The citation metric every business is reading wrong about AI visibility

TL;DR

• When AI describes your brand, your own pages get cited 16% of the time. When it recommends brands in your category, just 0.9%; we measured an 18x drop across the same brands in the same week.
• Per engine it’s starker: on ChatGPT your site’s citation share falls from 40% to 7% between the two prompt types. On Perplexity and Claude it drops to zero.
• The reason why one number misleads: describe-prompts are won on your own pages (homepage, About, FAQ); recommend-prompts only on third-party sources (listicles, reviews, press). This means different strategies with different budgets.
• The fix is to stop trusting the blended number and split it by engine and by prompt type. If your dashboard can’t show that split, you’re paying for the chart, not the measurement.

A couple of weeks ago Camilla Wirth and I were staring at the same graph for a while and wondering whether something was wrong with our data extractor. We couldn’t spot an error but found a behavior pattern in AI engines’ responses that became worth sharing.

The metric we were looking at was the frequency of brand citations from brand’s own domain in response to questions about it. In other words, this is the metric that every “AI visibility” tool sells to you in the form of a single number on every dashboard.

Blended 16% citation average didn’t come as a surprise to us. The fact that, when split by prompt type, the very same metric turns into 16% in one direction and 0.9% in the other. That’s a 18x gap, concealed inside the single citation rate reported by AEO/GEO dashboards.

And why does it matter for businesses willing to show up in AI search results? Because these two parts of the gap require totally different content strategy, and the blended number is simply unable to tell you where the issue lies.

Brand-owned content is 16% of AI citations when AI describes you. 0.9% when it recommends you.

This is our main result from a 14,140 prompt responses test we ran across a set of twelve brands, with the same five AI engines in the same week. Measured only on responses where the brand was mentioned.

The difference is obvious after you see it. If a person asks AI about your brand, the most authoritative resource on “who does it represent” is your homepage and your About page.

What other sources could the model refer to?

If a person asks AI a recommendation about your brand’s category, your website becomes absolutely useless, with zero impact. AI doesn’t take it into account at all because of the very nature of the prompt: your homepage makes a case for you. Your competitor’s homepage makes a case for them. (Yes, it is perfectly okay that your homepage does make a case for you. This is its purpose. The problem is your site simply cannot perform this particular task.)

So, AI is forced to find third-party sources: shopping listicles, comparative articles, media coverage, reviews left by users.

Different question. Different sources. And your dashboard probably gives you no idea of that distinction.

The situation is even worse per AI engine

Here’s the break down by engine:

First time I got this view I thought there was something wrong with our data extractor. Two engines literally return zero on recommendation citation share. I reran the test three times before believing my eyes.

ChatGPT is the only engine that still mentions brand sites when recommending. Even there, the citation share drops from 40% to 7%. Perplexity and Claude cite your site with enthusiasm when describing it and the minute the user asks for recommendations, your domain becomes zero. Not just “almost zero”, but entirely zero. Zero of 4,595 recommendation citations across both engines during our test.

So, what happens if your dashboard shows you the blended number?

You are averaging apples and oranges. Two out of five major engines virtually never mention brand-owned content in recommendation prompts. The blended number can conceal that truth for months.

The action: split your AI search analytics by prompt intent AND by engine

OK, now this is what really bugs me about the way AI visibility tools are being sold right now. The standard report is one number. “Citation rate: 16%.” You log in. You see green. You assume everything is fine with your AI search strategy.

It may not be. You don’t know where the problem lies. And neither does your dashboard.

You need a 2×5 grid. Five engines down the side, two prompt types across the top. Ten numbers, not one. Anything less is the AI search equivalent of reporting email open rates without splitting by inbox provider. Technically a number, completely useless for figuring out what to do next.

Challenge your tooling. If it shows you one number, ask for the splits. By engine, by prompt type. If the dashboard cannot render that view, you’re paying for the chart, not the measurement.

After you get the splits, the further action becomes clear.

Recognition prompts (where your buyer has already typed your brand name into AI) are won on your own pages. Homepage, About, FAQ, Wikipedia. These should be the most concise and extractable descriptions of your brand. Worth direct investment.

Recommendation prompts require another approach. The buyer didn’t type your brand, they asked for your category. AI is comparing brands using third-party sources, full stop. No exceptions in our data. Shopping listicles, comparative articles, media coverage, reviews from people who have actually tried your product. Your site cannot compete here. Another game, with another budget: Digital PR, partnerships, listicle placements, the kind of work that won’t appear in your CMS.

These tasks are not the same. A “16% citation share” number that combines them into one does nothing for you. Worse, it can give you false confidence in the part of the funnel that works fine while the other part remains untouched.

You can find the full case study’s white paper here.

Summary

What you now know: the blended citation rate reported as a single number by most “AI visibility” tools is an average of two entirely different metrics. Brand-owned content is cited 16% of the time when AI is asked to describe a brand and 0.9% of the time when it is asked to recommend brands in a category, an 18x gap.

About friction AI

We built friction AI to measure how AI engines recognize, surface, and recommend brands across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overview. The platform publishes original research, like the 14,140-response study behind this article, on how AI search actually decides which brands to cite. Built by Joao da Silva and Camilla Wirth.