Related articles:

A major, almost existential challenge for many broadcasters lies in making programming both readily accessible and clearly identifiable as a brand from the mass of content out there in the streaming world. Technologies are emerging to help address the challenges, especially under the banner of AI, but only if they are embraced properly with an emphasis on higher contextual awareness.

A key point is that while broadcasters are at risk of being drowned out by the torrent of content from multiple sources across social media and streaming generally, and are certainly losing eyeball share, there is a counter trend towards authenticity, accuracy and trust. Broadcasters can tap into that if they can establish themselves as trusted purveyors of content, particularly for news and documentaries. There is the opportunity to position themselves as sources of accurate news that will be picked up not just by humans, but also AI bots as they tune in on contextual awareness.

It then becomes clear that technology intersects with content, regulations over prominence, and user behavior. This much has become evident from early forays into findability by broadcasters, which have unearthed issues to be resolved.

Getting Lost In The Noise

German broadcaster ZDF was one of the first to implement EU findability rules introduced as part of the latter’s Audiovisual Media Services Directive. These findability rules apply to two components – media platforms and user interfaces – but not what might be called media intermediaries such as app stores, search engines and social media.

An important point is that media platforms include not just broadcasters and pay TV operators, but also OTT service providers. This has the obvious implication that broadcasters are competing with other service providers and will not necessarily enjoy a privileged position when it comes to others’ user interfaces. Their success therefore depends on the primacy of their content, not just their position within a UI.

In fact, one lesson ZDF has learned is that having a prominent spot on a UI alone does not guarantee much public access. The broadcaster thought it had solved the issue of making hundreds of services all findable on third party platforms with limited space on the screen. It had addressed this by deploying a dedicated tile at the UI selection level, so that users supposedly could not miss it. Then by opening the tile, users could access the apps of all services deemed of public value, including ZDF’s.

In practice though most people just ignored the tile in the first place, even when they came in via the UI. A further issue was that many users still came in with remotes, many of which still house dedicated buttons for the big streamers, such as Amazon, Disney+ and Netflix.

ZDF deduced three key lessons from this experience. The first was to limit the number of specified public value services to ensure meaningful visibility, on the grounds that it is impossible to stand out in a large crowd. The second was to redefine privileged positions on UIs and remote controls to avoid relegating public value content to ‘dead zones’.

Presumably this might mean giving broadcasters such as ZDF their own dedicated buttons. Indeed, the third lesson was to adapt the rules to the reality of content disaggregation, so that individual programs are not just prominently displayed but also linked back to their original platform. This addresses the issue of broadcaster brand recognition, which is essential for maintaining privileged status as a Public Service Broadcaster (PSB).

Taking Advantage Of C2PA

While ZDF was in the middle of such teething troubles, France Télévisions appeared to be enjoying greater success deploying the Coalition for Content Provenance and Authenticity (C2PA) standard. This coalition was founded in February 2021 by Adobe, Arm, BBC, Intel, Microsoft, and digital content verification specialist Truepic, having been spawned from a merger between Adobe's Content Authenticity Initiative (CAI) and the joint Microsoft/BBC Project Origin. The aim was to establish a royalty-free and open standard for digital content provenance, aiming both to combat disinformation and provide verifiable information about content origin, creation, and edits.

France Télévisions’ objective was to meet the challenges of modern news operations, including content discoverability, through automation of the C2PA workflow. Each piece of content was to be cryptographically signed to confirm origin from France Télévisions. It has been a success in part because it had a more clearly defined objective than ZDF had with its broader initiative, aiming for proof of provenance and prominence, but not much more.

How Machine Learning Is Optimizing Search

Broadcasters also face the issue of optimizing search within their own content libraries and inside their own portals. The BBC has been investigating how machine learning can be employed to search its news content more effectively with greater user satisfaction. News content search poses different challenges from entertainment archives which are more historical and cover a longer period. News is always rolling and has a shorter shelf life, so search algorithms have to aim at much more of a moving target.

The BBC has been experimenting with AI techniques to improve the relevance of news search, calibrating against human feedback as well as automated testing techniques. It is currently incorporating data signals for search and discoverability within both its iPlayer TV portal and Sounds audio streaming and download site.

The machine learning operates by detecting which search signals, whether based on similarity of search, what is currently trending, or other metrics, are most predictive of what users will click on next. The BBC is aware that a given user does not behave in exactly the same way across all platforms, making different choices on mobile apps than when sat in front of a TV, or even viewing on a laptop, so the predictions are filtered on this basis.

Results are also improved when the system can probe deeper into users’ intentions or desires, beyond the basic keywords. This enters the realm of semantic search with an application of machine learning to understand what kind of content a user might be interested in through an analysis of that user’s broader natural language queries.

This is a clear objective for video service providers generally, not just broadcasters, involving greater personalization. The BBC says its next step is to investigate how Generative AI might be employed to deepen this personalization and understanding of users’ intentions.

NEO And Retrieval Augmented Generation

As a member of the EBU (European Broadcasting Union), the BBC also benefits from a larger ongoing project at the EBU called NEO, which the organization describes as a generative AI tool focused on Retrieval Augmented Generation (RAG). The aim is to take generative search output from LLMs (Large Language Models) up a level by applying advanced natural language processing to eliminate dubious or inaccurate sources. NEO comes in as the tool for analyzing natural language questions from users, as well as providing more intuitive and accurate synthesized answers.

The challenge is eased by limiting searches purely to trusted content provided by EBU member broadcasters. This exploits the EBU’s unique status as a loose coalition of broadcasters with 123 members from 56 countries, along with associate members from 20 other countries, some outside Europe.

As a result, NEO will also help these broadcasters liberate their own news content and make it more discoverable. This will be done through the EBU’s internal aggregation platform called “News Pilot”, which receives about 3,000 articles every day from its public service media members, and now has around 4 million articles in total.

Successful application of AI to overcome the deficiencies of traditional keyword-based search requires careful implementation. Part of this is to confine searches purely to trusted news content, hopefully avoiding the risk of searches returning irrelevant trivia. But it is also important to constrain the application of AI within strict boundaries, according to Alexis Alleman, EBU Data Scientist in Natural Language Processing, and co-author of the EBU’s seminal paper on NEO, “From Proof-of-Concept to Serving Audiences at Scale: The EBU NEO RAG Architecture”.

NEO’s algorithms were forced to generate responses strictly from the content presented and avoid using knowledge gained during its training process. Its goal is to avoid hallucinations where algorithms appear to jump to conclusions that are ultimately wrong.

The EBU’s Four-Part Process

There are four key components in the NEO technical stack. First is the News Pilot repository already mentioned. Second is a tool called EuroVox, which translates incoming text and audio content into English and other languages simultaneously, given that the EBU serves a multilingual base. Third is the aforementioned RAG, restricting the AI algorithms to articles that have previously been retrieved.

But the most novel component is the fourth, Dynamic LLM Selection, which allows the system to switch between LLMs according to the task, aiming to optimize both retrieval speed and precision. That is significant given ongoing debates over ethical AI, especially in regulated industries such as broadcasting, and provides the scope for differentiating between LLMs on the basis of ethical considerations.

The approach also allows flexibility on performance grounds. Not all tasks require the very largest LLMs capable of supporting 100 billion parameters; some are more efficiently executed by smaller models that may have been optimized for particular tasks. By incorporating dynamic LLM selection from the outset, the EBU hopes NEO will be able to evolve with the field as part of a flexible platform.

The EBU also avoids the term AEO (Answer Engine Optimization), which has emerged to add confusion to the field by suggesting it represents a radical advance over GEO (Generative Engine Optimization), when it is really just the same thing. Perhaps the term serves to underline the focus on responding to queries, in which case it might more properly be considered a subset of GEO.

What is beyond doubt is that GEO is a major step forward from SEO, the main point being to enhance rankings in searches conducted with more advanced semantical techniques to stand out in AI-driven results. Yet broadcasters and providers of news in general, whether print, audio or video, are finding that AI seems to reward good old fashioned journalistic practices.

This applies particularly to news, where journalists are taught the inverted pyramid method where key points are at the top, followed by supporting details of ever lower importance, but still valuable for the full picture. GEO based search also favors articles with summaries that are clearly written, which reduces the risk of misunderstanding by the algorithms while also increasing the chances of being prominently displayed.

AI search increasingly rewards clarity because that helps it establish its own ranking. Setting out clearly the who, why, what and where of a story is critical, and in that sense, the algorithms are not so different from people.

Such considerations extend beyond stories to web pages and whole websites themselves, since algorithms also engage with these in responding to queries. Broadcasters will therefore benefit from revising and updating their websites for the AI search era.

There are now tools such as Google's Rich Results Test which can assess web pages for clarity to appeal to AI algorithms. But there will always be room for improvement as even the websites of the big tech players themselves have been found wanting.

Moving Targets

Broadcasters have some grounds for optimism given their pedigree and trusted status in news and documentary production. But they will need to grapple with the methods and techniques of AI search optimization and associated metadata generation.

But at least there are now various collaborative initiatives designed to expedite AI based search across multiple platforms and sources, even if it remains to be seen how they will come together. There are also some key unifying threads emerging, such as the Story Object Model (SOM), an open standard designed to share contextual information across the production pipeline, in broadcast and digital environments generally. The aim here is to create a unified understanding of story context that could bring together multiple tools under the banner of GEO, including the EBU’s NEO.

Nevertheless, the field of content findability under AI is still in flux – and broadcasters should study developments keenly over the next few years as it matures.