How Multi‑Agent LLMs Can Auto‑Optimize E‑Commerce Product Titles

Keywords in product titles are crucial for search visibility; using large language models (LLM) to automatically optimize titles has become a core capability in AI‑driven e‑commerce operations.

Background

Effective keywords boost exposure without additional fees, but manual optimization requires professional effort and constant monitoring of hot search terms.

Generation Schemes

1.0 – Direct LLM Generation

Candidate words are generated from similar product tokenization, hot queries, or competitor titles, and an LLM creates new titles.

Experiments showed low output quality, large title changes, and low merchant acceptance, leading to the design of a rule‑based approach.

2.0 – Rule‑Based Generation with Multi‑Agent Selection

Multiple traffic words are generated, then an LLM evaluates and selects the best ones for addition or replacement in the title.

Traffic Word Generation

Sources: recent high‑PV queries, competitor title points, product’s own traffic words.

Relevance measurement: weighted semantic similarity between added words and category/title, plus a small model for fine‑grained similarity.

Custom dictionary: filtered high‑frequency Chinese terms from platform tokenization tools.

Example JSON Output

{
  "titleList": [
    {"newTitle": "染色绗绣绣花亮片立体叶子珠管绣面料跨境童装婚纱礼服网布网纱", "newWord": "绗绣"},
    {"newTitle": "染色绣花花边亮片立体叶子珠管绣面料跨境童装婚纱礼服网布网纱", "newWord": "花边"},
    {"newTitle": "染色绣花印花亮片立体叶子珠管绣面料跨境童装婚纱礼服网布网纱", "newWord": "印花"},
    {"newTitle": "染色绣花盘花亮片立体叶子珠管绣面料跨境童装婚纱礼服网布网纱", "newWord": "盘花"}
  ]
}

Six‑Thinking‑Hat Framework for Keyword Selection

The process uses white (facts), red (intuition), black (criticism), yellow (optimism), green (creativity), and blue (management) hats to generate, evaluate, and select keywords.

Final recommendation from the framework: 立体 .

Engineering Design of Multi‑Agent System

Two main components: Agent clusters (handle LLM requests) and Pipeline clusters (orchestrate agents, manage context).

Agent Class

public Object execute(Object params, HashMap<String, Object> context) {
    params = preExecute(params, context);
    Object result = request(params);
    return postExecute(result, context);
}

Section Class

public Object execute(Object params, HashMap<String, Object> context) {
    if (CollectionUtil.isEmpty(agents)) {
        return params;
    }
    return agentExecute(params);
}

Pipeline Class

public Object execute(Object params) {
    AgentSection now = head;
    while (now != null) {
        params = now.execute(params, context);
        if (checkPoint.contains(now.getName()) && (params == null)) {
            return null;
        }
        now = now.next();
    }
    return params;
}

Example initialization shows how white, yellow/green, black, and blue hat sections are assembled into a pipeline.

Results and Metrics

Authorized merchants who applied the optimization saw a 69% hit rate for optimized keywords, a 33% product coverage, and a 59.12% positive impact on exposure, with 58.63% being significant growth.

Additional Example: Fruit Peeler Title Optimization

Using the same multi‑agent and thinking‑hat approach, the system selected the keyword 多用途 to improve exposure.