Engineering a Generative Conversational Advertising System

1. Problem Definition: Conversational Advertising Requires a Different Approach

Advertising in a generative conversational product is not as simple as appending promotional text to the end of an answer. Conversations may span multiple turns, user intent is often implicit, task stages can change dynamically, and generated content is inherently uncertain. As a result, the advertising system must be designed as a controlled business workflow that covers request triggering, policy gating, candidate retrieval, ranking and bidding, generation and fusion, frequency and budget control, billing and attribution, review and compliance, and observability.

Compared with traditional display or search advertising, conversational advertising has three notable differences:

1. Stronger dependence on context. Users may reveal their actual needs only after several follow-up questions. Ad relevance cannot be determined from the latest utterance alone. It must also consider the conversation topic, task stage, historical preferences, and safety boundaries.

2. Higher risk during generation fusion. Once advertising information enters generated content, it must be clearly labeled and traceable to controlled sources. The model must not freely invent product benefits, claims, or commitments.

3. Greater sensitivity to user experience. Users come to conversational products for assistance. Ad density, frequency, and interruption can directly affect retention. Commercial returns must therefore be constrained by quality scores, relevance thresholds, and experience costs.

The technical goals can be summarized as follows:

• Controllable requests: Clearly define when an advertising request may be initiated.
• Explainable relevance: Ensure that ranking criteria can be audited.
• Auditable billing: Make impressions, clicks, and conversions traceable and idempotent.
• Measurable experience: Monitor negative feedback, skips, session interruptions, and related indicators.
• Traceable compliance: Maintain complete records for consent, data masking, review, and campaign suspension.

2. Overall Architecture and Data Flow

A maintainable conversational advertising system typically consists of the following layers:

• Access layer
• Context understanding and safety layer
• Candidate retrieval layer
• Ranking and bidding layer
• Generation and fusion layer
• Frequency and budget control layer
• Billing and attribution layer
• Review and compliance layer
• Data and experimentation layer
• Observability layer

The synchronous path should complete only the decisions required during the user request lifecycle. The asynchronous path should handle event persistence, billing reconciliation, offline training, and auditing. Keeping these paths decoupled prevents commercial logic from slowing down the primary conversational experience.

A typical data flow is as follows:

1. The access layer receives an ad-slot request and generates a request identifier.
2. The context understanding layer extracts the topic, intent, task stage, emotion, and sensitivity, then outputs a de-identified feature package.
3. The policy gate evaluates user consent, scenario suitability, advertising density, and safety boundaries.
4. The retrieval layer uses keywords, semantic search, rules, categories, and other methods to produce a candidate set.
5. The ranking layer combines relevance, quality score, bid, and experience cost to calculate ranking scores.
6. The bidding and budget layer normalizes bids, smooths budget consumption, and applies frequency controls.
7. The generation and fusion layer uses controlled templates or approved assets to produce clearly labeled ad segments.
8. The billing and attribution layer records impressions, clicks, conversions, and negative feedback, while applying idempotent deduplication and billing reconciliation.
9. The experimentation and monitoring layer continuously evaluates both commercial and user-experience metrics.

3. Key Module Design and Implementation Considerations

3.1 Context Understanding and Safety Checks

This module should not pass raw conversation text directly into the advertising pipeline. Instead, it should first identify and mask sensitive information, then extract only the minimum features required for ad decision-making.

Possible outputs include:

• Topic labels
• Intent confidence
• Task stage
• Commercial intent strength
• Safety level
• Scenario identifier

The policy gate should run before candidate retrieval. If the user has not provided consent, the conversation is sensitive, the user is a minor, or the content is high risk, the system should skip the advertising request immediately. Moving compliance checks forward avoids unnecessary computation in downstream modules.

3.2 Multi-Path Retrieval and Candidate Governance

Candidate retrieval can use common techniques such as:

• Keyword matching
• Semantic vector search
• Rule matching
• Category expansion

After candidates from multiple retrieval paths are merged, the system must:

• Remove duplicates
• Filter out assets that have not passed review
• Filter categories that conflict with the current context
• Preserve retrieval-source information for later explanation

The goal of retrieval is not to maximize the number of candidates. It is to maintain relevance and safety within a controlled computational budget.

3.3 Ranking, Bidding, and Constraints for High-Value Delivery

A ranking score is typically determined by relevance, quality score, normalized bid, and experience cost. High-premium delivery essentially means that an advertiser is willing to bid more for highly relevant, high-value intent. However, the system must not sacrifice quality or user experience in exchange for a higher bid.

Candidates with low relevance, low quality scores, or excessive frequency should be blocked by the policy gate.

The following illustrative pseudocode expresses the relationships among these factors. The parameters are examples only and do not define specific product rules:

def pass_policy_gate(context, ad, policy):
    if not context.user_consent:
        return False
    if not ad.audit_passed:
        return False
    if frequency(context.user_key, ad.category) >= policy.max_frequency:
        return False
    if budget_remaining(ad.campaign_id) <= 0:
        return False
    if experience_density(context.session) >= policy.max_density:
        return False
    return True
def rank_candidates(context, candidates, policy):
    ranked = []
    for ad in candidates:
        if not pass_policy_gate(context, ad, policy):
            continue
        relevance = semantic_relevance(context, ad)
        quality = quality_score(ad)
        bid = normalize_bid(ad.bid, policy)
        experience_cost = estimate_experience_cost(context, ad)
        score = (policy.relevance_weight * relevance
                 + policy.quality_weight * quality
                 + policy.bid_weight * bid
                 - policy.experience_weight * experience_cost)
        ranked.append((score, ad))
    return sorted(ranked, key=lambda item: item[0], reverse=True)

3.4 Frequency, Budget, and Experience Density

Frequency controls should cover at least the session, user, and category levels. Budget controls should support smooth spending and emergency campaign suspension.

Experience density limits how often ads appear within a single session. The system may also reduce ad delivery automatically when negative feedback increases. High-premium delivery must not bypass frequency or experience thresholds; otherwise, short-term revenue may lead to long-term loss of user trust.

3.5 Generation Fusion and Ad Labeling

Ad segments should be generated from approved assets and controlled templates, with a clear distinction between organic content and commercial content.

The system should perform asset-consistency validation before generation and safety review after generation. The model must not invent prices, performance claims, product promises, or qualification information. When confidence does not meet the required threshold, the system should return an ad-free result rather than force an ad into the response.

4. Data, Billing, and Attribution Engineering

The event model should include, at a minimum:

• Request identifier
• Scenario identifier
• Summary of contextual features
• Ad identifier
• Ad position
• Event timestamp
• Consent indicator
• Event type
• Idempotency key

Impression, click, conversion, and negative-feedback events require unified collection, deduplication, and retry or backfill mechanisms. Billing reconciliation should use reconciliation jobs and discrepancy alerts.

Attribution windows, deduplication rules, and statistical definitions must be versioned. Otherwise, changes in delivery strategies may cause metric definitions to drift over time.

At the infrastructure level, general-purpose services such as object storage, message queues, container platforms, logging services, and monitoring and alerting systems from Baidu AI Cloud can be combined to build the event pipeline, offline analytics, and observability foundation.

5. Privacy, Compliance, and Experience Governance

The advertising system should follow the principle of data minimization. It should use only the features necessary for relevance assessment, mask sensitive information, and avoid passing raw conversation context directly into the advertising pipeline.

The system should retain records for:

• User consent
• Content and asset review
• Ad delivery
• Campaign suspension

These records support auditability and traceability.

Generation fusion must clearly label advertising content and avoid misleading expressions. For sensitive categories, minor-related scenarios, and high-risk content, the system should establish policy gates, manual review, and emergency suspension mechanisms. This ensures that commercial capabilities operate within controlled boundaries.

6. Observability, Experimentation, and Implementation Roadmap

Monitoring metrics can be divided into five groups:

• System metrics: Request success rate, candidate fill rate, and error rate
• Policy metrics: Relevance distribution, frequency-control hit rate, and budget consumption
• Commercial metrics: Delivery and conversion-related outcomes
• Experience metrics: Negative feedback rate, skip rate, and session interruption
• Compliance metrics: Review interception rate and missing-label rate

Experiments should use stratified rollouts and long-term metric observation rather than focusing only on short-term click-through performance.

A phased implementation can proceed as follows:

1. Establish the event model and policy gates.
2. Validate relevance and user experience with a small amount of traffic.
3. Gradually increase traffic while improving billing reconciliation and automatic degradation.
4. Enter a continuous optimization cycle based on monitoring, experimentation, and audit results.

Overall, the core of conversational advertising is not a single ranking model. It is an auditable engineering system in which policy, data, compliance, and user experience work together.