How I'm Using AI in Publishing: Amazon Vendor Central's Hidden Gem

One of the hidden metadata gold mines on Amazon Vendor Central is the giant data dump of search terms it provides each week. This is a beast of a CSV file (you can download a filtered book-only version here) that contains over 1 million rows of very rich search and sales data across the entire U.S. site. But in its raw form, this CSV is just too dense to be efficiently usable (and, hence, a perfect use case for AI).

ChatGPT’s answer to a request for relevant search terms for our new adult coloring book using the Amazon Vendor Central data.

Training ChatGPT on this list is as simple as uploading the file (in this case I used a TXT version) and asking questions of it.

While this is a rudimentary use case for AI and this type of data—and there are plenty of paid sites that can provide you with even more useful analyses—I still think this is one of the most exciting windows into what AI will help editors and marketers do in the months and years to come.

For example, if you go back to that first query about coloring books for adults, you’ll notice that “halloween coloring books for adults” is currently ranked #30,971 in all search terms for all products on Amazon. Depending on when you’re reading this, that might not sound impressive; but right now, it’s August and it’s currently 93 degrees in NYC (yes, I know Michael’s is probably decked out for October 31 already). To this marketer, late October just isn’t on my radar and, so, I find it invaluable to our company to tap into objective consumer data rather than rely on my own ignorance. And when AI gets powerful enough in the years to come, I will be feeding these massive datasets to an AI assistant every week and asking it to identify search terms that are just now starting to trend—long before human marketers, like me, think of including them in metadata or placing advertising buys on customer interest that still seems several months off.

By combining multiple weeks of search ranking data, we can see how “Halloween coloring book for adults” trends over time. Based on the 2023 data, the best time to push discoverability will be mid-September.

(If you are looking for how to download Amazon Vendor Central’s search terms, here’s a video)

In Defense of Publishing’s Alternative Business Models

I want to preface this by saying I know that this conversation is old and has gone in endless loops; but I spent last weekend thinking a lot about the ways in which publishers can structure how they work with authors. The publishing houses I oversee, Ulysses Press and VeloPress Books, are considered traditional indie publishers—we pay our authors an advance before publication and then royalties on net revenue as the author’s book sells. There is nothing innovative or interesting about this business model.

What irks me is how stodgy and rigid traditional publishers are in the face of more creative business models (with the emphasis on business). We are, after all, in business to make money and provide for our employees and ourselves.  

Getting Authors Lost in Semantics

If you break down the traditional publishing model into more mundane business terminology, this is what it looks like:  

An advance is an interest-free loan that, technically, must be returned if it isn’t earned out. (Traditional publishers’ reluctance to “claw back” unearned advances has less to do with our magnanimousness and more to do with failing to be good fiduciaries to our own businesses.) Moreover, since the average advance is now somewhere around $5,000, that means that our authors are much better off getting a job at Starbucks than working on a book. That $5k boils down to getting paid for eight weeks of work at minimum wage (NY). If you know a talented author that can write, edit, revise, and market their book in that amount of time at $15/hour, please contact me immediately.  

Royalties, on the other hand, amount to a small percentage of deferred revenue paid once every quarter or six months. Very few people would sell their business or personal IP in other industries for an interest-free loan plus a promise of deferred revenue (with no guaranteed minimum). But traditional publishing puts this model on a pedestal.

Embracing the Alternatives

Hybrid, profit-share, author-services-bundled-with-traditional-publishing (or whatever you want to call the alternatives to the traditional model)—they are just different ways of structuring a business, making money and providing for a company’s stakeholders. Yet if you try to nominate a book for an award as a hybrid, you are required to voluntarily stigmatize your author’s book as coming from something other than normal.  

In the hybrid model, authors aren’t paid an advance and are asked to share in both the costs and the rewards. Some of the best publishers I know are hybrids. They take enormous care in producing, printing and marketing their authors’ books. And their authors are rewarded for their work, risk and investment. In return, these hybrids get to produce exceptional books without driving themselves out of business. And, yes, there are predatory hybrid publishers; there are also plenty of bad actors on the traditional side as well, all of which do enormous disservice to their authors’ writing careers.

Considering hybrid “less than” is a slander used by well-established traditional publisher types that were afforded the opportunity to build large backlist cushions during decades (or centuries) when printing, rent, distribution, and overhead were all cheaper and selling books was easier. In my opinion, thumbing their noses is a defensive tool aimed at preventing new competitors from entering the market and taking their authors or challenging their titles’ dominance.  

But the reality is that there is no right or wrong way to operate a business—just successful or not. In fact, iterating on the traditional publishing model with common business strategies could be a boon for both authors and publishers. For example, if, instead of an advance and royalty, a publisher offered their authors equity stakes in the whole publishing house, those authors would have the opportunity to profit from shares (albeit slowly diluted over time) in a vibrant company rather than just holding a quickly degrading stake in a single product. In exchange, publishers would immediately cut acquisitions costs and royalty payouts and have more cash on hand to put into salaries, project execution and getting books to customers. There are myriad ways to operate in this business that very few companies have tried out.

Innovate or Die

Since 2010, the median cost per unit at Ulysses (including freight) has risen 75%. Our median list price has risen 3% over that same period. That’s not sustainable. (Sponsored self-plug for Perfect Bound—the driving reason our COGS dropped in 2023-2024.)

Publishing books is not going to get easier. There is endless competition for our readers’ attention. There are fewer profitable bookstores and retailers willing to sell books. And the cost of goods has steadily increased while list prices have stagnated. It is time to put the absurd barriers, distinctions and arguments to bed and accept an equal and “all of the above” approach to making books. That is how our industry to survives.

Search Generative Experience Arrives at Amazon

The AI tool I’ve been expecting for a year has come to Amazon and its has been christened “Rufus”. In September of 2023, I discussed this eventuality on Publishers Weekly’s webinar, “Artificial Intelligence: Revolution and Opportunity in Trade Publishing” hosted by Thad McIlroy. You’ve undoubtedly already seen generative search in action on Google and other search engines. Look up “how many publishers are in the United States?” and the first thing you get is an AI-crafted “no-click” result eliminating the need for you to browse of the website results below it.

Search Generative Experience delivering “no-click” answers on Google

Generative Search is Already Impacting Online Industries

It’s estimated that Search Generative Experience (SGE) may lead to a 40-60% decline in organic traffic for many websites[i]. Right now, SGE appears to already be negatively impacting industries where no-click results are most common. One prime example is the beauty industry with SGE answers being provided for nearly 95% of user queries[ii].  Allure.com, where nearly 80% of website traffic comes from organic search, has seen a 20% decline in organic click through visitors over the past 3 months alone[iii]. For online companies that depend on advertising, online sales, and affiliate clicks, the decline in free traffic will have a significant impact on revenue.

The Potential Impact of Amazon’s Rufus

Amazon’s AI shopping assistant is not geared toward answering questions but toward improving purchase conversion. In theory, putting AI toward improving customer experience and conversion sounds harmless to the publishing industry. But if you question what Amazon’s goals might be in the long run, both sides don’t necessarily come out winners.  

Examples of results given by Amazon’s new beta version gen-AI search assistant, Rufus.

Gauging Amazon’s Incentives

Amazon potentially has a lot to gain from switching from its traditional search experience to one that is steered by ever-more-intelligent AI models. Here are some of the considerations:

1.        Better Customer Experience/Conversions: Amazon’s number one goal is to ensure that when a customer visits their website, they make a purchase. Until now, that has meant surfacing the closest products that will satisfy a search (with a whole lot of ads in between). The result rankings are thanks to a complex algorithm that takes into consideration everything from keywords and customer reviews to sales volume.  Rufus’ goal is to eliminate all the scrolling and comparing a customer must do and deliver a highly customized result with just three options.

2.        Streamlined Operations: If Rufus can home in on the mostly likely purchase every customer will make, that means more predictive inventory planning, smoother fulfillment, and, potentially, fewer SKUs held in each warehouse. In 2023, Amazon’s fulfillment operations cost the company approximately $90.6 billion[iv]; just a ten percent improvement on those expenses would give Amazon savings equal to the entire market cap of a company like Pinterest. Now that’s incentive.

3.        Positive ROI on AI: Amazon has invested billions of dollars in developing their AI infrastructure, including a $4 billion investment in Anthropic in addition to developing their own chips, Inferentia and Trainium, for AWS. Eventually, these sizeable investments will pay dividends as Rufus and its following generations train on more and more data from the world’s largest ecommerce platform.

4.        Advertising Dollars: On the other side of the scale, however, is Amazon’s mammoth advertising platform, which generated about $47 billion in revenue in 2023[v]. As with Google, it’s yet to be seen how paid search plays a role in a future where AI and search are completely intertwined.

The worst case scenario for the publishing industry is if Incentives 1 and 2 play out in an aggressively negative way. While Amazon has made a killing as “The Everything Store”, it may want to evolve into the “Everything that Actually Sells Store”. In which case, winners and losers in the publishing industry will become rapidly apparent.

The Winners

Category Killers and Bestsellers: Any publishing company that has enjoyed steady sales, great reviews, and consistently high search results for a title on Amazon stands to reap strong revenue going forward. If a book is considered a category killer or was showered with rave reviews or awards, it may no longer have to compete with sponsored competitors or fresh titles. That may end up locking in those titles’ competitive advantage in a certain genre or category for good.

High-Quality Books: The easiest way to make a title highly defensible in the new world of AI-powered search will be to focus on quality. Ideally, AI-driven search would do away with keyword stuffing and search gaming and become a meritocracy of product where the best books win (at least more of the time).

Forecasting: If SGE creates a defensive moat around a publisher’s strongest books, it would make forecasting inventory and sales, both notoriously difficult in publishing, a lot more accurate.

The Losers

Midlist Catalogs: If SGE consistently surfaces only the top books in a catalog, it could severely impact publishers’ midlist books, especially if the titles are in a crowded and competitive category.

Discoverability and Metadata Strategies: Fostering discoverability, which has already become more challenging over the past few years, is going to require even more effort and savvy. Publishers are going to need to work overtime to optimize well-written descriptions, add richer content, drive Amazon purchases from outside, and A/B test constantly to find out what works and what doesn’t. Across large catalogs, that will take enormous resources (although AI will probably help). And what’s more, the rapid evolution of LLMs will require constant vigilance and adjustment from marketing teams as the AI evolves.  

Self-Publishing: Small indies and author-publishers will be placed at a further disadvantage if snagging a position within generative results becomes more difficult and more work. In addition, the era of quick-to-market books aimed at selling through gamed SEO and keywords will slowly fade.

How We are Adapting

Since the early 2000s, Ulysses Press has successfully leveraged data, statistics and market research to acquire on-trend and niche titles. While data is still very much going to inform our acquisitions decisions, our editors will undoubtedly need to become more selective about the titles we invest in as we move into 2025 and 2026. In addition, we will need to pour more resources and time into uncharted Generative Engine Optimization strategies.

Whether all these AI-driven changes come to fruition is beyond me. Consumer habits when it comes to online shopping are stubborn (we’ve been content with the same Amazon user experience for two decades now). But shopping behaviors do evolve—20 years ago I wouldn’t give websites my credit card; now I’ll buy toilet paper by yelling at a cylinder in my living room. I honestly don’t know what shopping on Amazon will look like in five years, but if it’s anything like what I expect, it will upend my business model. And that’s something I’d like to be prepared for.

[i] https://searchengineland.com/how-google-sge-will-impact-your-traffic-and-3-sge-recovery-case-studies-431430

[ii] https://www.searchenginejournal.com/google-sge-organic-traffic-impact-divided-by-verticals/514800/

[iii] Similarweb.com

[iv] https://ir.aboutamazon.com/news-release/news-release-details/2024/Amazon.com-Announces-Fourth-Quarter-Results/default.aspx

[v] https://www.statista.com/statistics/1305698/amazon-advertising-revenue/

How I'm Using AI in Publishing: Automating Ulysses Marketing Assets

The Problem

One of the biggest time-sucks I’ve found in the marketing of entire catalogs of books is producing the visual assets needed for everything from Shopify merchandise images and social media carousel cards to product images for the Amazon A+ Standard Comparison Chart.   These media are critical to allowing your publishing house to maximize title awareness and create free cross-promotional opportunities, keeping readers within, what I call, the “publisher ecosystem”.

But producing these assets normally requires a few minutes each (in Photoshop or Canva). And for a catalog of 1,000 titles, that means about 8 days-worth of work! That project can cost your publishing house thousands of dollars to do manually in house.

The AI Fix

With ChatGPT, we can reduce those 8 days down to about 20 minutes of actual work (and less than hour of the computer working behind the scenes). In this use case, I asked ChatGPT the following prompt to create square Facebook carousel title cards complete with a light drop shadow using Python:

Here is the code it generated:

ChatGPT writes Python code to produce automated marketing assets

All I had to do was match up the name of the folder I was directing it to and click “Run”. The code iterates through that folder, filled with cover images, and quickly produces perfect, uniform social media cards. Check it out:

Python code running and creating assets in realtime.

Examples of the finished products for Shopify and Amazon’s A+ Comparison Module

Of course, I can also have it create A+ comparison cards at the same time, simply by dropping in the additional code for more cards with the different aspect ratios and margins.

Click on the image to download the .ipynb file

The Result

All in all, this project allowed me to build out every combination of assets we might need going forward (and with the uniform naming conventions). It took less than 30 minutes. These are the types of truly startling productivity gains that you can achieve with AI—start with the tedious tasks that are necessary but take a lot of time and AI can pay massive dividends almost immediately.