How Many Sources Does an AI Answer Actually Use?
Across 100 commercial queries through one engine, the answer was strikingly stable: about six cited sources per answer, never a couple, never a flood.
- Traces
- 100
- Citations
- 610
- Avg/answer
- 6.1
- Read time
- 4 min
When an AI answer engine responds to a commercial question, how many sources does it actually lean on? Not in theory — in the answer the user sees. We logged every citation across 100 commercial and GEO/SEO queries run through one engine (Claude Sonnet 4.6 with web search), and counted.
610 citations across 100 answers — an average of 6.1 per answer. Every single query came back with sources cited; not one answered from nothing.
The headline isn’t the average. It’s the stability.
The width is narrow and consistent
You might expect answer width to swing wildly — two sources for a simple question, twenty for a sprawling one. It didn’t. The per-answer count clustered tightly around six. The engine reaches for a small, consistent handful of sources and stops.
And it held across subjects. Grouping the 100 queries by topic, the average citations per answer barely moved:
| Topic | Queries | Avg citations |
|---|---|---|
| GEO | 24 | 5.9 |
| Software | 20 | 5.8 |
| Marketing | 17 | 6.5 |
| Technical SEO | 15 | 6.7 |
| AI tools | 12 | 5.8 |
| E-commerce | 6 | 5.8 |
| Content | 6 | 6.3 |
A 0.9-citation spread across seven very different topics. Whatever decides answer width, it isn’t the subject.
Why this matters for GEO
If an AI answer cites roughly six sources, then being cited is a competition for about six slots — and the field size is predictable. That reframes the goal. You’re not trying to crack a top-three of organic results; you’re trying to be one of ~6 sources the model considers worth quoting for that question.
Two practical reads:
- Zero-citation answers were absent here. All 100 commercial queries produced a cited answer, so “the model just won’t cite anyone for my space” is, at least for this kind of query, not the usual failure mode. If you’re not cited, someone else filled the slot.
- Width is stable, so depth of competition is, too. You’re not fighting a flood. Six slots, re-decided per query — which is exactly why who fills them is so concentrated.
Honest limits
N = 100, one engine, one query basket (English, commercial/GEO/SEO-weighted). This describes this run; a different basket would move the average a little. The figure to carry is the shape, not the decimal: an AI answer cites a small, steady handful — about six — every time. (An earlier 30-query pilot put the average at 6.9; the larger sample settled it lower, at 6.1 — the shape held, the decimal firmed up.) The raw per-query citations behind every number here live in the repo (src/data/panel/raw-latest.json) for anyone who wants to recompute them.