Repeating the identical prompt helps least: adding a sixth to tenth repeat to a five-repeat design barely moved the error. Varying what you ask and where you ask it helps more. So that’s our rule: two or three phrasings of each question, more than one engine, every week, with rates read from the pooled answers.

Then read every change against your own spread. Run-to-run noise, retrieval churn and a real model update look identical from a single check: a model change moves many questions on one date and stays, while noise moves one question and reverts. When an engine ships a major model change, the IAB’s advice is to reset the baseline and report the before and after as separate series.

Log them separately, with their sources. For each run, record whether you were mentioned, whether a page of yours was cited, whether you were recommended, and every URL the answer leaned on. A spreadsheet protocol works: a fixed question set, repeated runs per engine and one row per answer.

Share of voice needs a definition before it needs a dashboard. Appearance rate is the share of runs that name you. Share of voice is your appearances divided by every tracked brand’s appearances in the same runs. Report both per question and per engine, with run counts, because a blended number can’t tell you where you lost.

Then read the sources. In Profound’s analysis of 11.84 billion citations, about 57% pointed to company websites, yours and your competitors’, and the rest to editorial coverage and platforms like Reddit and YouTube. A citation gap becomes a content brief only when a page you own could fill it. The rest is work for reviews, partners or the product.

In two places, and neither is complete. For visitors, GA4’s default channel group has an AI Assistant channel for referrals from sources like ChatGPT, Gemini, DeepSeek, Copilot and Grok, while Google’s AI Overviews and AI Mode count as Organic Search. Visits that arrive without a referrer still land in Direct, so add a channel for other assistant domains and treat the total as a floor.

For machines, use your server or CDN logs. Crawlers and agents don’t run your analytics script, so count them in the logs, verify the bots, and split them by purpose: indexing, live retrieval and training. Keep the scale in proportion: Cloudflare reports 52% of crawler requests are now for AI training, and a crawl is not an audience. A fetch proves access, not that anyone saw an answer.

When you compare conversion, put rates beside their raw counts. AI-referred visitors often convert well because the assistant qualified them first, but the volume is thin, and a rate built on forty sessions is a guess.

It can, but measuring visibility can’t prove it did. Treat revenue as a chain of separate records: an answer that names you, a visit, a sign-up or sales conversation, activation, qualification and revenue. Each link has its own system of record, and you join them one at a time instead of multiplying them into a score.

Label every conclusion by what supports it: observed, attributed, inferred or causal. Attribution assigns credit under a rule you chose; only a holdout or an experiment earns “caused”. Add a free-text “How did you hear about us?” field to your forms, because much of this research happens where no tracker can see.

Keep context apart from demand. Reports that more ChatGPT conversations are commercial describe activity inside a vendor’s sample, not buyers or pipeline. For an ROI view that doesn’t invent a price per citation, measure the chain link by link and weigh what you spend against the questions where you hold a measured presence.

Ask what sits under them. Every visibility number rests on a sample, a denominator, a counting rule and a mix of engines, and a benchmark is only as good as the methods behind it. Before you trust a figure, ask:

• How many questions, runs and engines produced it, and over which dates?
• Is it reported per engine, with its spread, or blended into one number?
• Can you see the underlying answers and citations, and export them?
• Does it count mentions, citations or recommendations, and does it say which?

The IAB asks the same of providers: disclose responses per query, show results per platform and how much they vary, and document how any composite is weighted. Hold research to that standard too. The study behind most GEO claims is often quoted from the one table that has no error bars. Our AI search statistics hold every figure to that standard, and list the popular numbers that fail it.