Most teams working on AI visibility worry about being absent. In our data the sharper problem is who shows up instead. We looked at almost 100,000 source citations behind AI answers about a sample of brands, and when an assistant answers a question about a company, it leans heavily on that company’s rivals. Of the outside sources cited in answers about a brand, around 40 percent were the brand’s direct competitors, and for most brands the figure ran between a third and a half. When AI describes you, competitors are a large share of what it cites.
Sort the outside sources by whose site they sit on and the split is clear: competitors take around 40 percent, and other third parties (independent media, niche blogs, communities) the rest. We measure citations to a brand’s own pages separately, and they are a much smaller, distinct stream; the story here is the outside sources, because that is what the model reaches for when it talks about you. This is the companion to our engine-overlap study, which asked how the five engines differ from one another. This one asks who actually gets cited.
What we measured, and how small the sample is
We record the source URL behind every citation in the AI answers we track, for a sample of brands, across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode, over roughly three months this spring. That is almost 100,000 citations to outside sources, resolving to more than 10,000 distinct domains. We tagged each cited domain as one of the brand’s tracked competitors or an unrelated third party, using each brand’s own competitor set.
Be skeptical with us about scope. This is a small, mixed sample of brands spread across roughly a dozen categories from SaaS to retail to travel, tracked on the competitive questions that matter to them. It is enough to see a strong pattern, not enough to call it a law of the web. Where larger public studies exist, we compare against them below rather than pretend our slice is the whole picture.
Competitors own a large slice of your citation graph
Across the sample, around 40 percent of the outside sources cited in answers about a brand were that brand’s own tracked competitors. This was not one or two crowded categories dragging the average up. The typical brand sat a little over a third, the middle half ran from roughly a third to just over half, and nearly nine in ten brands had at least a quarter of their citations going to competitors. When an assistant fields a question about you, competitors are a large share of what it cites.
That is a different problem than the one most AI SEO advice addresses. The goal is usually framed as getting mentioned. But being mentioned in an answer that also cites four of your rivals is not the win it sounds like. The competitive surface is the citation list itself, and right now your rivals are winning a large share of it.
AI describes you from other people’s pages
The model is not reading your homepage to describe you. The sources it pulls into an answer about you sit almost entirely outside your own site, and a large share of them are your competitors. Larger studies point the same way on the owned-versus-third-party split: Profound, analyzing millions of citations, reports that roughly 80 percent of cited sources sit on third-party domains. Your own pages matter, but they are a minority of what shapes your AI presence, and you do not control the rest of it.
The wider map, and where our sample is too thin to trust
Two more patterns showed up, and both come with caveats worth stating plainly.
First, the citation graph in our sample is wide and flat: more than 10,000 domains, and about 43 percent of them cited exactly once. But that flatness is partly an artifact of a small sample spread across unrelated categories, where each category pulls in its own niche sources and nothing concentrates. The largest public analysis we know of, Profound’s look at millions of citations, finds the opposite on broad queries, with the top ten domains alone accounting for more than a third of all citations. Read our flatness as “no shared shortlist across these particular categories,” not “AI has no favorites.”
Second, source type. Among the recognizable platforms, video led: YouTube was the single most-cited named platform in our set, which lines up with 2026 reporting that YouTube has overtaken Reddit as the top social source for AI citations. The review and marketplace sites that B2B teams pour budget into were small in our data. But our figure is a share of all citations, including the long tail, while most published studies report share among each platform’s top sources, which is a different denominator and runs much higher. The mix also splits hard by engine: other analyses show Wikipedia dominating ChatGPT’s sources and Reddit dominating Perplexity’s. Our companion study covers that per-engine split.
What we could not conclude
We measured where citations point, not why any single page was chosen. We did not test page features like word count, schema, freshness, or named authorship against citation outcomes; that needs a controlled join we have not built yet. The Princeton GEO research found page-level effects like that in a controlled setting, and nothing here contradicts it, but we did not reproduce it. And the sample is small enough that the exact percentages will shift as we add brands. Treat the direction as solid and the decimals as provisional.
What to do with this
- Audit who AI cites about you, not just whether it cites you. The gap you can act on is the competitor pages winning citations you are not.
- Win the third-party pages, because that is where the model reads. Reviews, comparisons, trade coverage, and video are doing more for your rivals than their homepages are. The pages that win those citations share a structure, which we broke down in how to write content AI platforms cite.
- Treat your own site as necessary but not sufficient. It is a small, separate stream of citations. It cannot carry your AI presence alone.
- Track per source and per platform. The split differs by engine, so a single visibility number hides the move you can actually make. The end-to-end loop is in our GEO playbook.
How we ran it
The dataset is the citations to outside sources attached to the AI answers we track for a sample of brands across the five platforms named above, collected over about three months this spring, almost 100,000 citations in total. We resolved each citation to its domain and tagged it as a competitor or an unrelated third party using each brand’s own competitor set, then aggregated and checked the competitor share brand by brand so the headline is not riding on a couple of outliers. Citations to a brand’s own pages are tracked separately and are not in these figures. The companion engine-overlap study looks at the same corpus from the per-engine angle. We refresh this analysis on a regular cadence and keep this page updated.
