Hi there.

Generative AI has become more sophisticated and therefore more accepted as the future of technology and something we need to send time on. Gartner has predicted that by 2026, traditional search engine volume will drop 25%, with search marketing losing market share to AI chatbots and other virtual agents.

Aside from the change in top-of-funnel traffic, we need to take chat seriously. It’s also the most obvious product use case within news. Today I want to summarise the research I have been doing by giving a practical guide to some of the decisions you will need to make if considering building a chatbot. 

Is this helpful? Drop me a note if this helps in your product development and/or you have other angles you’d like to see covered. I’m at Jodie.hopperton@INMA.org.

Thanks, Jodie

Everything we know so far about building chatbots for news

At a recent Webinar, I asked people who was working on chatbots. It’s slightly biased because the audience I asked had already subscribed to a Webinar about chatbots but ….

So let’s start by looking at a couple of chatbots to draw some conclusions — or at least questions you should be asking yourself before you build — on the product itself.  

Let’s take Bild as an example and break down their chat product. It’s a standalone chat, not embedded into other pages. At the very top in yellow, you see an alert: “You are now chatting with ‘Hey_,’ your AI helper. We are still in the testing phase. The answers are not from Bild. Please keep in mind that an AI can sometimes be wrong.”

They are very clear that this is not a finished product. It’s interesting to me that they are seemingly using the open Web for their responses. I’ll get back to this point later.

Below this alert near the top is a section with images and headers. These are popular questions that have already been asked. To note here: This is not only a chat — it’s content in its own right as a user can be voyeuristic, seeing what others have asked. This also brings the cost down as queries aren’t re-run. 

Next you’ll see that they give prompt ideas, which ChatGPT and other offers also do. Some kind of prompt seems to be essential to nudge people along the journey. A blank slate can be scary for people; prompts take them from zero to one. 

Bild’s chat is open for anyone to use.

Which users is this for?

• It could be gated for subscribers/members for engagement and retention. 

• It could be for logged-in users only as an incentive to get first-party data. 

• One option, as we were told at Skift during the recent INMA Media Subscriptoins Summit, is that it could be in front of the paywall as an incentive to sign up. Users get a short answer for free, but if they want to dig deeper, they need to sign up. 

• Or, of course, it can be free and open to more users. 

During our recent Webinar on chatbots, Lucky Gunasekara, CEO of Miso.AI, wrote in the chat: “One tactic we saw in the wild is using chatbot usage metering to drive anon visitors to register and then subscribe. We got early data from one site doing this and the conversion funnel was about 35% from Day 1.”

Where does it fit with the user journey?

• Most experiments seem to be standalone. We’ve seen it with Bild, Forbes, and Planet Money (the latter of which you can learn about in our recent members-only Webinar). This acts more like search.  

• Or on the article page (reportedly both Blick and Ippen are focusing on this). This could be helpful to users as they have a base (story) to go from. They are then likely to ask related questions such as going deeper on a specific point or asking for explanations.

Now, as we dig into the content itself, let’s look at how Forbes and Bild compare in how they respond:

Note how Forbes gives a summary and then goes to related articles. This is the same as we saw with Planet Money during their demo in the Webinar. Bild takes a different tack, answering everything in the chat and suggesting responses (none of which are further prompts). 

As a news organisation, I would assume that you want to point people towards your original journalism. We know this is true of Forbes and Planet Money, plus how Ringier is looking at this. 

How do we present content?

1. Interrogate your own journalism first. Show the sources and encourage people to go deeper to look at the original content (note this can be words, audio, or other formats).

2. If the chatbot doesn’t know the answer, programme it to say so. Planet Money was very conscious that they wanted the chatbot to admit when it didn’t know something rather than make something up (aka hallucinate). They created an alert that was shown if they didn’t have anything and only then offered to search the open Web. 

3. Where you don’t have content available, you may want to search the open Web using an LLM. There is an inherent risk that you can’t control, but on the flipside it can be clearly marked and doesn’t run the risk of frustrating a user. One AI company I have spoken to is looking to build a central repository of shared content from reliable news sources, which, if it contains a vast amount of data, could be a great compromise.

4. Consider the timeliness of content. Clearly you want to surface the most recent results first. You also need to be clear on whether it is information to be relied upon, e.g. a stock price or a breaking news story, and, as above, consider if you want to link to an external resources. 

5. Consider including possible follow-up prompts.

Whatever you decide to do, label it. With Bild, you see clear labeling that this is experimental. Others, Forbes, and Plant Money seem to only pull from their own journalism.

It’s a two-way street

The queries you get can also inform your newsroom or other departments. Is there an angle to the story that is missing? Is something unclear? Or is there a subject that people want to hear about that has not yet been covered? Chatbots could be an excellent listening device as well as offer users two-way conversations. 

An internal chatbot

The other example that we haven’t talked about here is that a chat doesn’t need to be for users. It can be for your own journalists who want to dig into the archives to see what has already been written. This could be step one of testing before anything is released externally.

How do you build a chat product? 

We’re not yet in a place where there are many tried-and-tested solutions to build chat. Here are the main options:

1. You can build your own, basing it off larger LLM such as ChatGPT. This is what Planet Money did.

2. You can use a third-party out-of-box solution. We’ve seen very few, but you can check out On Platform (for larger orgs), Miso AI, and Turing (a Brazilian startup used by Rede Gazeta).

3. You can use a hybrid with components from big providers such as AWS, Google, Chat GPT, etc.

What is the cost and ROI?

As my colleague Sonali wrote recently, the ROI of chat isn’t yet clear.

• Cost: It’s expensive right now. As Bernd Volf, managing director for Ringier Media Tech, pointed out at the Webinar: “If you’re successful, it could bankrupt you.” Steve Henn at Stanford, who built the PlatMoney chatbot, said by using CHatGPT 3.5, he got costs down to US$6 per 1,000 queries. Either way, this is something everyone is struggling with. For example, a Google GenAI search costs 25X that of a regular search. Just know that costs are likely to come down so even if you don’t want to launch now, you can prepare for when you do. 

• Revenue: It’s unclear but let’s look at the options:

• In extreme cases you could charge, although doing this specifically for news seems unlikely to get a lot of traction at this point. 

• This could be seen as a cost to engage and retain subscribers.  

• As per the Skift example above, this could be seen as a lead generator for subscribers. 

• And, of course, there is advertising/sponsorship. I have not yet seen use cases, but I am fairly sure they will come once we’re out of the experimental phase and showing real traction.

• So when will companies get to break even/profitability? This is the question that’s on everyone’s mind. If you have an answer, please let me know.

Last but not least

It doesn’t end here. Bernd Volf told us this is the precursor to audio. And I am with him 100%. 

If we are able to converse with our computers using questions and responses, and as audio technologies improve, it is only natural that this moves to audio. The devices and habits are already there as audio is built into many things: phones, smartwatches, earphones (AirPods have changed the game), and hundreds of millions of smart speakers in homes, which are currently underused playing music and setting timers. 

As consumers get used to chat, we’ll see audio get a new life.

Date for the diary: INMA’s World Congress of News Media starts April 24 in London

The lineup is incredible. The networking opportunities with steller attendees are second to none. And I would love to see you in person. Come join us in my original hometown to talk about the news business, sustainable revenue models, Artificial Intelligence, organisation structure, leadership transformation, building high-impact teams, and much much more. 

More about the World Congress can be found here. And if you’re coming, hit me up to let me know as I’ll be organising some off-schedule drinks for product and tech people. 

About this newsletter 

Today’s newsletter is written by Jodie Hopperton, based in Los Angeles and lead for the INMA Product and Tech Initiative. Jodie will share research, case studies, and thought leadership on the topic of global news media product.

This newsletter is a public face of the Product and Tech Initiative by INMA, outlined here. E-mail Jodie at jodie.hopperton@inma.org with thoughts, suggestions, and questions. Sign up to our Slack channel.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .