More Searches, More Sources? How Fan-Out Depth Shapes AI Citations
The short version
Across 13,647 AI answers from 69 experiments, the more fan-out searches a model ran, the more sources it cited in 11 of 12 models, but the number of brands it named barely moved in most of them.
Citations rise with fan-out depth in 11 of 12 models (within-model rank correlation +0.22 to +0.73).
Brand mentions mostly don't follow. GPT models gain brands with depth (+49% to +79%); Grok, Gemini 3.6 Flash and Claude Sonnet 5 stay flat (-10% to +2%).
One model reverses the rule. Gemini 3 Flash Preview cites 25% fewer sources in its deepest fan-out third than in its shallowest, and 82% of its experiments show a negative link.
Why fan-out depth matters
When an AI engine answers a prompt, many models first rewrite it into several smaller web searches, called fan-out queries. In our data the median answer ran 4 fan-out queries, and 53.7% ran four or more, with a maximum of 36.
A common GEO assumption is that more searches mean a wider net: more pages read, more sources cited, and more chances for a brand to appear. We tested that assumption directly, separately for citations and for brand mentions, and model by model.
How we measured it
We took the most recent successful visibility execution of every visibility experiment that ran between 26 October 2026 05:34 UTC and 2 November 2026 01:22 UTC. That is 24,013 results from 110 experiments, 30 organizations and 16 model names. Client details are deliberately left out.
We kept the 13,647 results (69 experiments, 4,388 distinct prompts, 12 models) that recorded at least one fan-out query. The other 10,366 results recorded none: four engines never log fan-out queries (Google AI Mode, Perplexity Sonar, ChatGPT and Copilot via a scraping provider), and some models skip fan-out on a share of prompts.
For each result we counted fan-out queries, cited URLs and named brands as stored in the platform. We then did three things:
Computed a within-model rank correlation (Spearman, computed as Pearson on average ranks) between fan-out count and citations, and between fan-out count and brands.
Split each model's results into three equal groups by fan-out count (low, mid, high) and compared average citations and brands.
Repeated the correlation after subtracting each experiment-and-model average, so a result is only compared with other answers from the same experiment and model.
Finding 1: deeper fan-out goes with more citations in 11 of 12 models
The correlation between fan-out count and citations count is positive in every model except Gemini 3 Flash Preview, and it is strongest in Grok 4.20 (+0.72), Claude Sonnet 5.5 (+0.73) and GPT-5.6 Sol (+0.66). Models with very few results (132 to 143) show large coefficients with wide uncertainty, so read those as directional.
Model | Results | Fan-out ↔ citations (rank ρ) | Fan-out ↔ brands (rank ρ) |
|---|---|---|---|
Claude Sonnet 5.5 | 143 | +0.73 | +0.20 |
Grok 4.20 | 1,760 | +0.72 | +0.01 |
GPT-5.6 Sol | 1,066 | +0.66 | +0.52 |
Claude Opus 5.5 | 143 | +0.64 | +0.23 |
Gemini 3.6 Flash | 1,111 | +0.58 | -0.02 |
GPT-6 Luna | 132 | +0.55 | +0.43 |
GPT-5.6 Luna | 2,743 | +0.42 | +0.35 |
GPT-6.1 Sol | 143 | +0.41 | +0.29 |
Gemini 3.5 Flash-Lite | 546 | +0.32 | +0.02 |
Claude Sonnet 5 | 3,589 | +0.26 | +0.10 |
Claude Opus 4.8 | 118 | +0.22 | +0.18 |
Gemini 3 Flash Preview | 2,153 | -0.23 | +0.10 |
In plain terms, comparing a model's shallowest third of answers with its deepest third, GPT-5.6 Sol cites 98% more sources (4.64 to 9.18), Gemini 3.6 Flash 45% more (8.49 to 12.30), GPT-5.6 Luna 39% more (5.48 to 7.60) and Grok 4.20 38% more (33.71 to 46.42). Claude Sonnet 5 moves only 7% (7.35 to 7.90).
Finding 2: brand mentions follow depth only in the GPT family
More sources do not automatically mean more brands. Across all 12 models the brand link is much weaker than the citation link (pooled within-experiment ρ of +0.10 for brands against +0.26 for citations, n = 13,647).
The split is by model family. In the two largest GPT models, the deepest third names 49% to 79% more brands than the shallowest third (GPT-5.6 Sol 7.01 to 12.52, GPT-5.6 Luna 8.24 to 12.31). Pooling every GPT result, the deepest third cites 52% more sources and names 56% more brands (n = 4,084).
For Grok 4.20 the same comparison goes the other way: citations rise 38% while brands fall 10% (24.57 to 22.16). Gemini 3.6 Flash gains 45% more citations but only 2% more brands (13.30 to 13.55), and Claude Sonnet 5 is flat on both (+7% citations, -5% brands). For those engines, extra searches widen the reading list without widening the answer.
Finding 3: Gemini 3 Flash Preview runs the most searches and cites fewer sources
Gemini 3 Flash Preview is the clear exception. Its shallowest third of answers averages 4.57 fan-out queries and 6.64 citations; its deepest third averages 12.11 fan-out queries and 4.96 citations, a 25% drop in sources for 2.6 times the searching.
This is not one noisy experiment. Looking at the 33 experiments (15 organizations) where it ran, only 18% show a positive link between fan-out depth and citations, and the median per-experiment correlation is -0.16. Its sibling models behave differently: Gemini 3.6 Flash is +0.58 and Gemini 3.5 Flash-Lite is +0.32. By contrast, GPT-5.6 Sol (11 experiments) and Grok 4.20 (24 experiments) are positive in 100% of their experiments, with median correlations of +0.48 and +0.70.
Our data cannot say why. One plausible reading is that on this model, long fan-out lists mark hard or ambiguous prompts that the model answers from fewer pages, but we did not test that.
Finding 4: pooled averages hide all of this
If you ignore the model and pool all 13,647 results by fan-out count, the picture looks like a hump, not a trend:
Fan-out queries | Results | Avg citations | Avg brands |
|---|---|---|---|
1 | 714 | 5.94 | 13.47 |
2-3 | 5,600 | 8.79 | 13.70 |
4-6 | 2,955 | 23.21 | 16.72 |
7-10 | 2,411 | 7.99 | 11.31 |
11-20 | 1,826 | 7.12 | 12.43 |
21+ | 141 | 9.37 | 13.62 |
The 4-6 spike (23.21 citations) is a composition effect, not a sweet spot. Grok 4.20 averages 4.64 fan-out queries and 39.84 citations per answer across 1,760 results, several times any other model, which is consistent with it dominating the middle buckets. Once the model is held fixed, the relationship is positive almost everywhere, as in Finding 1. Any benchmark that averages across engines without controlling for model will mislead.
What to do with this
Do not read a long fan-out list as a visibility win. Deeper search raises citations, but only the GPT family also raised brand counts in our data.
Track citations and brand mentions as separate outcomes per engine. A page that earns citations on Grok or Gemini 3.6 Flash may still not get the brand named.
Treat Gemini 3 Flash Preview on its own terms. Long fan-out there signals fewer sources, so do not use its depth as a proxy for reach.
Compare engines only with the model held fixed. Pooled benchmarks conflate model behavior with search depth (Finding 4).
Limitations
Correlation, not causation. Prompts that trigger more searches may be harder or broader, which also changes citations. Subtracting experiment-and-model averages reduces this but does not remove it.
One week, latest run per experiment. 26 October to 2 November 2026 only; earlier runs and repeat runs were not used.
Counts as logged. Citations are cited URLs, not unique domains, and brands are those the platform extracted. Fan-out counts depend on what each engine exposes.
Small samples for new models. Claude Opus 5.5, Sonnet 5.5, Opus 4.8, GPT-6 Luna and GPT-6.1 Sol have 118 to 143 results each; treat their coefficients as directional.
Excluded results. 10,366 of 24,013 results had no fan-out queries recorded, including every result from four engines, so findings apply only to answers where fan-out was visible.
Not weighted by organization. Experiments from larger organizations contribute more results, and the prompt mix is whatever customers chose to run.
Tercile groups split ties by row order, so group boundaries within a model are approximate.
