We Analyzed 1,000 AI Answers. Here’s What We Found About Brand Citations.

How We Ran the Study

The unit of the headline is 1,000 governed question runs: 20 brands × 50 prompts (10 branded, 40 non-branded). Running that set across three platforms produced roughly 3,000 platform-level answer records, which is the number in the downloadable dataset.

• Prompts were stratified across informational, commercial, comparative and navigational intent, then frozen for the cycle.
• Each prompt ran three times per platform, with web search on and personalization off; we coded the median answer.
• Every answer was scored for mention (named or not), citation (source URL), absorption depth (0–2) and the rank of the first brand mention.

The full rules are in the measurement methodology, and the raw rows are free on the Data page.

Each answer was coded independently by two reviewers, with disagreements arbitrated against the written rules; inter-reviewer agreement stayed above 90%. We logged the model version, date, region and settings for every run so the snapshot can be recreated later.

Finding 1: Citations Concentrate on a Few Sources

Citation supply is not evenly distributed. Perplexity casts the widest net, while ChatGPT is far more selective — and that selectivity changes what “winning” looks like on each engine.

This mirrors the peer-reviewed study From Citation Selection to Citation Absorption (arXiv:2604.25707), which found across 602 prompts and 21,143 citations that Perplexity and Google cite the most sources per prompt while ChatGPT cites fewer. In our own 20-brand sample, the same pattern held: a long tail of sources appeared once or twice on Perplexity, while a short, recurring set of high-authority domains dominated ChatGPT answers.

Practical takeaway: on breadth-first engines, coverage and modular content get you listed; on selective engines, authority and trust decide whether you are listed at all.

Finding 2: Mention, Citation and Absorption Tell Different Stories

The most common mistake is treating “we showed up” as a single metric. In our data, the three layers separated constantly.

• Mention without citation was common on ChatGPT: brands were named from memory with no link — good for awareness, unverifiable for the reader.
• Citation without absorption was common on Perplexity: a brand’s link sat in a stack of sources while the answer text drew nothing distinctive from the page.
• Deep absorption — where the answer repeated a brand’s specific number, definition or procedure — was rare everywhere but disproportionately drove which brand was named first.

The arXiv authors put it bluntly: high-influence pages are longer, more structured, semantically aligned with the question, and richer in extractable evidence. A link does not buy that; evidence does.

Finding 3: Non-Branded Prompts Are Where Brands Disappear

The widest gap in our sample was not between platforms but between prompt types. Most brands scored reasonably when a prompt contained their name and poorly when it did not.

• On branded prompts, the AI usually recognized the brand; the failures were outdated facts or confused positioning.
• On non-branded category prompts, the same brands were frequently absent altogether, with the answer recommending one or two better-evidenced competitors.

That gap is the discovery problem in one number: the AI knows who you are when asked, but never introduces you to someone who has not already heard of you.

Being named first beats being named at all

In answers that compared several options, the first brand named received the bulk of the descriptive language and the most specific claims, while later mentions were often one-word list entries with no supporting detail. Recording the rank of the first mention — rather than a simple yes/no — turned out to be one of the most predictive fields in the entire dataset, and the one teams most often overlook.

How to Read These Findings

A note on scope. This is a snapshot of 20 brands at a point in time, not a census; model versions and indexes shift weekly, and the sample leans toward B2B and consumer software. We report descriptive patterns, not causal claims — publishing a page that resembles a high-influence page does not guarantee absorption. Treat these findings as hypotheses to test on your own category, which is exactly why the prompts and raw data are public.

What This Means for Your Brand

Three actions follow directly from the data:

1. Report the three layers separately. A single “visibility” number hides whether your problem is awareness, citation or influence.
2. Fix non-branded first. Build answer-first pages for the category questions buyers ask before they know your name.
3. Optimize for absorption, not just links. Put definitions, statistics, comparisons and step-by-step procedures where an engine can extract and repeat them.

You can reproduce every chart with the published dataset and the standard’s scoring sheet.

Frequently Asked Questions