Citation share is not a separate study for us. It falls straight out of the same instrument we use to grade the nine trackers on the bench: a fixed 150-prompt panel, fired at ChatGPT, Perplexity, Google AI Overviews, Gemini, and Copilot every 48 hours, with every raw answer logged and diffed against the prior run. Grading a tracker asks whether it reported that run correctly. This page asks a different question of the identical data: which domains the assistants actually cited.
What we count as a citation
A citation is any domain an assistant surfaces as a source, link, or reference inside an answer, whether through a structured citation list (where the engine emits one) or a URL detected in the answer body. We deduplicate at the root-domain level, so a vendor’s blog subdomain and its marketing site count as one citation, not two. Because Google AI Overviews renders inside a search results page rather than a chat window, we read its citations off the rendered block itself.
The current reading
These numbers come off our own 150-prompt panel, not a market-wide crawl, and update on the same 48-hour cadence as the grade board above. Treat them as directional, tied to our panel’s mix of prompts and verticals, not as a census of the open web.
- Reference and encyclopedic sources lead citation share on informational prompts in our panel, though their share has been easing as assistants pull in more recent, narrower sources.
- Large trade publishers and review sites hold the next tier, especially on comparison-style prompts (“X vs Y”) where a buyer is weighing named options.
- First-party vendor domains rarely break through on informational prompts but show up far more often once a prompt turns branded or transactional, meaning the buyer already has a shortlist in mind.
- Forum and community threads appear more often on narrower, more technical prompts than on broad category questions, and their share has grown across our reruns this year.
Why we track this at all
Every tracker on our bench claims to help a brand understand its presence in assistant answers. Citation share is the ground truth we check that claim against: a tracker that reports strong “mentions” but never surfaces which domains actually get cited is answering half the question. The measurement protocol documents how citation depth factors into each tracker’s grade, and the grade board is where that grading actually lives.
What we have not verified
This reading comes from one panel, one region, in English. We have not checked whether citation share looks the same across other languages or geographies, and a 150-prompt panel is not large enough to make claims about long-tail categories outside the verticals we test. Where a pattern held across every rerun in our test window we say so above; where it moved between runs we call it a lean, not a finding.
Every figure here is checked against the most recent completed run, and when a reading no longer matches, the page is corrected within two weeks.