Prompting patterns for marketing with LLM evaluation
AI Marketing Blueprints #2
Outline
Introduction
In series #2 of the AI Marketing Blueprints, we explore essential prompting patterns for marketing tasks and evaluate the performance of leading LLM models through qualitative human preference testing.
Our analysis focuses on four common marketing tasks, each showcasing a powerful prompting pattern that marketers can immediately apply:
A systematic 6-step recipe to structure prompts like a management consultant
Claude thinking protocol for critical communications writing
Meta-Prompting with ChatGPT to generate system prompts that improve outputs
Leveraging LLM role-playing capabilities for brainstorming and interviews
To identify the most effective approach for marketers, we evaluated these patterns across 4 leading models from 3 major foundation model providers. The results demonstrated clear performance leaders for each task (Exhibit 1).
Exhibit 1: Prompting patterns for 4 common marketing tasks with LLM evaluation results
This comparative analysis revealed several crucial insights:
First, no one size fits all, each model excels in specific tasks, highlighting the importance of choosing the right tool for the job.
Second, the foundation model landscape is hyper-competitive. Marketing teams need to regularly evaluate and potentially combine multiple models to optimize their workflows and applications.
Finally, we anticipate the rise of specialized, fine-tuned models targeting specific steps of the marketing value chain. This evolution will provide marketers with increasingly sophisticated tools optimized for their unique needs.
Task #1: Marketing insights - Structure prompts like a management consultant
Task:
As the head of marketing of an alcohol free wine online shop, you would like to keep up to date on the market trends for alcohol free wine.
A 6-step recipe to structure your prompt:
1. Give a detailed summary of market trends for alcohol free wine. Create a bullet list of no more than 5 key trends.
2. Examples of trends include health and wellness, sustainability, and premiumization. Do NOT use exactly these names but only as inspirations.
3. Only include things you can find support for in the text and include facts & numbers by searching through named industry reports and references.
4. Answer in the voice of an experienced marketer with 20-year expertise.
5. Answer to me as if I am an online retailer for alcohol free wines, with detailed insights.
6. Answer in the following structured format for each trend:
- Name: Give a brief name to describe the trend;
- Context and Background: Brief introduction and explain why the trend has emerged;
- Market Segmentation: main segments that the trend applies to;
- Product or Service Innovation: Detail any unique product or service offerings, what are the key features incl. taste, quality, or price;
- Challenges and Limitations: Barriers to Adoption or market saturation;
- Real world examples: with case studies and notable brands or leaders.
Step explanations:
Provide a clear goal and instructions
Provide context and examples as needed
Ask to mitigate hallucinations
Ask to answer from the viewpoint of relevant expert
Describe the audience the answer is for
Provide desired output format
Example outputs (ChatGPT o1):
LLM evaluation:
Exhibit 2: Marketing insights with the 6-step recipe prompts
Key learnings:
Structured prompts using the 6-step recipe improves the outputs significantly compared to generic prompts.
Reasoning-Based models outperform generic models: Results show that a “reasoning mode” yields stronger analytical and insight-focused answers than more general LLM setups.
ChatGPT o1 leads in overall performance: Of all models tested, ChatGPT o1 delivers better outputs, scoring best on both structure and depth of data support. While Claude 3.5 performs decently, it typically provides less structured responses and fewer data-backed insights compared to the other models.
Task #2: Lead outreach emails - Critical communications writing
Task:
As a founder of an alcohol free wine shop, write a lead outreach email to a corporate event planner at a large consumer goods company.
Prompt:
Write a prospecting email for your alcohol free wine collection. Write in a concise, warm tone and highlight the following points:
- focus on the benefits of alcohol free wine in corporate event;
- you have a large curated collection of alcohol free wines suitable for events, for example, the award winning STRAUCH collection;
- you have 25+ year experience in the wine industry and has worked with many companies;
- Then close with a clear Call for Action and ask for an 15 min discovery call.Write to a corporate event planner at a large consumer goods company.
Write in the voice of an experienced founder/owner at an online alcohol free wine shop.
Example outputs (Claude 3.5 Sonnet thinking):
LLM Evaluation:
Exhibit 3: Lead outreach emails with critical communications writing
Key learnings:
Claude Thinking Protocol excels in smooth, sophisticated, and natural outputs, making it ideal for tasks requiring fluidity and human-like coherence (e.g., creative writing).
ChatGPT outperforms in contextual understanding, particularly for nuanced, multi-layered discussions (e.g., technical explanations, long-form analysis) where retaining thread depth is critical.
Combining outputs from different models yields optimal results.
Task #3: TCO analysis - Prompt generation for analytics tasks
Task:
Build an Excel model for analyzing 3-year TCO (Total Cost of Ownership) of implementing and running a Data Reporting solution.
Meta prompts:
Follow the instructions of OpenAI prompt generation to use a higher level model to generate system prompts based on a description of tasks (original prompt).
Original prompts:
Build an Excel model for analyzing 3-year TCO of implementing and running a Data Reporting solution. Perform a comprehensive break down and list all possible cost components of one-time implementation and recuring costs. Structure the outputs into a clean tabular format. The Data Reporting solution has 3 main cost scenarios: 1. Raw Data: Microsoft Azure, Compute: DataBricks, Dashboard: Microsoft PowerBI. 2. Raw Data: Google GCP, Compute: Google Big Query, Dashboard: Google Looker. 3. Raw Data: Amazon AWS, Compute: Amazon Redshift, Dashboard: Tableau. Assuming there is 1TB of raw data and grows at 10% YoY. Only include facts and numbers from trusted vendor websites and references. Answer in the role of an IT executive with 20-year hands on experience. Answer to me as if I am the business decision maker of a large consumer brand.
Example outputs:
LLM Evaluation:
Key learnings:
ChatGPT’s meta-prompting enhances prompt generation and refinement, making it valuable for strategic prompt engineering.
DeepSeek delivers precise, actionable structures with granular details (e.g., cost breakdowns) that are readily usable for analytics tasks.
Always verify facts/numbers from tools like DeepSeek with trusted sources to ensure accuracy, especially for high-stakes decisions.
Task #4: Brainstorm & interview - Ask LLM to play a specific role
Task:
Brainstorm questions to help understand a persona and interview persona to get specific answers to the questions
Prompts:
Flipped interactions:
I am a potential customer in the Persona “The Health-Conscious Millennial”, for an alcohol free wine online shop.
Ask me questions to help you understand my detailed preferences and profile information.
Ask questions until you have gathered sufficient information to provide personalized contents (copy, image, video) that will appeal to me.
Interviews:
You are now a potential customer in the Persona “The Health-Conscious Millennial”, young woman in your 20s, single, working in Berlin, Germany. Answer above questions with detail and examples if you can.
Example outputs:
Questions generated:
Answers:
LLM Evaluation:
Key learnings:
ChatGPT and DeepSeek excel in structured, comprehensive outputs, offering logically organized and thorough responses.
DeepSeek generates granular detail, delivering rich data points, numerical specificity, and concrete examples.
Summary
We’ve developed a blueprint outlining four distinct prompting patterns designed for common marketing tasks, highlighting how the best results often come from combining the unique strengths of different LLM models. For those building applications in specialized domains like marketing, this presents a prime opportunity to harness rapid innovation from an increasingly competitive foundation model landscape.
