Jun 2026 Data

Does the Topic Change Who Gets Cited?

Across seven topics the number of sources an AI answer cites barely moved — but 91% of the domains it cited belonged to a single topic. Same appetite, almost entirely different worlds.

Traces
100
Topics
7
Domains
462
Read time
4 min

We split the 100-query panel into seven topics — GEO, software, marketing, technical SEO, AI tools, e-commerce, and content — and asked a simple question: does the engine behave differently depending on the subject?

The answer splits cleanly in two. How much it cites barely changes. Who it cites changes almost completely.

Count: flat

Average citations per answer, by topic:

TopicQueriesAvg citations
GEO245.9
Software205.8
Marketing176.5
Technical SEO156.7
AI tools125.8
E-commerce65.8
Content66.3

Roughly six, everywhere — a 0.9-citation spread across seven subjects. The engine’s appetite for sources is a near-constant, independent of subject (we unpack that stability separately).

Sources: disjoint

The domains, though, hardly overlap between topics. Of the 462 unique domains cited across the run, 422 (91%) appeared under a single topic. Only 40 domains straddled two or more subjects, and just five reached across four. The sites cited for “best CRM for startups” and the sites cited for “how to reduce server response time for SEO” are essentially two different worlds. Even the small recurring head splits along topical lines — the broad-coverage SEO tools bridge a few subjects, but the specialist sources cluster tightly on their own.

What this means for GEO

The operational takeaway is sharper than it first looks:

  1. “AI authority” doesn’t transfer. Being cited consistently in one topic buys you little in an adjacent one — nine in ten cited domains never leave their subject. The citation surface is rebuilt per topic, from a different candidate pool.
  2. Pick the topic, then compete in it. Because the sources are topic-bound, your competition for a citation slot is the handful of domains that own your subject — not a general-authority leaderboard. That’s a smaller, more tractable target.
  3. Breadth is expensive; depth compounds. Spreading thin across subjects means starting from near-zero in each. Owning one topic deeply is what puts you in its recurring head.

Honest limits

N = 100, one engine, seven coarse topic buckets, English commercial/GEO/SEO queries. The topic split is ours, not the model’s, and a finer grouping would shade the numbers. But the two-part finding — flat count, 91%-disjoint sources — is the structural read worth carrying. The per-query, per-topic citations are in src/data/panel/raw-latest.json.