Signal from Noise: AI in Marketing Research

During the holiday season, I got to catch up with friends, colleagues, and former associates. This post about AI was born from one of those interactions. While it's longer than usual, I'm not a fan of sequels, so here it is in one sitting.

DISCLOSURE: The following is based on a true story, but names have been changed to protect the not so innocent :)

SCENE: I hadn't seen John for a few years since we were co-workers, but we had kept in touch via LinkedIn, i.e., the auto-populated "congratulations," "how are you," "thanks for your reply," emoji, etc., so I was glad that he agreed to meet me for lunch.

LOCATION: a bustling cafeteria in a multinational pharmaceutical and life sciences company

ME (smiling widely): Good to see you! Congratulations on your promotion to VP of marketing research! How long has it been since we worked together, 10... 15 years? How are things going?

JOHN: Thank you. Yes, time flies; what happened to your hair? I head up marketing research for one of our consumer brands these days. I'm also responsible for looking at marketing research innovations and deciding which ones can apply to our business.

ME (disturbed but ignoring the "hair" comment): What have you found?

JOHN (looking pleased): AI will be huge for us. A few days ago, I brought BiggestBlue AI to pitch our team on their marketing research offerings. Machine learning, training algorithms with our data, neural networks, etc.—you name it, and they can do it. I was impressed with everything and can't wait to use them. 

[SCENE: End. Fade out]

---

Before answering "Where's the AI?" we should understand "How's the AI" in marketing research.

Visual representation of an AI neural network classification model used in market research

When assessing the significance of an AI/machine learning model, consider its training-learning process as the critical path. The diagram illustrates neurons, depicted by the boxes responsible for assigning and refining weights to intricate, non-linear data shown on the right. The connecting lines between neurons signify forward and backpropagation of information, enabling the model to learn and adapt through 'training' and 'test' datasets. Click here for a dynamic view of the model analyzing diverse datasets.

---

AI Through First Principles: Finding the Critical Path to Solving Marketing Research’s Core Problems

John's not alone in getting swept up by the AI tsunami and the FOMO (fear of missing out). In this next section, I'll delve into the role of AI/machine learning in marketing research, focusing on its applications in surveys and interviews. Something my data science professor always said and has stuck with me all these years, "90% of the work is preparation, i.e., ETL (Extract, Transform, and Load), the 10% is the reward when the data are ready for analyses."

1) Extract

Data scientists compile and utilize data from various sources, such as databases, online repositories, and real-time streaming feeds, to develop AI/machine learning models. In comparison, marketing researchers predominantly gather data through quantitative and qualitative interviews. Therefore, marketing researchers often face limitations in the quantity and diversity of data they collect, typically tailored to meet specific business goals.

Why is this important?

AI/machine learning algorithms thrive on extensive data, iteratively analyzing them to discern relationships between variables and enhance precision. The greater the data volume, the more robust the testing, training, and learning processes. Conversely, marketing researchers often face constraints due to less survey data in studies for specific business objectives. Even those with large sample sizes may not fully optimize learning and generalizability due to the lack of comprehensive records, diversity, unexamined variable relationships, and incomplete/missing data.

2) Transform

Most marketing research data are categorical, typically nominal or ordinal, and tend to describe attributes or qualities rather than quantities. Moreover, handling missing data is also critical to the accuracy of AI and machine learning models. To ensure optimal performance, marketing researchers must numerically encode categorical variables and impute missing values before feeding the data into a model.

Why is this important?

Since algorithms depend on comprehensive datasets for precise predictions, it's essential to carefully examine the data and ensure it aligns with business goals. Transformation methods are crucial when selecting the best methods for encoding and imputing categorical and missing data, with the ultimate goal of effectively integrating them into AI/machine learning models.

3) Load

Creating a central repository and integrating data across sources is crucial in ensuring data is available, accessible, and preserved for each type of analysis—descriptive, predictive, or prescriptive. It is a critical business intelligence and data management component for various analytical professionals and their preferred tools.

Why is this important?

Marketing research survey data is best optimized with other data sources, thus enabling advanced predictive analyses. Therefore, marketing researchers, statisticians, and data scientists can easily access the data. Each may use different analytical tools and applications (R, Python, TensorFlow, etc.) that suit their needs and the data types they choose to work with.

---

Now, the original question: "Where's the AI?"

Within the ETL (Extract, Transform, Load) framework, I've highlighted AI and machine learning applications that meaningfully enhance marketing research studies to date, ranked in order of their value-added impact.

Natural Language Processing (NLP)

NLP involves the development of algorithms and models that enable computers to understand and interpret text and human language in a meaningful and helpful way. Examples include sentiment and topic analysis, coding of open-ended responses, and speech-to-text transcription. Again, with the critical path of training and learning mentioned earlier, plenty of text sources (websites, social media, reviews, articles, etc.) are available to train language models. It's not limited to the text found in any given survey, and as a result, the relevance and accuracy of such models are constantly improving. So, where is/can it be applied?

• Brand Tracking: This is likely the most suitable marketing research methodology for applying AI/machine learning models. This method more closely aligns with the ETL framework—continuous data collection leading to higher volumes of data for model training and tuning, multiple sources of unstructured data (surveys, reviews, call centers, etc.), and repetitive activities primed for automation. Analyze and code open-ended responses in continuous surveys and interactions to identify sentiment trends, inform attribute development, and enhance key driver analysis.

• Product development: Harvest online conversations to gain valuable insights into customer needs, preferences, pain points, and sentiments about existing products or services. This approach is especially beneficial for early-stage ideation and exploration. An intriguing piece of marketing lore suggests that the frequent pairing of "Doritos" and "tacos" in various online discussions helped inspire the creation of Doritos Locos Tacos.

• Audience Segmentation: Analyze and encode open-ended, enabling a more thorough analysis and categorization of open-ended responses that improve cluster analysis and model inputs. This approach can also enrich and vividly depict the personas that emerge from the data.

• Focus group and in-depth interviews: Utilize transcription and text-to-speech conversion for enhanced efficiency. This approach allows for thorough sentiment analyses, topic modeling, keyword extraction, categorization, etc."

Predictive Classification Model

Through an ensemble learning methodology, marketing researchers can integrate passively and actively gathered data to construct and refine classification models, such as Logistic Regression, Decision Trees, Random Forest, Naïve Bayes, etc. Where is/can it be applied?

• Online Panel Sample Quality: The classification algorithms are trained to categorize respondents as either "good" or "fraudulent." Any given survey contains categorical and continuous variables marketing researchers can use to refine the model's accuracy. Passively collected data include time spent on the survey overall/specific questions, IP addresses, device fingerprinting (form factor, operating system, browser, location, click patterns, time and day taken, etc.) Actively collected data pertain to questions answered, such as mismatching responses, unlikely situations, demographics, firmographics, etc. The data types—passive/active and categorical/continuous can be combined to develop a robust classification and sample scoring model.

  
  

  Use a Receiver Operating Characteristics (ROC) curve to assess model performance in identifying legitimate and fraudulent respondents. Marketers can modify the tolerability ROC thresholds (specificity and sensitivity) and the True Positive versus False Positive rates based on the research needs and business decisions. Doing so allows a balance between achieving more completed interviews and applying less stringent criteria.

Generative AI

Generative AI and NLP are closely related in applying deep learning neural networks, text generation, and language understanding. Such NLP models, powered by conversational interfaces, typically undergo training and learning using extensive collections of textual data. Where is/can it be applied?

• Content Ideation and development: In marketing research, it functions as a back-end support tool for design and implementation, facilitating brainstorming and feedback for questionnaire development, crafting messaging, formulating product descriptions, and generating visual stimuli throughout the research process.

Computer Vision

An area of AI/machine learning, it focuses on enabling computers to interpret and understand visual information, including images and videos. Like language models, there are plenty of sources for training Computer Vision models (online photos and videos, natural scenes, aerial, etc.). Where is/can it be applied?

• Focus Groups, In-depth Interviews: Facial recognition technology enabled by Computer Vision enhances qualitative studies in marketing by allowing the researchers to assess respondents' reactions and emotions. This insight is valuable in understanding consumer sentiments towards products, advertisements, and retail experiences.

---

Conclusion

The encounter with John underscores three critical implications of AI in marketing research:

1. AI offers transformative potential but faces challenges in adaptation within a traditional marketing research framework.

   
   

2. The success of AI in marketing research relies on the ETL framework, emphasizing the need for extensive, diverse data collection, accurate data transformation, and effective data integration.

   
   

3. The practical applications of AI, including Natural Language Processing, Predictive Classification Models, Generative AI, and Computer Vision, highlight its potential to significantly enhance various aspects of marketing research, from data analysis to consumer behavior understanding.

These applications demonstrate AI's growing impact in shaping, innovating, and future-proofing your marketing research program. However, as likely the case with John's, when cutting through the noise, if it sounds too good to be true, it probably is.