The Domains One Engine Keeps Coming Back To

Run a hundred commercial queries through one AI engine and it cites 462 unique domains. The striking part — covered in the long-tail study — is that 85% of them appear exactly once. But a head recurs. We pulled the domains cited across three or more of the hundred answers. There are 23.

semrush.com (12) · medium.com (7) · ahrefs.com, arxiv.org, blog.hubspot.com, neilpatel.com, searchengineland.com (6 each) — then a longer tail of tools and trades down to three.

That’s the recurring head. One domain — Semrush — straddled twelve of the hundred answers; Medium seven; then a cluster of SEO tools and publishers at six. Below them, another sixteen domains recurred three, four, or five times.

It’s the SEO-tool ecosystem

Ask someone to name the sites an AI “always cites” and you’ll hear Wikipedia, the big news brands, maybe Investopedia or GitHub. That’s not what recurs here. Wikipedia shows up in just three answers; the household news brands, essentially never. What recurs instead is the working toolkit of the SEO and content-marketing trade:

• The tools — semrush.com, ahrefs.com, conductor.com, zapier.com, clickrank.ai. Software vendors that also publish heavily.
• The trades and blogs — searchengineland.com, blog.hubspot.com, neilpatel.com, medium.com. Where the field writes about itself.
• The reference layer — arxiv.org, developers.google.com, en.wikipedia.org. Surfacing on the more conceptual and technical questions.

What the 23 share isn’t fame. It’s broad topical coverage of this query set — each publishes across enough of the questions to be the relevant source more than once. Recurrence comes from coverage, not from a brand halo.

The short list was mostly noise

There’s an honest correction buried here. An earlier version of this study used just 30 queries and found only five recurring domains: arxiv, searchengineland, stackmatix, digitalapplied, and techradar. Two of those didn’t survive the larger sample. At 100 queries, digitalapplied.com was cited exactly once — it fell all the way into the one-time tail — and techradar.com appeared just twice, below the recurrence line. They were never structural; they were sampling luck.

Meanwhile the domains that were structural (Semrush, Ahrefs, HubSpot, Neil Patel) barely registered as a “head” at 30 queries and only resolved into one at 100. This is the same lesson the long-tail number taught us when it moved from 93% to 85%: a recurring head read off a small sample names the wrong domains. You need scale before the head is trustworthy.

What this means for GEO

1. Recurrence is the signal worth chasing — but only once it’s stable. A domain that recurs across several queries in a large sample of your space is a structural authority. One that tops a 20-query pilot might just be noise.
2. The head is a working ecosystem, not a wall of mega-brands. It’s tools, trade blogs, and reference sites — many of them reachable. In a young field, the recurring head is winnable by whoever covers the topic most usefully and publishes consistently.
3. Map your own head at scale. Run your space’s real queries — enough of them — and see which domains recur. Those few are your actual competition for the citation slot, not the household names you assumed.

Honest limits

N = 100, one engine, one English-language commercial/GEO/SEO basket. A different query set would surface a different head. The point is the structure: across any basket, expect a tiny recurring head and a vast one-time tail — and don’t trust the specific members of that head until the sample is large. You can browse the full ranked list — and re-run it — from src/data/panel/raw-latest.json.