Despite being located in the capital of the most expensive country in Europe (and the second most expensive in the world after Bermuda), Google’s offices in Switzerland are – from the outside – surprisingly nondescript.
Slate grey bar the occasional flash of colour with the Google logo, it sits on the former site of a brewery but could be any office in any industrial park in any major city.
Inside, of course, is a very different matter. An explosion of geek-centric interior design, each of its seven floors has a theme – like “sky” (it’s the top floor), “sports” and “jungle”.
There are canteens and cafés on most floors, offering staff free food and drinks during the day. There is a massage room. There is a meditation room. There is a gym. There is a “power-nap” room whose only source of light is from backlit fish tanks. There’s a room set up for bands to play, with giant posters of Prince, Madonna, The Beatles and Bob Dylan on the walls.
There are reclaimed cable cars fitted out as meeting rooms (of course there are) and private pods to relax in or make calls from. There’s a full-time hairdresser, so large are the staff numbers now. Yes, there’s also a slide leading into the ground floor canteen.
On the “jungle” floor, there’s a room packed tightly with giant potted plants, like a mini botanic garden, and casually sitting by a meeting pod is a full-size plastic crocodile with a plastic doll’s arm lodged in its jaws.
(Keen fans of heavy handed visual metaphors for relationships between the tech industry and the music industry will have a field day with that one.)
Zurich was Google’s first engineering site in Europe, opening in 2004 with a staff of two. It’s now up to 1,800 employees and YouTube takes up the top three floors of the building, housing Google’s biggest engineering team outside of the US.
This, our guides tell us, is where “the Content ID magic” happens. (There will be several references to “magic” during our time here, incidentally).
HOW IT WORKS: THE COPYRIGHT TOOL
Music Ally is in Zurich to speak to the team behind Content ID and to see how it works as well as how it is evolving. The engineering team works from around 11am to 7pm (to be able to do conference calls with head office in California) and alongside Content ID, they are working on improving upload speeds, refining the video platform’s analytics and extending the tools offered out to creators.
Its algorithms are, as one might imagine, right at the cutting edge and the teams are increasingly leaning on machine learning to push it further still, so that the technology adapts to how people are actually using it. This is powering both recommendations and also how Content ID itself works.
YouTube is, of course, the whipping boy for parts of the record business – led by organisations such as the RIAA, IFPI and BPI who accuse it of accelerating a widening “value gap” which is also holding back the growth of full subscription services. YouTube, they say in their lobbying to both Washington and Brussels, is slowly suffocating their businesses and is gaming safe harbour exemptions to let itself off the hook for much of the damage it is seen to be causing.
The scale of YouTube is both its blessing and its curse for the music business. On the one hand, it is the most powerful marketing platform they have (which they can use for free); on the other hand, it is delivering the lowest ARPU of any licensed streaming service out there.
Their attacks on YouTube come at a time when the majors are all renegotiating their contracts so there is no smoke without fire. That said, even if the labels are taking negotiations into the public arena, that should not negate their criticisms of YouTube entirely.
MYSTERIOUS SPLITS
Unlike Spotify or Apple Music, for example, there is no minimum per-stream rate on YouTube. It is based on an ad-share model and that is subject to market fluctuations. Ad rates will be high during peak times (like Christmas) and lower at other times (January); as a result, it is very hard to make concrete earnings forecasts.
YouTube will not give specifics on the splits, but says the content owners always get the larger share. One company Music Ally spoke to recently said it was on a 55/45 split with YouTube and hinted that the majors could be expecting a 60/40 split as an absolute minimum.
YouTube says that record label income is split roughly 50/50 between the official content they have uploaded themselves and the claims (via Content ID) they make on UGC content (ranging from anything from a straight rip of a song to a track being used in the background of a cat video).
The problem, as the labels see it, is that there is far more UGC content than official content and it pays a much lower rate.
The ad inventory sold on UGC videos is lower than on official videos as advertisers place a premium on being able to drop in ads against what YouTube terms “trusted sources”. So if 50% of label YouTube income is from UGC, it is inherently built in that the ad rate will be lower and contribute more to the value gap. This is a key and unspoken part of the wider debate.
The IFPI says Content ID is a busted flush; YouTube begs to differ. The IFPI recently quoted a figure that suggested anything between 20% and 40% of music content online is falling through the cracks of YouTube’s Content ID system. YouTube says this is an “erroneous” claim. “We looked into those figures and found no basis for them,” it says, adding that Content ID is “97.7% precise”.
This is, in a large part, the trigger for YouTube bringing journalists to Zurich. They want to show off the Content ID technology and explain why it is far more robust than its critics (the major labels and their trade bodies) are claiming.
THE BIG NUMBERS
YouTube has 1bn regular viewers and view time has been growing 50% each year for the past three years. Mobile views doubled in the last year.
As consumption grows, so too does creation, fueling the beast even more. There are 400 hours of content being uploaded every minute to YouTube and the company says it is doubling its server space each year to be able to shoulder this.
Content ID swings into action as soon as a video is uploaded and will identify if copyright material is used and respond to the copyright owner’s terms (block the video, track the video, make a claim and monetise the ads around the video, set territorial restrictions and so on) before it goes live.
With so much content being uploaded, Content ID is expected to do far, far more in 2016 than it did in 2007 when it was first introduced.
In the earliest days of YouTube, copyright owners would lean on the DMCA to take down unauthorised user uploads. In 2007, YouTube introduced the first version of Content ID that was designed to automate the process as much as possible and give content owners much greater flexibility.
Today the service claims that only 2% of content management happens via the DMCA, with the other 98% happening through Content ID.
In the past nine years, YouTube has invested $60m in Content ID, starting with audio matching in 2007 and following this with video matching in 2008 and melody matching (to snag cover versions or live versions of songs, for example) in 2010.
Since its launch, it has paid out $2bn to all content owners. Some in the music industry regard this as “found” money, such as Scott Ackerman of TuneCore. Others see it as a massive devaluation of their content and akin to death by a thousand cuts.
COMMUNITY POLICING
Outside of the automatic scanning, YouTube says it is “by and large community policed”. As soon as any video gets flagged by a user, it is reviewed by a human rather than an algorithm. Content needs only one flag for it to be put through such a review.
Fabio Magagna, product manager at Content ID, went through how, exactly, it all works. This is based on proprietary technology developed in-house at YouTube and the company says it has no plans to white label it or license it out to other video services.
It now has 50m active reference files in its Content ID database and this is built up by content owners sending YouTube a reference file and its related metadata which YouTube then fingerprints so that the system will automatically make a match if a user uploads something with their content.
For video matching, the technology is much more advanced out of necessity. All official videos from content owners are sliced into frames, with each frame individually fingerprinted.
All fingerprints can be identified using heatmaps which can compare exactly how much of the official video content is used in a UGC video. The closer the angle is to 45-degrees, the straighter the line and the brighter it is, the more obvious it is how much of the original video has been used.
The UGC community has been working to try and fox the Content ID system, but it has adapted and can identify videos even if they have been mirrored/flipped, turned from colour into black and white, had their aspect ratio changed, been colour shifted, had light effect adds or any combination of these. It is also able to identify videos shot on a camera (from a TV or cinema screen, for example) and make matches.
YouTube claims it is, with advancing precision and evolving machine learning, “almost perfect” here. It does, however, admit it’s an “ongoing game” with uploaders trying to trick them. If an uploader gets three strikes, their account is suspended, although strikes disappear after six months. If an uploader is “in bad standing” they get limited access (e.g. no access to live streaming etc.)
The one thing it cannot identify is if someone uses ripper software to download a video or convert it into an MP3. While this is breach of the T&Cs of YouTube, its technology is powerless to spot this or prevent it from happening.
CONTENT ID FOR COPYRIGHT OWNERS
For content owners, Content ID gives them three options when a UGC video draws on their copyrights. YouTube claims that monetising is comfortably the biggest decision, with 95.5% of all UGC videos with copyrighted material being claimed by the copyright owners.
Just 4% of videos are blocked (for music this is often around pre-release leaks or other acts, such as Adele or Taylor Swift, who only want official content on the site as opposed to album rips). Finally, a tiny 0.5% of videos are tracked, where the content owner is not concerned about monetising the video but does want access to the usage data.
The company is doing rights management at huge scale (remember that stat about 400 hours of video being uploaded every minute). “It’s not perfect,” its engineers admit, “but it’s pretty robust.”
To explain how it is actually being used by content owners, Britta Sölter, director of strategic partnerships for Athletia Sports, and Gideon Mountford, director of video international at Believe Digital, talked about how they deploy Content ID every day.
Athletia works as an MCN for over 200 partners and sports organisations, including the Bundesliga in Germany. In any weekend, she claims there are 2,500 to 7,000 unlicensed videos going up online of football matches from the Bundesliga that they rely on Content ID to deal with.
The Bundesliga doesn’t add videos at the weekend as they want to give broadcasters exclusivity for a period on all matches, so will ration out clips through the week to drive online views.
Content ID is able to process live streaming so can work to identify unlicensed streaming of matches and block them if the copyright owner so requests. Tracking unlicensed content relies on either Content ID or keyword searches (across multiple languages). “If Content ID can’t find it and keyword searches can’t find it,” argues Sölter, “you’ll never find it.”
BELIEVE DIGITAL ON CONTENT ID
Mountford says that Believe set up a dedicated video department in 2009 as it was proving such a big part of its distributed labels’ business.
For music, the biggest video type after official videos is fan-generated lyric videos which most owners simply monetise and add campaign features that will direct users from a UGC video to the official channels of the artists in question.
He feels, however, that YouTube should to let them edit and cusotmise the (templated) messages here to allow them the same functionality they have with annotations. He says the recommendations from Believe to its clients is to simply claim and monetise all UGC videos and only apply the blocking option when it is related to a pre-released track.
“On a key release, we might advise the client to potentially block Content ID to focus all views on one asset that they are then able to take around to the media to show them that this video is doing well [so the UGC uploads don’t cannibalise views from the official video],” he says.
“After the promotion, we would probably advise the client to allow Content ID [so permitting UGC videos and monetising them]. By allowing Content ID to monetise on a catalogue track, it essentially gives lie to the long tail. It gives the song legs where traditionally that promotion period has ended. By having a Content ID policy that is monetising at the end of a campaign, we are able to drive additional revenue and awareness of the product that wasn’t previously there.”
Mountford was effusive about the algorithmic and machine learning leaps forward as Content ID has developed in recent years. In the past, remixes and mashups could slip through the net, but they are increasingly being caught. Even if tracks have been sped up, they can still be identified.
THE BLAME GAME: CONTENT ID AND MONETISATION
If UGC is a moving target, then Content ID is furiously keeping up and turning to machine learning to make itself smarter still.
A real trend in digital music in the past year has been blame shifting. Spotify will point to YouTube and say too many people find the service too good and this is stopping them from paying for subscription services.
Apple Music, meanwhile, points to the huge numbers of free users on YouTube (1bn) and Spotify (70m) and argues it is a better actor in this space as all its users are paying a monthly subscription (conveniently sidestepping the controversy over its three-month free trial when it launched last year).
So it follows that YouTube has some blame shifting of its own. Positioning itself as a social network as much as a streaming service, it claims that Facebook has limited search and enforcement options with regards to infringing videos on its platform while Twitter/Instagram/Vine have no automated detection or search as well as limited enforcement.
Of course, the demand for videos on all these platforms is a drop in the ocean compared to YouTube’s userbase, but it is insistent that it is not painted as the bad actor here.
Compared to them, however, YouTube’s Content ID system is light years ahead. This, somewhat inevitably, will not end the mounting criticisms facing it from certain (and significant) quarters of the music industry.
If YouTube is hoping to win a PR war here (or at least limit the damage), Content ID is just one part of it. The other side to address is monetisation.
Content ID lets content owners claim revenues generated around their own content. The content owners will perhaps now be less concerned about the “How?” of Content ID and would rather YouTube addressed the “How Much?” side of things.
Tools: platforms to help you reach new audiences
Tools: Kaiber
In the year or so since its launch, AI startup Kaiber has been making waves,…