AI Video Ad Pre-Testing Platform: Predicting Brand Lift Effects with Synthetic Audiences

This article introduces the open-source project AI Video Ad Simulation Core, which uses AI Agents as synthetic audiences, combined with multimodal LLMs and chain-of-thought reasoning, to predict lift metrics across seven dimensions of the brand funnel before actual deployment. This platform addresses the pain points of traditional advertising's 'deploy first, test later' model—high costs and delayed feedback—by adopting a dual-world control design to enable causal inference, providing a low-cost and efficient pre-testing tool for advertising creative iteration.

Traditional ad effect evaluation follows a costly feedback loop of 'produce-deploy-wait for data to return'. Each round of creative iteration consumes real media budget, and feedback often misses the optimization window of the current campaign, leading to resource waste and decision delays.

The platform initializes AI Agents with demographic and psychographic state variables as synthetic audiences to simulate target audience responses. It adopts a dual-world control design: for each audience, a pair of twin Agents is generated (Treatment group watches the ad, Control group does not), sharing baseline attributes with only ad exposure as the intervention variable. Through steps like Call0-A (segment baseline), Call0-B (individual baseline), and Treatment/Control Call1/2, the brand lift estimate is obtained by comparing differences between the two groups.

The system measures seven dimensions along the brand funnel:

• Upper-level cognition: Brand awareness lift, ad recall lift
• Mid-level attitude: Information relevance lift, favorability lift (0-1 unipolar scale)
• Lower-level behavior: Consideration lift, purchase intention lift, search intention lift It also calculates Headroom Lift (remaining space lift rate) and outputs a deterministic four-section comprehensive report to ensure consistency.

The platform consists of a seven-layer pipeline:

1. Ad pipeline: Video standardization, shot segmentation, ASR, OCR, visual feature annotation
2. Persona generation: Generate twin Persona pairs (sets of state variables) from demographic distributions
3. Simulation environment: Build session context (device, time, geographic location, etc.)
4. LLM-Agent simulation loop: Chain-of-thought reasoning for baseline estimation and causal response
5. Statistical analysis: Paired difference estimation, confidence interval calculation, power analysis
6. Diagnosis: Root cause attribution and improvement suggestion generation
7. Integrated output: JSON result contract for downstream use

The project openly discloses three major limitations:

1. Lack of real-world calibration data: Estimates are anchored to LLM judgments and not calibrated against real campaign results; only useful for relative comparisons
2. Chain-of-thought self-reinforcement: Call2 relies on Call1 narratives, so biases may compound—needs mitigation via funnel consistency constraints
3. Manual adjustment of Persona affinity constants: Not fitted to real deployment logs; only guides the direction of relative comparisons

The core values of the platform include:

• Reducing creative iteration costs: Test multiple variants before deployment to avoid waste of ineffective materials
• Rapid market testing: Quickly predict effects when there is no historical data for new markets/audiences
• Accurate causal attribution: Dual-world design accurately estimates the causal effect of ads
• Diagnostic insights: Identify the impact of ad elements on effects and provide optimization directions

AI Video Ad Simulation Core combines causal inference, multimodal LLMs, and engineering design to provide a brand-new pre-testing tool for the advertising industry. Although there are limitations, with improvements in model calibration and validation, such systems are expected to play a greater role in marketing science, promoting the transformation of ad effect evaluation from 'post-hoc measurement' to 'pre-hoc prediction'.