Echobox is designed to be better than all but the best social media teams, a feat that is achieved by the perfect combination of content, image and timing. 

Content is, of course, utterly key, which means the first task for the software is to determine which articles should be disseminated for maximum success.

“A big publisher will publish up to 1,000 articles each day, but you only get to share maybe 30 or 40 on Facebook. So the question is: out of those 1,000, what are the best 30 or 40?” asks Amann. 

“Prior to Echobox, publishers would have teams of editors going through all that content manually, looking at analytics screens as well and looking at feeds of their competitors on Facebook and Twitter; looking at what's trending, doing all this manual analytics work to decide what to actually share.

“The way we solve this is we predict the virality of content in order for the human not to have to do it anymore. What we use for that is a deep learning deep neural net and we train with the model of historical data, so we know how well certain articles worked in the past, and then we predict future articles using those models.”

For each article that is selected, the right image and text also needs to be found to ensure maximum success. 

“We read through all the sentences in an article, and we can measure the virality of every single sentence in an article and add up every single word as well, and then we pick the best sentence to be used as a share message,” says Amann. 

“Also we pick the best image as well, so out of all the images that are available you can A/B test different images, and we basically pick the image with the highest virality.”

However, the right content alone won’t guarantee success. Timing is also hugely important, as different types of articles will do better if disseminated at different times. 

“For instance, some politics pieces might work well in the mornings, whereas other types of pieces might work well in the afternoons, and we can calculate this by the minutes of the day,” he explains. 

“That's a granularity at which we can define what the best time is, and in order to do this we use a genetic algorithm, so we analyse all the content that we chose to share, and we test it out across many different times of day with the genetic algorithm. The optimal output is essentially: what combination of all these articles and times would generate the most traffic for our clients?”

However, in order to do this effectively, Echobox also needs to understand the difference between different types of content. 

“When we say we're building an AI that understands content like a human, we humans – when we read an article – we notice certain things that are very hard for a machine to notice,” explains Amann. “For instance, we would know whether an article is written, [on a subject] like a terrorist attack, where it needs to go out quickly. 

“We also will know intuitively if an article is evergreen content, where it is an article that can go out at any time, and especially editors know that very well when they read their content.”

As a result, Echobox is designed to have this same intuitive understanding, where it knows if an article needs to be shared immediately, or can be held back for later.

“We predict if something is breaking news or whether something is evergreen content, or if something is todayish content that is between the two, and for this we use different methods, and we use lots of national language processing as well,” says Amann. “So we analyse the article, we analyse the keywords, we compare those with labels that we've collected from our clients of historical breaking news and historical evergreen content, and this gives us a good prediction of what type of content something is within just a few seconds.

“So we have the ability to classify content a lot faster than a human ever could.”

While Echobox has standard systems for its automation, a significant part of its success is down to the fact that its service is heavily tailored to each publication, to appropriately reflect their editorial style. 

“We found early on that for this to really work well we have to build a model for every client separately. So each client has their own model, and it’s calibrated according to the audience that actually reads that content,” says Amann.

“So the audience of, let's say, The Guardian, has different preferences than an audience like The Sun or Deadline Hollywood, for instance.”

This tailoring is achieved by gaining in-depth access to a company’s existing social media accounts, seeing what success human social media teams achieved for the publication and using that data to do the same faster, and typically with even more success.

“Any algorithms that are built, any machine learning methods you use, are only as good as the data you actually have,” explains Amann. “Whenever we sign up a new client, we basically ask them to give us access to all their data, so they give us access to their entire analytics software, and we scrape a lot of the data, and this is extremely granular real-time data. 

“Other companies that try and optimise for social media presence, often they will look at purely public data, so they will look at how many likes and reshares and so forth there is, but that data is not granular enough to actually build a really good model like we have, and to actually achieve what we have achieved.”

As a result of this approach, Echobox beats all but the most brilliant human social media teams, and even in those cases the software can achieve the same level of success in less time.

“The one metric we use to define whether our technology can outperform humans is pageviews back to websites, so traffic, and the other one we look at is time saved,” he says. “Now some publishers will have really good social media teams in place already, and when they started using our technology they didn't see a traffic increase, because they are already doing a really good job, and this does happen. 

“However those publishers will benefit from a lot of time saved, so instead of doing a lot of analytics and spending hours and hours doing other things, they can spend more time thinking about what copy to use and thinking about more creative things than doing analytics. 

“A lot of publishers, however, they don't have good social media teams in place, or they are understaffed for social media, and those typically would be the ones that would reap the biggest benefits from using our technology, including traffic increases and also time saved as well.”

Echobox has seen considerable success in the field of publishing, but there is only so far that a company focusing purely on this market can grow. 

“We recognised from the beginning that the publishers space, even though it’s an amazing space to be in, if you want to build a great company just selling to publishers won't be enough,” explains Amann. “It's a small market and unfortunately publishers are struggling financially.”

As a result, Echobox is now looking to expand beyond publishing, with a variant of its software designed to meet the needs of brands and marketers – a field that dwarfs publishing in terms of scale and earning potential. 

“We're actually now in the process of launching a new Echobox specifically designed for brands and marketers, and it will be a bit different, so less about automation and more about insights and analysing different copies, analysing different images,” says Amann. “We've started doing some trials with some agencies who give us lots of different images and we basically score everything for them and tell them which is the best image. 

“We've also started doing this with musicians as well – with well-known artists who have different photos of concerts of themselves and their agency will post them on Facebook – and they want to know what is the best image to post on Facebook. 

“So that's a new product we're going to launch very soon, and it's going quite exciting.”