Which domains US AI answers actually cite

Published July 28, 2026 · CiteAngle · Measured July 27, 2026, United States market · 50 category and brand questions observed across all 15 answer surfaces in one full pass · External studies cited here are linked inline with their publication dates · Figures are logged in our public claims registry

What does the cited set look like? A short head and a very long tail

The run was a single full pass. We took 50 questions a US buyer might type or say when shopping for AI search work. Each went to all 15 answer surfaces on the same day, and we kept every source link in every answer. That came to 8,662 citations, pointing at 1,987 separate domains. Own measurement

Two numbers tell you more than any average. The ten most-cited domains hold 2,353 citations, 27.2% of the total. At the other end, 1,088 domains show up exactly once in the whole set — 54.8% of every domain the answers touched.

Ranks and counts come from our own run of 50 questions across 15 answer surfaces, United States, measured July 27, 2026. The type column is our grouping, not part of the data.

The first row needs an asterisk, and the asterisk is the interesting part. youtube.com drew 1,152 citations, 13.3% of the whole run. But 1,040 of those came from the two video answer surfaces, which return video by design. Drop those two surfaces and youtube.com falls to third, behind semrush.com and searchenginejournal.com. So read the table as proof that surface mix drives the ranking. It is not a content instruction.

Reach across surfaces is rarer than volume. Of the 1,987 domains, 1,390 showed up on only one of the fifteen. Two reached eleven: semrush.com and searchengineland.com. A domain that shows up in one AI answer tells you almost nothing about the next one.

How much of this sits on platforms you can post to?

Operators care about a narrower question than "what gets cited." They want to know how much of the cited ground is open to them. That means pages they can post today, without waiting for an editor. We counted citations to nine such platforms: youtube.com, linkedin.com, instagram.com, facebook.com, reddit.com, tiktok.com, x.com, quora.com and medium.com.

For this count we held out the five brand-name questions and used the 45 category questions. That way the number describes the category, not one company. Those 45 questions gave 8,228 citations, and the nine platforms took 1,631 of them, or 19.8%. Own measurement

That fifth is not spread evenly. Of the 1,631 platform citations, 1,308 came from the two video answer surfaces. On the remaining thirteen surfaces the nine platforms accounted for 323 of 6,908 citations, or 4.7%. In this category, posting to platforms is a video-surface play. The text answer surfaces were built almost entirely out of ordinary web pages.

Do all the surfaces cite at the same rate? Not remotely

Citations per surface run from 14 to 2,675 in the same run, on the same questions, on the same day. This is the part a single blended "AI visibility score" averages away.

Google News answers alone carried 2,675 citations, drawn from 635 separate domains. That is 30.9% of everything the run collected. The YouTube surface carried 836, all from a single domain. At the thin end, the Claude surface held 14 source links across its 50 answers. Read that as a point about units, not a league table. A citation on one surface is a different event from a citation on another, and a blended average of the two hides the one thing you can act on. Where each of these surfaces sits in the US market, and what Google and Bing report about them, is mapped in our US AI search surface map.

If you rank well, are you in the answer? Published evidence says only partly

This data breaks a habit — the belief that AI answers just restate the top of the organic results. seoClarity compared AI Mode citations against organic rankings for the same queries on US desktop. Only 19% of AI Mode citations came from the query's own top 20. The rest came from outside it (transactional queries, data from September 2025). Industry study seoclarity.net

Being cited is also not the same as being used. One study looked at 602 prompts covering 21,143 citations and 18,151 cited pages. How often a source was cited and how much it shaped the answer came apart. The most-cited source is not reliably the one shaping what the answer says (arXiv preprint, revision dated April 29, 2026; correlational, not causal). Preprint study arxiv.org Which content edits move citation and which do nothing measurable is the subject of our read of the controlled citation evidence.

Does a domain's rank in this table mean it is a good place to publish?

No, and the table should not be used that way. It counts where 15 answer surfaces pointed for one US question set on one day. A count like this reflects the surfaces we sampled as much as the merits of any domain. That is exactly why the video split matters. The useful read is the shape — a short head, a huge tail, and a per-surface mix that changes the answer to "where do we compete" more than any single domain name in the list.

How do you read your own citation ground?

This page shows the ground of one category. Your buyers ask different questions, and the answers they see cite different domains. That set is knowable. Knowing it is the difference between guessing at a content plan and pricing one. A client audit runs the same questions again and again, not once. Panorama covers 50 queries × 7 runs × 14 engines. That is a 4,900-cell grid, run in full, with every cell's status shown and reported across the 15 surfaces above. It comes with repeated runs and confidence intervals, so you can tell a real shift from noise. Why one pass is the right tool for describing ground and the wrong one for ranking brands is worked through in our note on single-run point estimates.

Every citation in the deliverable keeps its source URL, its surface, and the answer span it came from. So the domain list above becomes a domain list for your category, with the surfaces that made it attached. Before that, stand up the free reporting the platforms already give you. Google and Bing's official AI reporting rails cost nothing and answer a narrower question well.

Method note: all first-party figures come from a single full pass over 50 US questions across 15 answer surfaces, measured July 27, 2026. Every source URL was kept and recounted from the measurement ledger on July 28, 2026. One pass across every surface is the right tool for describing what a cited set is made of. Brand-to-brand comparison and confidence intervals come from the repeated-run grid described above. External studies are linked inline with their own dates and scopes.