What AI Answers Actually Cite: 21,508 Citations Analyzed

Every guide on getting recommended by AI tells you the same thing: get into the roundups. Earn placements in other people's "best tools for X" posts, because that is what the models read.

That advice is half right, and which half depends entirely on which engine your buyers use. Here is the data.

The corpus

VerifiedDR runs tracked buyer questions against ChatGPT, Perplexity and Google AI Mode on a schedule and keeps every source each answer cites. In August 2026 it classified the whole archive by page type: 16,322 distinct pages across 21,508 citation events.

The first finding is a counting trap worth understanding before you read any other number. Measured per distinct URL, the mix reads 42% homepage, because 1,884 of those citations point at a bare hostname and each one is a single URL. Weighted by citation events, which is what a reader actually encounters, it looks different:

So the roundup advice is not wrong. Listicles are the single most cited page type across the archive. But they are 17.1%, not a majority, and the average hides the thing that matters.

ChatGPT and Perplexity are mirror images

Split the same corpus by engine and the two biggest slices invert completely.

ChatGPT reaches for the brand's own homepage a third of the time and almost never cites a roundup. Perplexity does the opposite. These are not variations on a strategy. They are two different distribution channels that happen to share a product category.

What this changes for an agent product

If you sell an AI agent, you are in a category where the buyer's first move is increasingly to ask another AI which one to use. That makes the split above a budget decision, not a curiosity.

If your buyers live in ChatGPT, your own pages do the work. A third of citations go to a homepage, and another 12% to product pages. The money belongs in making those pages answer the question directly: what the agent does, what it costs, what it connects to, who it is for. Not in outreach.

If your buyers live in Perplexity, one citation in five is somebody else's roundup and your own homepage barely registers. Getting named in the directories and comparison posts of your category is the actual work, and it looks more like digital PR than like content marketing.

If you do not know which one your buyers use, find out before you spend anything. Ten questions asked manually in both engines, logged out, with the date recorded, will tell you more than a quarter of guessing.

The measurement problem underneath

Agent products have a specific version of this problem. Revenue claims are easy to make and hard to check, which is why AgentMRR exists. Being recommended by an AI answer has the same shape: it is easy to assert that a model likes your product, and hard to show it.

Both are fixed the same way, by keeping a fixed question set and a dated record rather than a single score. Report the share of unbranded buyer questions where you appear, the cited sources you do not own, and any factual error the model repeats about you. Resist building one blended visibility number: generation is sampled and engines personalize, so the number moves for reasons you cannot explain and you spend the next quarter defending noise.

If you want the recurring version of this rather than a manual check, VerifiedDR tracks it across those three engines with every answer shown verbatim and the cited pages listed, on the same profile as the backlink and trust data that decides whether you are citable at all.