Using OpenAI GPT to Generate Product Marketing Content

Early Lessons

Background

About 6 months ago at BPM, we started testing OpenAI’s GPT large language model to see what it could do when we connected it to the Messaging Workbench – our software for capturing and developing customer insights and messages to take products to market.

Initially we used GPT3.5 and GPT4, and most recently we’ve been using GPT4o.

In this paper I’ve summarized what we’ve learned so far!

Key findings

• GPT can write excellent English but getting it to generate marketing content that is on point and accurate for a particular purpose is hard. Getting it to do this reliably, time after time, is also a challenge. But it is definitely possible, and the results can be astounding!
• The bar to generate product marketing content using GPT is high – due to the fact that extensive input knowledge (on customers, typical challenges, product capabilities, differentiation, etc.) needs to be uploaded to GPT, as the basis for the copy you want it to write.
• Interacting with GPT via the API is unlike interacting with any other piece of software! You send it exactly the same input data, you give it precisely the same prompts and… every time you get slightly different results in terms of the content it writes!
• However, get your prompts optimized and your input data organized correctly, and you can generate reliable product marketing content that makes the same key points every time, but just uses different words and phrases.

How large language models work

Naturally, I asked Microsoft Copilot (i.e. ChatGPT4) to write me a summary ;-)…

“Sure! Large language models, like GPT-4, work by learning patterns from vast amounts of text data. Here’s a simple breakdown:

1. Training: They are trained on diverse text sources, like books, articles, and websites. This helps them understand language patterns, grammar, and context.
2. Neural Networks: These models use neural networks, which are like complex webs of interconnected nodes. Each node processes bits of information and passes it on, mimicking how the human brain works.
3. Prediction: When given a prompt, the model predicts the next word or phrase based on the patterns it learned during training. It does this repeatedly to generate coherent and contextually relevant responses.
4. Fine-Tuning: They can be fine-tuned for specific tasks, like answering questions or writing essays, by training them on more specialized data.

“In essence, they learn from a lot of text and use that knowledge to generate human-like responses.”

Not a bad explanation!

We’ve found that GPT can write excellent English that’s indistinguishable from content written by a great copywriter. But getting it to generate content that is on point and accurate for a particular purpose can be really challenging, as is persuading GPT to do this reliably time after time!

Prompt engineering

GPT has some idiosyncrasies – it's a bit like working with a young, smart intern who will sometimes go off-piste and take a different approach to the one you requested!

So, the way you interact with GPT, through the instructions or prompts you send to the model, has a huge impact on the quality and reliability of the results you get. Your prompts have to be very precise and unambiguous, and you often have to repeat key instructions to ensure reliable results.

The process of optimizing your prompts so you get the results you want, reliably, is known as Prompt Engineering.

You can read a lot on the web about prompt engineering and how you need to break your instructions for GPT down into relatively small steps that follow a logical order. Our testing confirms this does work and hugely improves results and reliability.

Interacting with GPT is unlike writing to any normal software API. You give it the same input data and exactly the same instructions, but you get different text generated every time. But the ideas, concepts and value propositions are essentially consistent.

This makes it very hard to know when you’ve optimized your prompts for a particular task. In our use cases, which revolve around the generation of marketing content, it becomes pretty subjective when you’re judging whether this elevator pitch is better than that one!

Input data

OpenAI has spent perhaps as much as $100m allowing the GPT4 model to learn from gargantuan volumes of text from a variety of sources – some reliable, some less so.

You can ask GPT to write copy with minimal instructions and it will generate remarkably well written text. So, for writing a best man speech, submit some basic facts about the bride and groom, a few funny stories and a bit of guidance on the style of writing you’re after, and hey presto you get a pretty good speech back!

However, if you want GPT to write accurately about a fairly technical subject you have to provide it with relevant input data (sometimes termed external data) to augment its knowledge.

One approach is to train your own version of the GPT model on your own data, but this is expensive and not a practical option for most organizations. There is a feature called GPT Builder which allows you to configure your own GPT with certain instructions and give it files of text to refer to which can contain input data. This can work well for use cases where your input data is relatively static.

The other approach is to upload data to GPT during the prompting process. This can be done using JSON code or by uploading a data file to GPT as part of the interaction. Where you want to use different input data each time you use GPT to generate something or where your input data needs to be updated on a regular basis, this is probably the best way to go.

We’ve found that what data you chose to upload, the way you structure the data, and how you explain to GPT what each row and field of data means, makes a fundamental difference to the quality and accuracy of the results you get.

Why generating product marketing content is hard for a large language model

Writing product marketing content demands a deep understanding of customers, their challenges and how the product addresses them and delivers value. Understanding differentiation versus competitors is also important.

This means the bar to generate on point product marketing content using GPT is higher than for the generation of more general marketing content. Your input data needs to equip GPT with relevant, trusted knowledge on, for example:

• Themes and drivers in target markets
• Customers and their challenges / opportunities
• The product’s capabilities to solve problems
• How customer value is delivered
• Points of differentiation

You could simply upload some existing product marketing materials to GPT and ask it to produce the copy you are after. But we’ve found this often results in ‘word salad’ – copy that sounds nice when you first read it, but actually doesn’t convey any useful information! Also, any errors or unfounded claims hiding in your existing materials will be carried forward by GPT.

We’ve found from extensive testing that input data needs to be presented to GPT in a highly structured format, so you can:

• explain the purpose of every block of information, and
• you can help GPT associate particular insights and messages with your market segmentation e.g. industry segments, buyer personas, use cases, geographic markets.

Then when you ask GPT to write an elevator pitch for salespeople to use in front of (say) the Head of Therapeutics at a Pharma tech show in Singapore, it can write you a carefully crafted message exactly for this target.

What product marketing content can GPT generate?

The ability of GPT to read and analyze multiple lines of input data on market themes, customer challenges, product capabilities, value propositions, and points of differentiation and then synthesize this to generate summary copy for a particular purpose is truly astonishing!

By combining very carefully designed prompts with the relevant input data structured to optimize GPT’s understanding, this is where we are right now…

Proposition summaries (i.e. elevator pitches)

Boiling down input data for a given proposition (can be several thousand words) to generate a couple of summary paragraphs, is perhaps the most remarkable thing GPT can do! Here’s part of an example interaction…

You

Excellent thank you! Now could you write an overall summary for this proposition comprising 3 paragraphs. The first paragraph should set out the market context, drawing on content provided in Section 2 Market Context and Drivers of the document. The second paragraph should summarise the key challenges faced by customers and explain how the solution addresses these and delivers value, based on your own analysis of the Section 3 Value Propositions tables. The final paragraph should explain why this solution is different from competing options, based on your analysis of the Top Points of Differentiation provided in Section 4 of the document.

AI Co-author

AI Content Generation Proposition Summary

The use of AI for content generation is gaining traction due to the pressing need for marketing and sales departments to enhance their output quality and efficiency. The evolving landscape brings concerns around the reliability and accuracy of AI-generated content, especially with issues like hallucination. However, the increasing demand for scalable and high-quality content generation presents a significant opportunity. Marketing teams can see the potential in AI to elevate their productivity, while sales enablement teams focus on removing bottlenecks in creating product support tools using AI.

Customers face several challenges, such as insufficient marketing resources, skill shortages, and the rapid obsolescence of crafted messages. These issues result in delayed campaigns, lost market opportunities, and subpar product launches that do not generate the expected leads. The AI-driven Messaging Workbench addresses these problems by allowing users to generate compelling and accurate marketing content quickly, using signed-off Proposition Maps to ensure reliability. This capability drastically reduces content creation time, keeps messages up-to-date, and enables sales teams to access tailored, high-quality materials promptly.

This solution stands out due to its unparalleled accuracy and reliability in AI-generated content, virtually eliminating hallucinations by using controlled and verified knowledge sources. Additionally, the Messaging Workbench produces first-draft copy that rivals the quality of top-tier B2B copywriters, offering a competitive edge. By ensuring the generation of precise and persuasive marketing messages, this AI solution helps organizations to maximize their campaigns' effectiveness and achieve faster, more impactful sales outcomes.

If you have further specifics you would like to dive into, please let me know!

Web content

GPT can produce high quality copy for corporate websites but usually a few iterations are required to get the style and tone of voice correct.

Foot in the door emails

GPT does a really nice job on these providing you tell it the purpose of the emails and give it some ideas for topics (unless these are already in you input data). Here’s part of an example interaction…

You

Excellent thank you! I need to write an introductory email to Julie Peterson, Global Head of Intellectual Property at Sarepta Therapeutics. The objective of the email is to secure an introductory phone call with this prospect. From my web research, I believe that one of the challenges the Company faces is Developing a view of the innovation landscape and trends. I was given this prospect's name by Johanson Largas, Global Head of Intellectual Property at Legend Biotech. I will give you guidance for the format and content of the email in one moment.

?

AI Co-author

Great! Please go ahead and provide the guidance for the format and content of the introductory email. I'll draft it accordingly.

?

You

Here is my guidance for the format and contents of the email...? Total length: maximum 250 words with title, salutation and 4 paragraphs plus signature. Email Subject Line:?Maximum 8 words to include topic of interest to target persona that will get their attention.?Introduction:?First paragraph: Introduce me [put your name, title and company name here] to the prospect referencing the person who provided the introduction. Explain in a few words how my company typically helps companies like one the prospect works for.?Purpose:?Second paragraph: Explain the purpose of the call, which is to have an introductory conversation and then schedule a face to face or virtual meeting if the prospect is interested in continuing our discussions.?Value:?Third paragraph: Explain the potential value of the meeting to the prospect. Include references to the potential benefits we can provide to the prospect personally in their role and to their company, using typical metrics for someone in this role e.g. [put typical goals and KPIs for this persona in here].?Action:?Fourth paragraph: Ask for a specific time for the introductory call.?Signature:?[add your normal signature here].

?

AI Co-author

Draft Email

Email Subject Line:?Enhance Innovation Insights for IP Strategy

Hi Julie,

My name is [Your Name], and I am the [Your Title] at [Your Company Name]. Johanson Largas, Global Head of Intellectual Property at Legend Biotech, suggested I reach out to you. At [Your Company Name], we specialize in helping companies like Sarepta Therapeutics gain a comprehensive view of the innovation landscape and trends, crucial for strategic intellectual property management.

I would love to schedule a brief call with you to introduce ourselves and discuss how we can support you in enhancing innovation insights. The aim of the call is to see if there might be a fit for a more detailed discussion, either face-to-face or virtually.

Our solutions have the potential to directly benefit you in your role by simplifying the complexity of navigating large datasets and identifying new opportunities in the pharmaceutical landscape. Typical metrics we impact include innovation cycle times, efficiency of data retrieval, and overall strategic positioning in the market.

Could we schedule a call next week to discuss this in more detail? Please let me know a time that works for you.

Best regards, [Your Name] [Your Title] [Your Company Name] [Contact Information]

Please provide your name, title, company name, and contact information, and let me know if there are any adjustments you'd like to make to this draft.

?

You

That's exactly what I was after, thank you!

?

AI Co-author

You're welcome! I'm glad I could meet your expectations. If you need any further assistance or modifications to the email, feel free to let me know. Good luck with your introductory call!

?

You

Could you write me a follow up email, perhaps focusing on a different area of challenge or opportunity, in case Julie doesn't respond to the first one!

?

AI Co-author

Absolutely! Here's a draft for a follow-up email focusing on another challenge—specifically, data integration and management:

Draft Follow-Up Email

Email Subject Line:?Streamline Data Integration for IP Excellence

Hi Julie,

I hope this email finds you well. I wanted to follow up on my previous message regarding how [Your Company Name] can assist Sarepta Therapeutics in developing a comprehensive view of the innovation landscape.

In addition to that, I believe that efficient data integration and management can significantly enhance your intellectual property strategy. Sarepta Therapeutics, like many in the pharmaceutical industry, likely faces challenges with handling extensive datasets and ensuring that crucial information is accessible and actionable.

Our solutions are designed to unify scattered data systems into a single, streamlined platform, making it easier for teams to access and utilize essential information. This can lead to more informed decision-making and a stronger, more agile IP strategy.

Could we find a convenient time next week for a brief call to explore this further? I believe this discussion could reveal some valuable opportunities for your team.

Looking forward to your response.

Best regards, [Your Name] [Your Title] [Your Company Name] [Contact Information]

Feel free to provide your details and adjust the message as needed. Let me know if there's anything else I can help with!

?

You

Perfect, many thanks!

?

AI Co-author

You're very welcome! I'm glad I could help. If you need any further assistance, whether it's drafting more emails or any other task, feel free to reach out. Good luck with your outreach!

Tailored presentation slides

GPT can write the copy for bullet point slides covering, for example:

• the particular challenges faced by the prospect,
• how the product solves these and delivers value,
• and why the prospect should choose you.

As part of your prompts, you have to, of course, provide GPT with additional information on what you’ve found out about the prospect in question from your ‘needs discovery’.

Product fact sheets

Longer form items such as these are more challenging to generate using AI, as they demand a lot of detailed instructions in the prompts such as:

• the purpose and approximate length of each section,
• style (e.g. narrative or bullets),
• tone of voice.

Get all this optimized and we’ve found that GPT can write a decent first draft for you to edit – potentially saving a lot of time, as you don’t have to start with a blank sheet of paper!

More information

If you’d like to learn more about our testing of OpenAI GPT or chat about the application of large language models to product marketing more broadly, please reach out to me on LinkedIn Robin Griffiths or email me at [email protected].