本研究以 LLM 多模態 persona 模擬虛擬訪客觀看網頁與素材後的點擊決策，生成推薦系統的初始訓練資料，檢驗其在零真實訓練樣本下的機器學習可用性。以醫療資訊網站為個案，由五位虛擬訪客依三種累進覆蓋配置生成合成資料，訓練與真實資料組相同的 27 個異質演算法，並於同一真實測試集進行跨資料驗證。 離線回測顯示，真實資料最佳模型之策略效益達基準的 1.78 倍，合成資料最佳模型亦達 1.66 倍；在真實資料表現前十名的模型中，合成訓練平均保留約七成策略效益。結果顯示，LLM 合成資料的價值不在精確複製真實行為分布，而在真實資料到來前提供可跨模型使用的機器學習訓練訊號。本研究據此提出「合成資料啟動、真實資料校準、合成資料退場」的冷啟動路徑。生成式 AI 與 AI agent 正逐步改變使用者搜尋、瀏覽與點擊網站的方式，本研究方法亦可作為未來代理瀏覽情境下合成訓練資料研究的基礎。;This study uses LLM multimodal personas to simulate virtual visitors viewing webpages and creatives and making click decisions， thereby generating an initial training set for recommender systems. Its machine-learning utility is evaluated under a zero-real-sample training condition. In a healthcare information website case， five virtual visitors generate synthetic data under three cumulative coverage configurations. The same 27 heterogeneous algorithms are trained on the synthetic and real datasets and evaluated on a shared real test set. Offline backtesting shows that the best model trained on real data achieves 1.78 times the baseline strategy effectiveness， while the best model trained on synthetic data achieves 1.66 times. Among the ten best-performing models under real-data training， synthetic-data training retains approximately 70% of their strategy effectiveness on average. These findings indicate that the value of LLM-generated synthetic data lies not in precisely reproducing the distribution of real behavior， but in supplying machine-learning training signals that remain useful across models before real data become available. Accordingly， this study proposes a cold-start path of “synthetic-data initiation， real-data calibration， and synthetic-data retirement.” As generative AI and AI agents reshape how users search， browse， and click through websites， the proposed method also provides a foundation for future research on synthetic training data in agent-mediated browsing environments. Keywords: cold-start recommendation， CTR prediction， LLM synthetic data， multimodal persona simulation， cross-data validation
