We Replaced Four Facebook Ad Managers With OpenAI, Amazon Reviews, and Slack

TL;DR: We built a custom Slack Bot that analyzes Amazon reviews to create targeted FB ads, uses DALL·E for matching images, crunches ad data from Google Sheets, and predicts future ad performance. The total headcount of the creative team was reduced from five to only one in-house creative who controls and monitors the new workflow.

Overview of the project

Leveraging consumer sentiment, our team deployed a groundbreaking Slack Bot, designed to disrupt the traditional creative process in the ad space. Tapping into Amazon's data goldmine (aka. Amazon product reviews) we extracted pain points through advanced NLP analysis per product, informing our AI-driven creative generation which can be controlled by only one human in a Slack Workspace.

Business problems

The Business grappled with pinpointing target audience pain points, leading to subpar ad copy and mismatched creative assets. Ad image relevance often lagged, negatively impacting CTRs. Additionally, interpreting Facebook's algorithmic feedback loop posed a challenge, making ROAS optimization too complex for a small team of eight.

Objectives of the Slack bot integration

The Slack Bot streamlines ad ops by mining Amazon reviews for consumer pain points, fueling dynamic creative ideation. It taps DALL·E's API for imagery, tightly syncs with GSheets for real-time CPL and ROAS analysis to generate more ads based on the historical Facebook / Meta business suite export data.

Product Development

The Slack Bot conducts NLP-driven deep dives (Scraping and subsequent sentiment extraction) of Amazon reviews for each product. Our bot surfaces critical user frustrations for targeted creative messaging. It distills buyer feedback into actionable insights, refining value props for better understanding of buyer needs. This pain point extraction informs the generated ad copy, sharpening hooks to address consumer objections head-on, omitting the need for a lot of creative staff.

Integration with DALL·E API for image generation

We integrated the DALL·E API to turbocharge our creative asset pipeline. This additional feature unlocks a trove of hyper-relevant, data-driven visuals on-the-fly, massively slashing asset turnaround time. Ad sets now brim with bespoke imagery, tailored to campaign KPIs and dynamically optimized for peak performance marketing. We are currently working on an additional feature for the Slack Bot that ideates and writes UGC video scripts based on the scraped Amazon review insights and customer pain points.

Setup of report analysis from Google Sheets

We used the Google Sheets API for seamless data extraction. Then we added historical performance tracking to get cross-channel metrics for holistic view. This allowed for granular ad set optimization and enabled data-driven decision-making by ensuring CTR, CPA, ROAS analysis at scale through our conversational Slack Bot that's hooked up to the data.

How To Build A Facebook Ads Slack Bot Yourself?

Step 1: Develop a Data Extraction Process
To kickstart building your Facebook Ads Slack Bot, it’s crucial to systematize the extraction of consumer sentiment and detailed insights. Begin by integrating APIs to pull Amazon review data. Use tools like RainforestAPI to extract essential consumer feedback. Enhance this process by employing Google Cloud Natural Language for sentiment analysis and extracting detailed insights.

Step 2: Integrate Real-Time Analytics
Connect with data management tools and aggregate performance metrics using Google Sheets API. This tool helps in centralizing and summarizing data, making it easier to analyze ad performance. For advanced visualizations, employ the Tableau API at Tableau to create comprehensive dashboards that offer actionable insights into metrics such as CTR, CPA, and ROAS. This real-time data management will empower you to make data-driven decisions, leading to better ad campaign outcomes.

Step 3: Facilitate Human Oversight
Even with automation, human oversight is indispensable for fine-tuning ad campaigns. Utilize the Slack API at slack.dev to facilitate real-time communication, monitoring, and adjustments. Integrate the Slack Bot with a custom dashboard built using technologies like React.js and Node.js. This dashboard will serve as a centralized interface for the team to monitor ad performance metrics and make necessary adjustments.

Step 4: Refine Ad Messaging
To capture the attention of your target audience, refining ad messaging with AI insights is key. Harness the GPT-4 API to generate high-quality, tailored ad copy. This AI-driven approach ensures that the messaging resonates with identified consumer sentiments derived from earlier data extraction.

Step 5: Continuous Improvement and Scalability
As business needs evolve, so too should your ad strategies and infrastructure. Regularly update your system with new features and improvements. Utilize Docker to containerize new functionalities, ensuring they are seamlessly integrated and easily scalable. Orchestrate these containerized applications with Kubernetes, enabling efficient management of your infrastructure.

Sparring Time With Opsie!

‍

Q: Data Quality & Bias: How do you ensure the accuracy and reliability of sentiment extraction from reviews, considering that user-generated content can often be biased or manipulated by fake reviews?

Opsie: To ensure accuracy and reliability, we implement various data preprocessing steps to filter out noise and use advanced algorithms trained on balanced datasets to minimize bias. Techniques such as anomaly detection identify and mitigate fake reviews. Regular audits and updates maintain model accuracy over time.

Q: Overfitting to Negative Feedback: Might focusing excessively on "critical user frustrations" risk overfitting your creative messaging to a vocal minority, rather than representing the broader sentiment of your customer base?

Opsie: We balance the approach by analyzing a representative sample that includes both positive and negative feedback. Sentiment analysis tools provide a weighted average reflecting overall sentiment, preventing disproportionate focus on a vocal minority.

Q: Privacy Concerns: Have you considered the potential privacy implications of scraping reviews and analyzing them without explicit consent from the review authors, even if the data is publicly available?

Opsie: We anonymize personally identifiable information (PII) during analysis and comply with data protection regulations like GDPR to ensure ethical and legal standards are met.

Q: Dependency on External API: Do you have a contingency plan for scenarios where the DALL·E API becomes unavailable or significantly changes its pricing model, potentially increasing operational costs?

Opsie: We have a contingency plan that includes backup models and alternative APIs like MidJourney. Market monitoring for emerging technologies also ensures we are not solely dependent on DALL·E, mitigating risks related to availability or pricing changes.

Q: Technical Debt: Building a system with multiple API integrations can lead to significant technical debt. How do you plan to maintain and update this complex infrastructure over time?

Opsie: Adhering to best practices like modular coding, comprehensive documentation, and regular code reviews helps manage technical debt. Automated testing frameworks, dedicated resources, and support ensure ongoing maintenance and scalability.

Q: Over-Reliance on Automation: Are you concerned that over-reliance on automated insights and generated content could reduce the creative agency of your human team and potentially lead to homogenized ad content?

Opsie: A balanced approach involves human oversight in the review process. Automated insights are recommendations, not absolute decisions, preserving the creative agency of the human team and maintaining originality in ad content.

Q: Content Authenticity: How do you maintain authenticity in your ad messaging when using AI-generated content? Isn't there a risk that AI-generated content may lack the genuine touch of human-written copy?

Opsie: GPT-4-generated content serves as a draft, which is then reviewed and refined by our copywriting team. This hybrid approach ensures alignment with our brand voice and maintains genuineness in ad messaging.

Q: Ethical Considerations: Given that large language models can sometimes produce biased or inappropriate content, what measures do you have in place to review and filter the output from GPT-4?

Opsie: Robust review processes screen GPT-4 outputs for bias or inappropriate content. Filters and moderation tools prevent harmful or unethical content. Regular updates with diverse and inclusive data also help mitigate biases.

How Did We Replace Creative Staff?

The Slack Bot significantly uplifted ad performance metrics, streamlining the creative process for agile campaign deployment. Efficiency skyrocketed, truncating turnaround times and optimizing ad spend through precise predictive analytics. The total headcount of the creative team was reduced from five to only one in-house creative which is trained in controlling the workflow as the human-in-the-loop.