Media Briefing: Dotdash Meredith’s Jon Roberts on the AI agenda in 2025

Media companies prepare for legal battles with Trump administration, Fortune retracts bogus story and more

Q&A with Dotdash Meredith on AI agenda in 2025

Dotdash Meredith has ambitious plans for the development of its AI-powered technology in 2025. 

After signing a deal with OpenAI back in May 2024, Dotdash Meredith has helped give feedback on the AI tech company’s products and is using OpenAI’s technology to improve Dotdash Meredith’s ad-targeting tool D/Cipher.

Jon Roberts, chief innovation officer at DDM, spoke with Digiday to share what else is coming down the pipeline this year, including expanding its ability to draw connections across content – both internally and across the open web.

This conversation has been edited and condensed.

It’s been almost nine months since Dotdash Meredith signed a deal with OpenAI. What have you learned from working together so far?

We believe it’s important that when AI companies take your content that they pay for the training that they do with your content… [and] when they use our content, it needs to be properly attributed. OpenAI launched SearchGPT. It has clear citations in it for every source that it has. That’s a win… And that’s not just something that happened. We are working very closely with OpenAI on their new releases of user-facing products… There’s more stuff coming shortly that we’re in the weeds on with them. It’s really, really interesting to be close to these things as they are getting developed… It’s great to be able to be in the room, provide some expertise and help.

We moved all of the language understanding of D/Cipher onto OpenAI’s tech in September, which is a pretty quick run to re-platform a large product. All the D/Cipher campaigns that we’ve been running since then – Q4 onwards – have been using the upgraded tech. I think an easy way to understand what the difference is [now that we’re using a language model, is that it] doesn’t work with the words [on a page]. It actually abstracts back to the concept. About 70% of the connections that came when we upgraded it were pretty much the same as the natural language processing version… [It gives us a] deeper, richer understanding of our overall corpus.

What that’s also allowed us to do is take that and go map the premium open web…. That’s step one of the partnership with OpenAI — getting to that base camp milestone of a really rich understanding of content, not just at DDM but more broadly.

What does that mean? How does this process happen across the open web?

That’s our special sauce. We’re not ready to talk about it yet.

Can you give an example of what this looks like with Dotdash Meredith’s own content?

I realized that was kind of abstract and super in the weeds and not the easiest to summarize. 

When we’re looking at Dotdash Meredith as a whole pre-OpenAI deal, the way we run D/Cipher is we look at a topic… and we understand all of the content within that topic [from editorially-tagged] content. Then what we’re saying is, looking at all the things we talked about within this topic, what other topics across DDM are similar and relevant to users with this intent? We build those connections two ways. One way is to say: when we talk about the topic of investing in mutual funds, [for example], we talk about long term, low risk, buy and hold investing. You might not use the word “mutual funds” in our retirement content, it’s not the main topic, but long term, low risk, buy and hold investing is talked about a lot. So these two topics are related because they use a lot of those same words. That’s how we connected things previously. 

[But] people looking at reading about visiting museums when they’re looking at the topic of visiting Rome [have different intent than those looking at] visiting the Pacific Northwest… The OpenAI architecture… will solve things like, these two things aren’t related because this one’s on one side of the world and this was on the other side.

What impact does this have on DDM’s business?

This means we get more high quality opportunities for our advertisers. So what this allows us to do is say… we have a bunch of content that’s directly about caring for a pet… [and] this also overlaps into this other content across the business [and find that] it turns out people reading about stain removal from their carpets are also very highly correlated with being new pet owners. We can [advertise for] pet insurance on Real Simple… The deeper you understand the content, the better you can understand the targeting, the outcomes, and then iterate and innovate on the next phase. And in terms of campaign, from just being something you turn on and turn off to an actual experiment that allows you to test and learn about your audience, about the message and about the moment they’re in when they see it.

Are there any measurable improvements in DDM’s advertising business since implementing this in Q4 last year?

We’re [not in] a position to share [numbers] on the record. We’re smiling, so it’s qualitatively good. We’re very confident. It is better.

What’s the next step now? Are you trying to monetize any of this, this year?

D/Cipher campaigns are in about 50% of all of our direct sales campaigns. That’s probably about the right level. The IO-based campaigns that have this in it are doing great. They are our fastest growing clients by far. They are our biggest clients by far. So a large part of our advertising success and growth over the last year is founded in the rollout, the work, the improvement of the iteration of D/Cipher. We’re obviously not done. There is more to come from this, for sure. 

To what extent are you looking to incorporate other LLMs and do deals with other AI companies? 

As we see it, there’s three different categories of deal conversations going on. One is when people take our content and train on it, we would very much like to get paid for that…. [The second is], ‘We’re a large language business. It’s incredibly useful to use large language models to help us move words around.’ And then the third piece of this, which is all the new startups that don’t have their own language model, but are building the tech off these capabilities. 

For the actual foundational model work, the critical things we are looking for are, what is the quality of output for a particular input? We now have benchmarks that we have set for these types of questions.

On some level, language models are a commodity now. So if it’s cheaper and faster, that has obviously a value to us. [What we’re doing] isn’t real time, user-facing yet. So the speed isn’t critical at that level, just yet. When we can get high quality, but also very, very fast, that starts to unlock some new things that we can think about. So we’re going to cross some of those thresholds this year. Some language models, it can take you eight to nine seconds to summarize a page of content – that is obviously way too slow to be anything user-facing. With other language models, we now can prove that can be done in under half a second. That’s different. The quality is lower, but the speed is higher.

I would believe that within this year, speed versus quality is not going to be an either or any more. You’ll be able to get both. And that’s going to change some things. Change your use cases, make them more usable. 

What are some of the languages models that are doing a good job of this now?

I’m not just saying this because we’ve partnered with them, but GPT-4 is still the best large language model. It is a race.

What else are you trying to prioritize this year?

The real priority is getting all of this intelligence into D/Cipher. A really interesting opportunity is to go beyond just understanding the content, but also understanding the ad creative. If advertising is the right message, right person, right time, the content tells us right person, right time. Because we knew you were this person reading this thing at this moment, the better we understand the content, the better we understand the moment you’re in. With the ability to put media in and interpret it, you can then actually also get a good understanding of the ad creative. 

This isn’t like dynamic ad creative, it’s just saying: when you ran this ad against this person, in this moment, on this content, it worked because this message was really relevant to that person. But when it worked right over here, it was not as relevant and maybe if we tweaked it, this message would actually be a better message for that person at that time. Up until this point, we haven’t had a system to understand creative at scale against the content that it’s running against. Now we have beginnings of those capabilities, so we are actively working on understanding how to bring that level of interpretation and insight into our biggest partnerships.

Do you think that’ll happen this year?

That’s a hard prediction to make. We’re working on it.

What we’ve heard

“We’ve anticipated this. We’ve been planning for this. It’s something we thought was a distinct possibility and we have been telling our audience the other places where they can find us.”

– Micah Gelman, The Washington Post’s head of video, on the impending TikTok ban.

Numbers to know

400: The number of Washington Post journalists who are asking owner Jeff Bezos to swoop in and fix leadership problems at the paper.

$5.13 million: The amount The Guardian US has raised in its end-of-2024 fundraising campaign, up from $2.2 million at the end of 2023

87%: The percentage of digital media leaders that said their newsrooms are being fully or somewhat transformed by generative AI, according to a Reuters Institute report.

82: The age of Barry Diller, IAC chairman, who will take back control of the company that owns Dotdash Meredith as CEO Joey Levin exits.

45%: The amount the subscriber conversion rate fell when The Los Angeles Times lifted the paywall on wildfire-related stories. (Subscriptions were up over 259% compared to the past 30-day average.)

What we’ve covered

OpenAI, The New York Times debate copyright infringement in first trial arguments

• A judge listened to arguments from lawyers representing The New York Times and OpenAI, in a federal court hearing that kicked off the copyright infringement lawsuit trial brought on by the publisher.
• The lawyers debated “fair use” and “hot news” doctrines, as well as the process by which large language models work.

Find the main arguments from the trial here.

Over half of publishers reported revenue increases in 2024

• A Digiday+ Research report found that well over half of the 50 publisher professionals surveyed said their revenues were up at the end of 2024.
• That’s better than the prior year, when just over a third of publisher pros said their revenues were up in 2023 compared with 2022.

Read more from the Digiday+ Research survey here.

Five trends in the AI space in 2025

• Industry experts predict five trends in the AI space in 2025, including generative AI’s impact on search, an adoption of AI content production tools, and more open source models
• But the growth of AI implementation comes with a need to maintain consumer trust

Read more about the five AI trends here.

What we’re reading

Media companies prepare for potential legal and political fights with Trump administration

Media companies are preparing for potential legal and political battles from a second Trump administration, The New York Times reported. Some national reporters are increasingly using encrypted messaging apps like Signal to protect themselves, and some organizations are checking their libel insurance coverage.

The Washington Post deals with financial and internal turmoil

The Washington Post is reeling from a $100 million revenue loss last year, leadership challenges and a frustrated staff, according to The Wall Street Journal’s reporting. In order to reach an ambitious internal goal of an audience of 200 million, executives are discussing AI tools and news aggregation.

HuffPost editor-in-chief Danielle Belton resigns 

HuffPost’s top editor Danielle Belton is leaving the BuzzFeed-owned publication after four years at the helm, according to TheWrap. At least 30 HuffPost newsroom staffers were laid off.

Publishers can now livestream on Substack

Substack is making livestreaming capabilities available to publishers, after testing live video features last fall, TechCrunch reported. In the shadow of an impending TikTok ban, the feature lets users host live video if they have at least 10 free subscribers, and the videos can be re-published and shared on other social platforms.

Vox Media mixes layoffs with leadership changes

The publisher of Vox, The Verge and New York has gone through multiple rounds of layoffs in the last two months – including a third on Monday – and promoted Pam Wasserstein and Ryan Pauley to co-president positions, according to Business Insider.