Where the Value Lies in a Multi-Agent Pipeline for Generative Engine Optimization: Evidence from Five Products and a Field Deployment
Year: 2026
Type: preprint
Abstract: Answer engines such as Doubao and Gemini now answer many product questions directly, so organisations need content these engines will cite. Producing it is a business process: choosing questions, writing, checking, adapting to many platforms, publishing and measuring. We built GEO-Loop, a multi-agent pipeline that runs this process with 23 agent roles around deterministic state, an automated release gate and evidence-based publication records, and asked where its value lies. On 20 topics for five products in Chinese and English, a single prompt that knew the gate's rules passed 17 times, as often as the pipeline, but by leaving out sourced and dated facts; asked to revise itself, it fell to 7, mostly by inventing numbers. The pipeline's first drafts passed 8 times; repairing only the failed checks raised this to 17 while keeping their evidence, and 19 of 20 platform versions passed after one repair. In a 19-day deployment the pipeline produced 164 pieces on ten channels for about US$7, of which 43% could be verified as published. Its value lies in the repair loop and in the stages a prompt cannot do, not in better first drafts, and pass rates alone favour articles that omit evidence.
Authors: Han, Lintong, Wang, Mengran
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236259827
