Tracking Your Brand Across ChatGPT, Claude, and Perplexity

A working guide to measuring AI brand mentions, citations, and share of voice when there are no impressions, no rankings, and no logs to pull.

Updated on: 2026-06-15

A client asked me last month why their inbound demos suddenly mentioned three competitors by name in the first sales call. Not in a comparison context. In the "we were told to look at you and these others" context. None of those competitors outranked them on Google. Two of them barely had a blog. But when we ran the same buyer prompts through ChatGPT, Claude, and Perplexity, the pattern was obvious: our client was getting named in maybe 18 percent of relevant answers. The top competitor was at 64 percent.

That gap is the thing nobody has a dashboard for yet. And it is the thing this article is about.

There is no impressions log for an LLM. No Search Console. No "your brand was shown 14,200 times this week." The model generates an answer, the user reads it, the conversation closes. Whatever happened in that exchange is invisible to you unless you went and looked.

So tracking AI brand citations is fundamentally a sampling problem, not a measurement problem. You build a representative set of prompts your buyers would actually ask, you run them across the assistants that matter, you capture the full answer plus any citations, and you measure what came back. Then you do it again next week. The pattern over time is the signal. Any single answer is noise.

This is closer to how TV ratings used to work with panels than how Google Analytics works. The Signal AI team has been making this point for a while, and their argument for treating AI citations as a panel-based metric holds up well in practice. You are estimating, not counting.

Once you accept that, the rest becomes design work.

People conflate these constantly, and it causes real confusion in reporting.

Brand mentions in the generated text. The model writes a paragraph recommending tools, and your brand name appears in it. This is what most executives mean when they ask "are we showing up in ChatGPT." It is about being named, ranked, described, and positioned inside the answer itself.

Source citations and links. The model points to a URL as evidence. Sometimes that URL is yours. Sometimes it is a review site, a Reddit thread, a comparison article, or a press piece that talks about you. Perplexity surfaces these prominently. ChatGPT shows them when browsing is active. Claude cites in some modes and not others.

These two metrics diverge more than you would expect. A model can recommend your product enthusiastically and cite none of your pages. Or it can link to your documentation while describing a competitor as the better choice. You need to track both, and you need to report them separately, because the fixes are different. Mention problems are about brand presence in third-party content the model trusts. Citation problems are usually about your own site structure.

Conductor's framing of mention and citation tracking as two layers is one of the cleaner explanations I have seen, and it lines up with what we keep finding in audits.

This is where most tracking efforts quietly fail. People pick 20 obvious prompts, run them weekly, and watch a flat line. The flat line is not the truth. It is the consequence of a thin prompt set.

A useful prompt universe usually has four layers:

One more thing on prompt design: include local and language variations if your buyers are not all in one market. "Best email marketing tool" in English and the same prompt in Spanish or German will pull different brand lists. If you sell in those markets and you are only tracking English, you are missing half the picture.

After running this kind of tracking for a while, a few metrics consistently survive the "what does this actually tell us" test:

This is the part most tracking conversations skip, and it is the part that determines whether your numbers will move.

LLMs are not pulling from your homepage. They are pulling from a mix of:

This is why "just write more blog posts" is not a strategy. You need to be present and well-described in the sources the model already trusts, and you need your own content structured clearly enough that when it is pulled in, the model can extract the right facts. Ahrefs has a good piece on monitoring and winning brand mentions in AI answers that gets into the earned-side of this.

If you are trying to stand this up without buying anything yet, here is roughly what works:

For agencies, the white-label reporting matters more than the tracking itself. Clients do not pay you to send them screenshots. They pay you for a branded monthly report that shows their AI share of voice moving up and a competitor's moving down. That is the artifact that justifies the retainer.

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

Spend two days writing a prompt set. Make it specific to your buyers, not generic. Run those prompts manually across ChatGPT, Claude, and Perplexity. Read the answers. Note where you are mentioned, where you are not, who is mentioned instead, and what sources the models are citing. This will tell you more about your real positioning than any dashboard, because you will see the language the models use about your category.

Then decide whether you have a presence problem (you are not mentioned), a description problem (you are mentioned but framed weakly), or a citation problem (your content is not being pulled). The fix is different for each. Buying tracking software before you know which problem you have is how budgets get wasted.

After that, automate. Because doing this manually every week is the kind of work that quietly stops happening by month three.

How often should I run AI visibility tracking?

Weekly is enough for most brands. Daily is mostly noise unless you are actively running a campaign or watching a launch. The models do not change their answers minute to minute, and your prompt set is a sample, so over-sampling does not buy you precision.

Can I just look at my Google Analytics referral traffic from ChatGPT?

You can, and you should, but it only captures the people who clicked through. Most AI answer consumption ends in the answer itself. The user never visits your site. Referral traffic massively understates your exposure, which is part of why dedicated tracking exists.

Do AI mentions actually drive pipeline?

In the audits we have run, yes, but indirectly. Buyers arrive at sales calls already having a shortlist that came from an AI conversation. They do not always tell you that is where they got the names. The honest read is that AI mentions are upper-funnel and hard to attribute cleanly, similar to brand PR. Treat them that way in reporting and you will be right more often than not.

Is this just SEO with a new name?

No, and the people insisting it is have not run the prompts. The optimization targets are different (answer inclusion, not ranking), the trust signals are different (third-party citations and entity clarity matter more), and the measurement is fundamentally sampling-based rather than log-based. It rhymes with SEO. It is not the same instrument.