AI Visibility: How to Measure Whether AI Actually Recommends You

AI visibility is how often AI engines name your brand in their answers, and the number that matters is whether they recommend you unprompted, not whether they recognize you by name. Measure it per engine, across repeated runs, over weeks. There is no single trustworthy "AI visibility score," and the gap between being known and being named is where most brands quietly lose.

Most teams measure the wrong thing. They ask ChatGPT about their own company, get a flattering paragraph back, and conclude they are visible. That is recognition, and it is close to meaningless for acquisition. The question that moves revenue is the one a buyer actually types: "what's the best X for Y," with your name nowhere in it. This guide covers what to measure, the three ways to measure it, and how to read the result without fooling yourself. It is the measurement companion to the broader generative engine optimization playbook.

Recognition is not recommendation, and the gap is enormous

The single most useful study on this came from Victorious in Q2 2026. The team tested 175 brands across legal, healthcare, SaaS, financial services, and ecommerce, on eight platforms: ChatGPT, Claude, Gemini, Copilot, Perplexity, Google AI Overviews, Google AI Mode, and Meta AI. When asked about a brand directly, AI described it accurately 96% of the time. When asked a category question, the kind a buyer actually asks, 89% of those same brands never appeared at all.

Read that again. Nine in ten brands that AI can describe perfectly are invisible the moment a customer asks for a recommendation instead of a definition. Victorious also found a threshold: brands with fewer than 2,000 indexed web pages mentioning them were named in category answers just 3% of the time. And in those category prompts, 99.99% of the 49,391 citations pointed to third-party sites, not the brand's own domain. Visibility is earned off your property, which is why it tracks brand mentions more than backlinks.

So the first rule of measurement: test recommendation prompts, not vanity prompts. If the only question you run is "tell me about [your brand]," you are measuring the 96% that almost everyone passes, not the 11% that decides whether you get the customer.

Why one number lies: engines barely agree with each other

The instinct is to want a single score, one dial labelled "AI visibility" that goes up or down. Resist it. The engines cite almost entirely different sources, so a composite number averages away the only signal that matters.

Wellows analyzed 22.7 million citations from January to June 2026 across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode, covering 1.15 million questions. The fragmentation is the headline: 79.6% of the websites cited on a question appeared on only one of the five engines, and just 0.31% of sources reached all five. On the same question, Perplexity never touches 89.1% of what ChatGPT cites. Even Google's own two surfaces, AI Overviews and AI Mode, agree on only 24.5% of cited sites, and that is the closest pair in the study. Across companies, the average overlap is 7.3%.

What that means for measurement is concrete. Winning in Perplexity tells you close to nothing about ChatGPT. A rollup score that blends them hides the fact that you might own one engine and be absent from the next. Track each engine as its own scoreboard, and weight them by where your buyers actually ask.

The three ways to measure AI visibility

Every tool and method falls into one of three families. They watch different things and miss different things, so the honest answer is to run more than one.

Prompt-based trackers (Profound, Peec, Otterly, AthenaHQ and the rest of the AI SEO tools built for this) send a defined set of prompts to each engine on a schedule and record which brands and domains appear. They are the most direct way to measure recommendation across engines. The catch is sampling: they test the prompts you choose, and any single run is a snapshot of a system that picks its sources differently each time.

Network-layer measurement is the newer approach. Cloudflare added an AEO Visibility Dashboard to its AEO Suite on August 6, 2026, scoring how often AI assistants cite a site by observing real crawl and referral activity across its network, not by sampling prompts. It reports Citation Rate, Mention Rate, Prominence, Share of Voice, Industry Fit, and AI operator activity. Two honest limits: it is early access by request with no published price, and it covers only Claude and GPT, so Gemini, Perplexity, Copilot, and Grok sit outside it.

Native platform reports are free and often overlooked. Google launched generative AI performance reports in Search Console on June 3, 2026, isolating impressions in AI Overviews and AI Mode. As of September 2026 those two are combined into one view, and AI Mode rows can be filtered by user prompt, the first query-shaped signal Google has exposed. The limit is sharp: impressions only, no clicks, CTR, or full query data. On the Microsoft side, Bing Webmaster Tools added Citation Share, Intents, Topics, and Compare to its AI Performance report on June 16, 2026, all free and in preview. Citation Share is the standout, showing your percentage of citations on a query rather than a raw count.

Measure recommendation, prominence, and share, not just presence

"Am I cited, yes or no" is too blunt. Once you are tracking the right prompts across the right engines, log four things per engine:

• Recognition rate. Does the engine describe you correctly when asked by name? This is your floor. If you fail here, fix accuracy first.
• Recommendation rate. How often do you appear in answers to category and comparison prompts where you were not named? This is the number that maps to pipeline.
• Prominence. Where in the answer do you land? First named, in a comparison table, or a footnote citation? Position changes whether a reader ever sees you.
• Share of voice. Of the brands named on a prompt, what fraction are you? Bing's Citation Share metric measures exactly this for its own surface, and it is the cleanest way to track gains against competitors rather than in a vacuum.

Presence alone flatters you. A brand cited last, in a six-item list, on one of five engines, is "visible" by a yes/no tally and nearly invisible to a buyer.

Account for volatility before you call a change a win

AI citations churn hard, and this is where most measurement goes wrong. Techmagnate tracked 83,633 citations over eight weeks of personal-loan queries across ChatGPT, Google AI Mode, and Perplexity and found an average of 39% of previously cited domains dropped out week to week (ChatGPT near 45%, Google AI Mode under 32%). Yet a small core held firm: roughly 330 consistent domains captured about 93% of all citations. Somantra's larger analysis of 2.4 million citation records on ChatGPT and Google over seven months found 57.2% of cited domains appeared in a single month and never again, with only 2.7% persisting across all seven.

The lesson is not that measurement is hopeless. It is that a single reading is noise. Three disciplines fix it:

1. Run each prompt several times. The source set shifts between identical runs, so one answer is one draw from a distribution.
2. Track over weeks, not days. A gain that survives a month of 39% churn is real. A gain that vanishes next Tuesday was variance.
3. Watch consistency, not just peaks. Appearing once is easy and worthless. The brands that win are the ones in the persistent core, which is the same lesson behind why building on a single platform is fragile.

State a date on every claim, too. Model defaults, retrieval pipelines, and tool coverage in this space change monthly, so a measurement is only true "as of" when you took it.

A measurement routine you can actually run

You do not need a budget to start. Here is a stack that works with free tools and an hour a week.

1. Write 15 to 25 real buyer prompts. Mostly category and comparison questions ("best X for Y," "X vs Z," "alternatives to W"), a few recognition prompts to track your floor. These are your fixed test set.
2. Run them across the engines your buyers use. ChatGPT and Perplexity in search mode, Google AI Mode, Gemini. Log, per engine, whether you appear, where, and who appears with you. Repeat each prompt two or three times.
3. Pull the native reports. Search Console's generative AI report for AI Overviews and AI Mode impressions, Bing's Citation Share for Copilot. Free, and they cover surfaces your manual runs cannot sample well.
4. Cross-check with referral traffic. AI engines send real visitors; watch for traffic from chatgpt.com, perplexity.ai, and the rest, which converts well above traditional organic. The full breakdown is in the AI referral traffic guide.
5. Repeat weekly and diff the trend. You are looking for recommendation rate and share of voice moving up and holding, per engine.

When you graduate to a paid tracker or request network-layer access, it automates steps 2 and 4 at a scale you cannot match by hand. The thinking stays the same.

Measuring AI visibility well is less about the tool and more about the question. Test what buyers ask, not what flatters you. Measure each engine separately, because they disagree 90% of the time. And give it weeks, because one good day in a system this volatile is not a win. Do that, and you will know whether AI recommends you, which is the only version of visibility that pays. Then spend the effort on getting cited in the first place.