One Wide Prompt or Multiple Narrow Rounds? Budgeting an Evidence Portfolio for GEO

Given 24 document slots, should a system supply all 24 documents once, or supply two documents across 12 generations? New research favors the second allocation for evidence coverage, but it also consumes more tokens and nearly ten times the sequential latency in one representative comparison. This is a cost–coverage frontier, not a free optimization.

The August 24, 2026 paper *The Laws of Context Allocation* separated context width k from generation count T. With the same 24-slot evidence budget, the two-document, 12-round setup achieved 0.144 higher portfolio recall than the 24-document, one-round setup, with a 95% confidence interval from 0.119 to 0.170. The paper summarizes broad sequential-generation gains of roughly 16.8 to 20.5 percentage points and reports positive extensions on Qwen 14B and 32B models.

The benefit was not merely sampling the same answer more often. Repeated generation from a fixed context improved recall, but rotating fresh documents into each round added an absolute 0.087 to 0.140 over fixed-context resampling. Fresh evidence exposure, not paraphrase diversity alone, raised the ceiling in this experiment.

Separate three types of budget

The document budget is the total evidence a model encounters. The generation budget is the number of rounds. The validation budget covers attribution probes, human checks, and deduplication. Saying only “the same number of calls” or “the same token count” can hide a material difference on another axis.

In the paper's representative pair, the narrow multi-round configuration used about 1.5 times the tokens and took 17.5 seconds sequentially, versus approximately 1.7 seconds for the wide single pass. Leave-one-out probes added non-zero but parallelizable computation. A simple task needing one answer may not justify this cost. A research task requiring many independent facts, entities, or alternatives is a better candidate.

A comparable experiment for a brand team

Brands rarely control how a commercial answer engine allocates internal context, but they can control their monitoring budget. Choose questions that require multiple facts and predefine a fact checklist. Compare one broad question, several narrow facet questions, and a controlled follow-up sequence. Keep platform, date window, total runs, and scoring rules comparable. Track supported fact coverage, brand mentions, citations, duplication, cost, and elapsed time.

Do not simply concatenate all responses. Deduplicate equivalent facts, flag conflicting versions, and avoid feeding a newly discovered brand name back into the original-query condition; that would introduce answer-side leakage.

GEO Radar at https://www.georadar.top can maintain fixed facet-based question groups and observe multi-platform answers over time, serving as the collection layer for this budget experiment. It cannot control an external platform's retrieval window, and the paper's 0.144 gain is not a promise of brand visibility. Decide on cost per new, source-supported fact—not the longest answer.

Sources for this article

• arXiv, August 24, 2026, *The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search*: https://arxiv.org/abs/2608.23252
• arXiv HTML full text, including the factorial grid, evidence rotation, scale validation, and cost–latency analysis: https://arxiv.org/html/2608.23252v1
• Authors' code, data, and measurement tools: https://github.com/PeiYangLiu/ascp