Strategic publishing, AI, and the battle for heart and minds
Where commercial publishers have to make money from their books, strategic publishers like me have to make an impact. Our books are in a battle for hearts and minds.
And whether we like it or not (I certainly don’t), that battle is increasingly won or lost in AI chat windows.
More and more, people ask AI questions and get the answers they want without ever visiting an actual website. They’re even less likely to read a PDF or a book. AI is their oracle, and most of the time AI users won’t go anywhere else.
An illustration from The Rocket Book (1912) by Peter Newell. Aren’t we all these folk, watching AI tear like a rocket through everything, not quite sure what just happened?
So, as strategic publishers, how do we make sure that it’s our ideas appearing in those chats, and not someone else’s ideas? Or, worse, not some malformed notion cooked up by AI in a pathological drive to please?
The answer is simple, but it’s not cheap: we have to publish writing that is great for humans – clearly and creatively written, well-structured, carefully edited, cleanly presented, interesting, novel, accessible, and trustworthy. And we need to publish it on the open web as well-constructed HTML pages. That kind of content is an AI trainer’s favourite kind.
Fresh content
AI companies have stopped at nothing, least of all the law, to hoover up every bit of good writing they can find, because that’s what makes AI worth anything at all. They’ve triaged and ingested it all to feed their magnificent robots. And they absolutely can’t afford to stop now. They also can’t afford to feed AI’s own writing back to itself over and over: the models will collapse. They need fresh content, rich with human ingenuity, to keep their models alive.
AI companies are making deals with commercial publishers to license new content, so AI training on their books will continue. But strategic publishers have different incentives, and most lack the scale that makes such licensing worthwhile. As strategic publishers, we usually want AI to use our content and repay us in impact, not money. (Expecting money is nice, but can distract us from the impact that matters most.)
HTML, not PDF
When AI trainers go looking for fresh sources beyond what they’ve licensed, the writing they find first and ingest most easily is on the open web. Importantly, the low-hanging fruit is on HTML web pages, not in PDFs. Even mediocre HTML pages are naturally structured and tagged in ways that help AI engines understand their meaning and significance.
We don’t just want AI to use our content, but to use it well. Publishers who’ve been focusing on accessibility – the work of making their pages readable by people with disabilities – are way out ahead here, because accessibility features are enormously valuable to machine learning: things like alt-text descriptions of images, SVG diagrams, properly organised heading levels, and role tags. While that’s possible to some extent in PDF, once you’re putting in that amount of work you might as well be using HTML.
Without that extra work, PDFs are just floating blobs of text that make sense only to humans who can see their visual layout. To machines, as to people with visual impairments, they are mostly devoid of structure and context.
For a long time it’s been fine – not great, but fine – to publish research, white papers, booklets, and entire textbooks as PDFs that people download. Those PDFs didn’t get a lot of visitors, and who knows if anyone read them, but they satisfied our funders, and didn’t have much competition. Now every pocket-sized AI oracle is the competition: they are eating what little traffic those PDFs could attract, and they won’t even leave us the crumbs.
Be the best source
While only some people will click through to a website after AI answers their question, that traffic is still invaluable to publishers. AI services are forced to offer users those links – by user expectations, regulation, and competition. At the moment, their attempts to link AI answers to sources are disappointing: if I click from a Google-AI answer to its supposed reference, that reference is rarely as appropriate as Google would have me believe. But there are serious efforts underway to improve linking to sources, and to reduce hallucinated references. Assuming those things will improve, strategic publishers really need AI users to be sent to their content. And that’s most likely to happen if the source is a well-constructed, book-quality web page. It’s certainly not going to be a PDF.
As I said, the answer is simple but not cheap. It does cost more to produce excellent web pages at book-publishing quality than to lay out a PDF and make it downloadable. And it takes careful thought, skill, and effort to show to AI and to humans that your content is trustworthy. But, to paraphrase that old adage: if you think publishing web books is expensive, you should try not publishing web books.
Is this really, now, finally, the end of PDF? The end of print? Heavens, no. PDF and print are still good at all the things we love about them. It is the end, however, of pretending that they’re enough for serious strategic publishing.
No AI was used in the writing of this post. I’m especially grateful for George Walkley’s weekly newsletter as a source of publishing industry news on AI. A number of the links in this piece come from it. This post was first published on electricbookworks.com.