User Needs Model in the Age of AI Agents

Branislava Lovre:

Welcome to AImpactful. Newsrooms have always had to think about what their audiences need. But now there is a new layer. AI agents can find, summarize and deliver content to people. So what happens when AI systems become part of the connection between newsrooms and their audiences? And how can media make sure their content keeps its value when AI systems use it?

My guest today is Dmitry Shishkin, who helped bring the user needs model into newsroom practice and is now developing the agent readiness layer. Dmitry, welcome to AImpactful.

Dmitry Shishkin:

Thank you, Branislava. Thank you for inviting me.

Branislava Lovre:

Dmitry, it’s been ten years since the BBC World Service research that led to the user needs model. You were the person who brought that research into everyday newsroom work at the BBC and later helped introduce the approach to newsrooms around the world. So let’s start with the basics. What exactly is a user need, especially for someone hearing this term for the first time?

Dmitry Shishkin:

Well, a user need, put simply, is a person’s motivation to consume a piece of content. We know that people can be interested in different types of topics, but they can also be interested in different angles of the same topic. User needs are nothing but angles of storytelling, but nobody had thought of actually capturing the impact of those angles before, in the form of numbers, in the form of KPIs. And we have done it.

And I would say that this is intentional. It’s probably the most important thing to say about user needs. It’s that it captures the intent of a piece of content before the content is consumed. So somebody commits to a piece of content knowing exactly why they are consuming it, not what they are about to consume.

Branislava Lovre:

That original research found a clear pattern: around 70% of the content was update me content, but it brought only a small part of the traffic. So if we’ve known this for ten years, why do newsrooms still produce so many short and often very similar news stories?

Dmitry Shishkin:

Well, humans are creatures of habit, right? So it’s very, very hard for them to change the way they do things because they have done things a certain way. They have been trained in a certain way. And I’m really happy that a lot of universities, a lot of journalism schools are now starting to teach user needs models. So it kind of goes right at the heart of the journalism training. So habits will always beat insights in the same way that culture will always beat any structures, any strategy or anything like that.

So I always will say that the most important thing about this is that our metrics have for many, many years rewarded volume. So you always needed to do bigger things, reach more audiences, kind of make sure that you had the biggest reach possible, but in the future, with an AI-mediated world and commodity content, reach is not going to be the most important thing. The engagement metrics, the direct connection with audiences will be the most important metrics for the future.

So we need to slowly… well, we actually need to move fast from the old age to the new world. And we really need to make sure that the change that we bring into newsrooms with the introduction of user needs models is supported by some kind of organizational structures, by the governance, by the workflows, by all of those types of things. And because of that, successful newsrooms are not the ones that know about user needs, but successful ones are the ones that are able to actually implement and start using it.

Branislava Lovre:

Let’s imagine that a major story has just happened and every news outlet has already published the basic facts. What should a newsroom do next? Should it develop follow-up stories for different audience needs? Focus on different distribution channels or both?

Dmitry Shishkin:

It’s actually both because you can’t really achieve a lot by just doing one thing. If something breaks, you immediately need to start thinking about what’s next? What is the follow-up story? The follow-up story is dependent on the topic that you’re covering, on the channels that you’re going to distribute the story, and crucially, on historical use of different user needs on those specific channels. And I would also say another dimension: a lot of people are ignoring formats. So effectively you’re saying what we’re covering, how we’re covering, where we’re distributing it and user needs play completely kind of central role in that because you kind of always, always the default position is that one event, many angles, many angles equals user needs, different user needs perform differently in different contexts. And you really need to be aware of that.

And that’s why I think in the future and we’ll talk about a little bit later, algorithms will make you much more efficient and effective in making those decisions because you don’t need to remember and keep all of that information in your head saying, you know, for this particular story, these particular three types of user needs really work very well. But if you start looking at distribution strategies, time of day, platforms, whatever, all of those types of things, user needs change as well.

Branislava Lovre:

You say that the user needs model works best when it becomes part of everyday newsroom work, editorial meetings, story planning, tools, workflows and the way a newsroom measures results. So how can newsrooms introduce this in practice? And what would a basic version of this system look like for a media organization?

Dmitry Shishkin:

Well, firstly, don’t just workshop it. Embed the workflow into the newsroom because it’s one thing to say, let’s say, you know, Dmitry comes in for an hour and introduces the user needs model to you and that’s it. And then kind of assume that everything is already remembered and people are going to use it. Of course not. They’re not going to use it, so embed the new ways of working in your workflow, whatever it is. And you rightly say this is about structures, this is about tools, this is about ways of working, all of that.

Then obviously you need to start with a meeting where you constantly say, okay, well, instead of discussing what is being prepared for publication, you need to talk about how you are preparing to do it. So you always say, we will not do something about this topic, but be specific, i.e., when I was at the BBC, I was asking, and now with my clients I’m asking people to be as specific as giving me a headline because only from a headline we can actually see the signal about particular user needs here and there.

And then the obvious thing, something that you mentioned as well, and your question is about measurement as well. So you measure what changes. So you always will say let’s do it at the level of one topic for an organization. For example, let’s do it for a month and let’s see what happened during this month and let’s see the results after the month and then what was there before?

Branislava Lovre:

We’ve been talking about the user needs model, data and results, but AI agents add another layer. More people may get their news through AI agents that find, compare, summarize and deliver information for them. What does this change mean for newsrooms?

Dmitry Shishkin:

Well, imagine a beautiful situation where your newsroom has already adopted the user needs model to the best of its ability and you’ve optimized your content and you know that you are not covering content with default user needs, but you always take very specific decisions about what kind of content to create and everything else. And then also, let’s imagine a really beautiful situation where 80% of your audience pays you to consume your content directly. So you have a really strong relationship with individual people and they consume. But that also leaves 20%, you know, the Pareto principle, about 80/20. 20% of content will be consumed elsewhere one way or another and then that ‘elsewhere’ will be changing very, very quickly because of, you know, bots first, then agents. And agents are kind of far more intelligent bots, in terms of the time, they’re actually going to be executing the will of a human being.

So imagine even those 80% of people who are going to be consuming you directly. Let’s imagine they will not come to your website directly or they will be asking the agents to come and synthesize information and prepare something for them and everything else. So the question is, you might be doing an excellent job in creating editorial content and then agents come. And because your structured data is not good, then the agents will not be able to capture the most important things that they need to capture.

So because agents need structure, your editorial content, your editorial output needs to be structured as well. So I’m really talking kind of going deep into the taxonomy and your metadata and kind of creating those signals to the agents about the intent not only of every single article, because it’s not going to be enough, because imagine if your article is ten paragraphs and maybe paragraphs five and six are going to be exactly answering the question, exactly what agents came here to do. So if you’re exposing them to a ten-paragraph article, they might ignore it because they might think, well, I don’t have enough tokens to go and process all that information and everything else. So what can newsrooms do? How can they make their content work better with AI agents? We as media need to help agents in the future. If your task is task A, you want to cite or find something that is not going to be found elsewhere: original content, for example, quotes or, you know, provenance, all of that type of information. It’s our job to connect that information with what agents are looking for. So structure is very, very important.

The intent, as I said, needs to be very explicit, almost on a paragraph-by-paragraph basis. And that’s why I came up with this idea of agent readiness layer. It needs to become visible to machines in terms of how machines are going to process your data. So it’s not only about what you are putting together when you’re publishing content, but there has to be something done at the level of the whole organization. And this is what makes me very, very excited about the future because, you know, it’s one thing to create user needs for everybody else. It’s now another challenge to capture that data from user needs and make sure that it’s not ignored by agents.

Branislava Lovre:

You’ve just mentioned the agent readiness layer you’re developing. In your article, you describe it as a way to make sure the value and purpose of the content are not lost when AI agents find, summarize or transform it. Can you explain a little bit more about what this layer does and how it works together with the user needs model?

Dmitry Shishkin:

I spend a lot of time creating content, and I create a lot of nuances in my content and a lot of really interesting things. My anxiety would be that content will be either ignored because, you know, sometimes if my data is not clear enough, then the agents will just go elsewhere or it will not be ignored, but it will be flattened. And then out of ten paragraphs they will take one paragraph which is not going to be the most important one or the most original one or the most exclusive one. The article which I wrote about this actually asked the question whether AI agents have needs. And I don’t think that they do. But what they do have is they have a requirement to capture human user needs correctly and adequately. And because of that, I think we need on the media side, we need to do that work to make sure that this is as clear and useful to agents as possible.

So this is not a replacement for the user needs model at all. This is a complementary system. In the past, it was human-to-human consumption. Now it’s probably, you know, business-to-agents-to-customers type of situation. But in the middle, where the business needs to pass the information to agents before it is passed to humans, it’s our responsibility and duty to make sure that our editorial value is preserved.

Branislava Lovre:

In your framework, you explain that AI agents need to find the right information, check it and compare it with other sources. So what does that mean in practice for a real news article? What needs to be clear in the article itself?

Dmitry Shishkin:

I don’t think that a human necessarily needs to do anything on the media side of things, because I think that if we say that humans need to do another kind of, you know, selection from some kind of drop-down menus, you know, in your CMS that will be too complicated. It probably will have to be media-side agents. So agents who actually will interpret the content and they will say, you know, we have unpacked the content into different components and each component has its own signal. And because of that, I think what we really need to understand is that we need to be explicit about our journalism. Every single quote needs to be attributed. It needs to be super well structured about what it is and what it is not. And because of that, you kind of need to always say, okay, sources are named, that’s the thing, but if the source is something that you really consider the most important in your article, you really need to highlight that somehow maybe on the metadata level, maybe on the level that is not seen by real people or anything like that.

But, you know, behind the scenes you really need to say that, out of those ten paragraphs, paragraphs three and four are the most important ones because this is where original quotes are from and this is where journalists actually spoke to real people and everything. So that’s one thing.

If your article is written in an unclear way, so when your, you know, claims and counterclaims are not really clearly written, then there is a risk that agents will start flattening the content or start ignoring content that is not clear and it’s not clear what it is for. So you kind of need to have that very clear.

But at the same time, if there is uncertainty about something, you really need to be also very explicit about that because agents might come and say, I know that there is no real answer about this, but I know that I was well informed about that uncertainty. And because of that, I captured that information.

Branislava Lovre:

You’re describing an article not only as one whole piece, but also as different parts: facts, claims, quotes, numbers and other elements. So what kind of checks should happen before an article is published?

Dmitry Shishkin:

Well, firstly, you probably will need to consider some kind of guidance at the level of each organization where every article that is submitted goes through some kind of filtering, where the copy is then given back to the editorial team to say that, before we submit it for copy editing or for publishing, several things need to be clarified, for example, you know, and so you kind of need to always be very, very clear.

Like, for example, I’ll give a very, very basic example. If somebody has been asked to create an educate me story about something, and they bring the story and the educate me part is less than 50%, then the article cannot be codified as an Educate Me piece because it will have lots of other user needs inside and everything else. And let me be very clear, there is no such thing as a 100% clean, one-user-need article. You always will have a mixture, but one dominant user need needs to be over 50%. And that’s really important.

And AI can help us very, very nicely here because once you are already working with your content management systems, you paste your copy inside the content management system and agents trained on your guidelines and everything else will say, well, according to you, you are writing an educate me article, but it’s not an educate me article right now. You need to go and add some educate me content there. And this is just one example, but you kind of get my drift.

Branislava Lovre:

You’ve already talked about how important good metadata and well-organized data are for newsrooms. You’ve also written about a simple example: a two-minute vertical video for TikTok and a 45-minute interview for YouTube can both be labeled simply as video in a CMS. Why is that a problem? And what should journalists and newsrooms pay special attention to when organizing their data?

Dmitry Shishkin:

I’m really, really glad you asked me that question, because the more I work with user needs and the more I work with topics, I have always been very, very clear that if you start comparing the performance of articles according to topics and user needs, you already will achieve a lot. So you will say something like a business story will do better than an entertainment story. That is a very basic thing. We have done this for 25 years. No, nobody cares. If you say a business story that has been created as an educate me piece does better than a business story created as an update me piece, maybe that’s even better.

But once you add the third dimension, i.e. a format, you will see a business story created with an educate me user need but delivered as a first-person interview, in comparison to, say, a Q&A, a listicle or, I don’t know, a feature. You know, you start comparing different formats and the formats are, you know, there are probably somewhere between ten and 12 formats everybody’s using. And when I’m talking about formats, I’m not talking about audio, video or text. I’m talking about specific formats like an interview, write-up, explainer, opinion piece, data story, you know, all of those types of things.

And the reason they are so kind of content-type agnostic is that they can be delivered as video, they can be delivered as audio, they can be delivered as text. But you really need to understand what it is. And because of that, segmenting your content according to formats, according to user needs and according to subtopics of a topic that you cover is very, very important because it’s not enough to say this is a video about entertainment.

Like what? Okay, this is maybe an interview delivered as a sit-down, long-form video and maybe it’s inspire me content about music, not entertainment as a whole because I don’t know if you tell me that this story is an entertainment story, I don’t know. Is it about TV series? Is it about books or is it about opera?

Branislava Lovre:

All of this is about making news content clearer and easier. We’ve already seen publishers that publish versions of content. So is there a risk that while trying to make their work easier for AI systems to use, newsrooms start adapting to machines instead of people?

Dmitry Shishkin:

No, I don’t think so. I think that, well, kind of yes and no, because on the one hand, we talked about, you know, structured journalism, about writing much more cleanly and writing much more intentionally, but a lot of really good and important work will be happening behind the scenes. It will not be seen by a human, but it will become available to agents, kind of, you know, under the hood, so to speak. So it’s really important to remember that.

Branislava Lovre:

There is also a business side to this. If a media organization blocks AI crawlers, it may become less visible in AI systems, but if it gives them wide access, it may give away valuable content without enough control or payment. So how should media organizations think about this balance?

Dmitry Shishkin:

Well, you probably always will be quite selective about your choices. So it will not be a binary choice between, you know, black and white, yes or no. So it will be, you know, we allow some crawlers. We don’t allow some other crawlers. The issue here is that you still need to default to the audience coming to you being number one. So nobody has canceled that major task of saying we actually need to have person-to-person first-party data relationship between whoever consumes us and us. But we’re not living in the perfect world. You always will. As I said at the beginning, you always will have, if you’re lucky, an 80/20 type of situation where you always will need to be nuanced.

I applaud, you know, initiatives like SPUR, for example, the SPUR Coalition, where lots of media organizations come together specifically to start negotiating together with large language models and kind of understand what’s kind of the most important components of them. And actually, the agent readiness layer is something that SPUR really needs to consider because the thinking is already there. But instead of doing it for one, we could do it for everybody else here. It’s also the question that if you are negotiating, you know, being paid by one of the large language models, well, firstly, I’m quite suspicious about those deals and about those deals lasting and staying with us in the future and everything else. I’m really worried about single-language deals. So if they strike a deal with one company, one media company covering one language or representing their content in one language, then it’s kind of, you know, there is no point for them in actually doing any deals with anybody else in the same language.

Maybe English is going to be an outlier because they just need as much stuff from English as possible. But if you are going to be a single-language country, and if you are a media organization in a single-language country, then I would be very worried about how useful it is. How realistic is it that you are ever going to be earning any money from that? So I’m worried.

But what I’m less worried about is that, once we prove to people that different types of user needs are more valuable to people, because we have proven many, many times that if you are doing it, if you are looking at conversions, if you’re looking at engagement, if you’re looking at reach, forget about reach for the time being. But if you’re looking at engagement and conversions, then your content, the different user needs, can convert differently. So it’s not enough to just do content without user needs. User needs will help you be much more successful business-wise, and because of that, your negotiating power moving forward will be much, much stronger because you can say that, you know, all my give me perspective content and all my educate me content will be five times more expensive than my update me content or something like that. And that makes me excited.

Branislava Lovre:

You’ve just talked about how different types of content can have different value, so do you see niches as part of the solution?

Dmitry Shishkin:

I believe so. And niches, I mean, we already see some examples of, you know, B2B organizations doing very well. And why? Because their audience is very, very committed, very, very specific. It already knows what it wants. It knows why you exist in the market, what kind of job you need to do for them. But you really always need to ask yourself a question. It’s not about the output. It’s about user needs first. It’s about the audience and what you actually do for the audience and how you engage them and what you kind of do for them and whatever.

So I always will leave people with just saying, be very honest with yourself. And out of 100 articles that you publish every day, what is the percentage of articles that are completely replaceable? And I think the honest answer is, again, going back to 80/20, that only 20% of your content, if you’re lucky, is going to be irreplaceable.

Branislava Lovre:

You work with colleagues in newsrooms around the world and we’ve talked about many changes that are already happening. What do you think we can expect over the next year or two in the way newsrooms collect, produce and share content with their audiences?

Dmitry Shishkin:

Well, nichification will continue, for sure. So you kind of really need to be much more precise about what you cover and what you don’t cover. And my litmus test for that would be, or my rule of thumb would be, to always, always have a visualization of the topics that you cover in front of you, kind of on a month-by-month basis. And you can see that if you still have one particular topic and write about it, but only sporadically or not regularly, then maybe the question is: why not stop covering it altogether because it’s just not your thing and somebody else will do it better anyway.

But anyway, cleaning your data and making sure that your data is as robust as possible will set you up for a multimodal world much better because the current mistake with multimodality around media is that people put a button on your website and say every single article can now be read or listened to. And this is not multimodality. This is just a product feature nobody really asked for.

And because of that, you kind of need to say, I understand that this piece of content is about topic A. When we cover topic A for this particular audience, this particular segment of audience, this particular distribution platform, this is the best user need to apply for it. And also the formats to package the content into will be that and that is multimodality.

So you effectively will be saying we publish 100 pieces of content a day, but the constellation between the user needs, formats and topics will be individual. For all of those hundred pieces, there will be a hundred different treatments.

Branislava Lovre:

Is there anything we haven’t talked about that you think is important to add?

Dmitry Shishkin:

No, I think I will just finish by saying that you really need to think about what makes you indispensable, about indispensability. So one of the things I talk about a lot in presentations to media organizations, and not only media but other publishers, is the infrastructure of indispensability. Every company now is a publisher anyway, so it doesn’t really matter what sector it is in. It’s about what kind of things you need to put in place apart from just newsrooms and editorial, what other things you need to put in place for you to become best-in-class for that particular topic that you cover, that particular niche product.

And it’s always about a triangle of product, data and content and of course data in the middle. So you kind of need to always be very, very precise because you will never be able to make positive changes to the other two sides of the triangle if one of them is not working. So data, products and content always need to work together.

Branislava Lovre:

Dmitry, thank you so much. I think that’s a great place to end. Thank you for sharing your knowledge with us and for joining us on AImpactful.

Dmitry Shishkin:

Thank you, Branislava. Thank you for inviting me.

Branislava Lovre:

And a special thank you to everyone watching and listening. See you in the next episode.