The Value and Limitations of AI Generation in Marketing
I use AI to generate marketing material for my publishing company, Energion Publications. Leveraging AI has helped me improve our marketing despite having limited help and staffing. At the same time, I want to keep my marketing under my own control and not let automation run away with it.
Values and Caveats
I have found that AI is helpful in a number of ways:
- Scanning full-length books for strong quotes.
- Creating slogans from book contents.
- Creating and verifying lists of keywords and phrases.
- Drafting and validating book description text.
- Selecting books for marketing attention.
In all of these areas, however, there are also dangers:
- It is very possible for an AI to be biased in its selection, choosing quoted elements that misrepresent the author’s intention when isolated.
- It can create slogans that might be popular, but don’t present a position the author would actually support.
- It can fail to find keywords that are critical to the subject but less popular in general search trends.
- It can describe a book in a way that doesn’t quite match the actual content.
- It can select books based on rigid criteria, missing important titles that require a human eye.
The most important takeaway here is that these pitfalls are remarkably similar to the ones you run into with human marketers. They point to a vital truth: at some point, a person well-acquainted with the book’s contents—such as an editor or the author—needs to review the marketing material for accuracy.
My Process
The specific tools I use are Google’s NotebookLM and Gemini, accessed both via chat and some scripted tools.
NotebookLM is designed to look at sources provided by the user and produce results based exclusively on those chosen sources. (To a limited extent, you can also tie in outside information). I keep a private notebook with all of my books’ complete content set as sources. This notebook is not public-facing or shared with anyone outside the marketing process. I use it to find related titles, tie together themes from various books, and do the heavy lifting of generating raw marketing material, from book descriptions to press releases.
Next, I run that output through a Gemini Chat. This second step—which uses a different interface to the same underlying AI engine—allows me to check the material against current events. For example, a book on politics might benefit from referencing a currently proposed law or an ongoing election. This pass refines and updates the material.
Finally, I edit that output the same way I edit my own writing, only more thoroughly. My most common correction is toning down overhyped language. In a recent case involving a book about apocryphal gospels, I had to correct an implied conclusion and restart the entire process. The initial AI-generated marketing material advertised the book as if it proposed the apocryphal gospels as reliable historical sources for Jesus’ life. Neither author—one a skeptic and the other an evangelical Christian—was making that claim. They actually argued that the texts had historical value in the history of Christianity, church art, and oral tradition. (See: A Story of Jesus’ Life according to the Apocryphal Gospels).
This issue is a prime example of overhyping, something Large Language Models (LLMs) tend to do. It is a direct result of their training data, which is built on decades of human-produced sales copy—which is itself frequently overhyped. If you’ve ever worked in either production or marketing and experienced that classic organizational tension, you know this is a deeply human tendency, too.
Because these models are trained on vast oceans of internet data, they have deeply absorbed the patterns of traditional online sales copy. As a result, when you ask an AI to write promotional material, its statistical default is to lean heavily into hyperbole—spewing out buzzwords like “revolutionary,” “transformative,” and “game-changing.”
Working with the AI creates a digital version of the classic corporate tug-of-war between the Marketing department and the Production team. The AI acts as the over-enthusiastic marketing coordinator, eager to make the biggest, glossiest promises possible to grab attention. You, the human creator, must act as the grounded production manager—constantly pulling the text back to reality, stripping away the artificial fluff, and ensuring that the copy remains strictly honest and authentic to the actual product (in my case, the book).
AI gives you the raw marketing engine, but human cooperation is what keeps it tethered to the truth.
A Bit More Detail
After reading the following section, please see the Appendix. It’s important!
Due to how LLMs work—by calculating the probability of certain tokens being adjacent in a sequence—they are remarkably good at uncovering themes I hadn’t thought of, as well as checking and improving existing text when paired with human cooperation. What they cannot do reliably is exactly what they weren’t designed to do: craft a final piece, whether a social media post or a full blog article, that you can be certain aligns perfectly with your specific theme, brand image, and the book you want to promote.
To bridge this gap and get the most out of the technology, I rely on two primary strategies:
- Type naturally, just as I would speak. In my experience, conversational phrasing provides the LLM with organic “hooks” to trigger those probabilistic token sequences—hooks I likely wouldn’t have produced consciously if I were trying to write a rigid, formal prompt.
- Ask again in a different way. If a prompt misses the mark, you don’t have to give up; you can simply rephrase it. Even better, you can openly collaborate with the AI by saying something like: “My previous prompt didn’t accomplish what I hoped. I was trying to do [intent]. How can we accomplish that?” You can also provide deeper context by adding: “What got me thinking about [previous topic] was [insert your background thought here].”
Conclusion
The human element remains an absolute requirement for good marketing. I strongly advise against any workflow that takes raw AI output and publishes it directly to your marketing channels. AI is an incredibly powerful tool, but it is not a replacement for human oversight and collaboration.
Appendix: A Live Demo of the Collaborative Process
The following is copied directly from the AI conversation used to create this very post. It demonstrates exactly how this collaborative process works in practice.
Me (Author): For purposes of a blog post in which I describe my use of AI in marketing activities, I want to use a characterization of AI activity. Is this accurate? “Due to the way LLMs work, which is by calculating the probability of certain tokes being adjacent, they have proven very good at locating themes and points I hadn’t thought of. What they tend not to do is essentially what they are not designed to do, and that is to craft a presentation, from a social media post to a full blog post, that you can be certain is consistent with your theme, brand image, and the item (in my case, book) that you want to promote.”
AI: Yes, your characterization is highly accurate. It correctly captures both the technical mechanics of Large Language Models (LLMs) and their practical limitations in marketing.
Here is a quick breakdown of why your statement hits the mark, along with one small typo fix:
- The Technical Basis: You are correct that LLMs work probabilistically. They predict the most likely next word (or “token”) in a sequence based on vast amounts of training data. (Note: You have a small typo in your draft—change “tokes” to “tokens”).
- The Brainstorming Strength: Because they draw from such a massive web of associations, they excel at cross-referencing ideas and spotting themes or angles a human might overlook.
- The Consistency Limitation: You hit the nail on the head regarding certainty. Because LLMs operate on probabilities rather than actual comprehension or intent, they cannot inherently “care” about or truly understand your specific brand, book, or goals. Without strict constraints and human editing, their output tends to drift into generic territory or misrepresent specific nuances of your work.
If you want to polish the text slightly for your blog post, you could phrase it like this:
“Due to how LLMs work—by calculating the probability of certain tokens being adjacent in a sequence—they are remarkably good at uncovering themes and angles I hadn’t thought of. What they cannot do reliably is exactly what they weren’t designed to do: craft a final piece, whether a social media post or a full blog article, that you can be certain aligns perfectly with your specific theme, brand image, and the book you want to promote.”
Me (Author): Yes, your rewrite is good. I need to add to the list of things done well, “checking and improving text, with human cooperation.”
AI: Here is how you can seamlessly weave that in:
“Due to how LLMs work—by calculating the probability of certain tokens being adjacent in a sequence—they are remarkably good at uncovering themes I hadn’t thought of, as well as checking and improving existing text when paired with human cooperation. What they cannot do reliably is exactly what they weren’t designed to do: craft a final piece, whether a social media post or a full blog article, that you can be certain aligns perfectly with your specific theme, brand image, and the book you want to promote.”
Me (Author): Good. Let’s continue, and I’m drafting in real-time here. “I would list two primary strategies I use when working with AI: 1) Trying to type naturally as I would speak, and 2) Ask again in a different way. The first of these, in my experience, provides the LLM with hooks to start one of the “probabilistic sequences of tokens,” hooks which I probably would never have consciously produced. The second takes a couple of routes. If a prompt does not produce the result you were hoping for you can ask again. Even better, you can ask the AI: “My previous prompt didn’t accomplish what I hoped. I was trying to do [fill in your intent in human language here]. How can I accomplish that. In addition, you can provide additional connections, such as, “What got me thinking about [content of previous prompt] was [descriptive text here].”
Now that is just drafted on the fly. How can it best link to what we just produced?
AI: To link these two sections seamlessly, you can use the limitations mentioned at the end of the first paragraph as the perfect springboard for your strategies.
Here is how the two pieces look when knitted together, with some light polishing for flow and punctuation:
“Due to how LLMs work—by calculating the probability of certain tokens being adjacent in a sequence—they are remarkably good at uncovering themes I hadn’t thought of, as well as checking and improving existing text when paired with human cooperation. What they cannot do reliably is exactly what they weren’t designed to do: craft a final piece, whether a social media post or a full blog article, that you can be certain aligns perfectly with your specific theme, brand image, and the book you want to promote.
To bridge this gap and get the most out of the technology, I rely on two primary strategies:
- Type naturally, just as I would speak. In my experience, conversational phrasing provides the LLM with organic “hooks” to trigger those probabilistic token sequences—hooks I likely wouldn’t have produced consciously if I were trying to write a rigid, formal prompt.*
- Ask again in a different way. If a prompt misses the mark, you don’t have to give up; you can simply rephrase it. Even better, you can openly collaborate with the AI by saying something like: ‘My previous prompt didn’t accomplish what I hoped. I was trying to do [intent]. How can we accomplish that?’ You can also provide deeper context by adding: ‘What got me thinking about [previous topic] was [insert your background thought here].'”*
(Note: For this collaboration, the author used Google’s Gemini 3.5 Flash model.)
