For over 10 years now, when Facebook first introduced the Newsfeed concept to their product, the company has been extremely active in improving the back-end that drives the central part of an intensely personalised experience. Lacking a better term - this back-end has been nicknamed the Facebook Algorithm - a highly complex set of processes that factor in thousands of previous interactions, preferences or social connections to predict the most engaging story that their users would click, like or share.
The Newsroom allows publishers to apply some of these practices that helped Facebook create the world's most engaging newsfeed product.
Newsroom's Content Delivery Algorithms are built from a collection of settings, permissions and auctioning parameters that, when executed, shape the experience of your users. The following processes take place right before one widget is filled with one or multiple cards, depending on its layout.
All of the 6 steps presented below are typically processed in under 0.02 seconds.
1. Creating Content Request
The Content Request is the equivalent of a real-world business RFP (Request for Proposal) - effectively stating the requirements for a story to be considered and then presented in a widget for one visitor, in a highly competitive auctioning process.
These requirements are relatively straightforward. Here's an example:
- Content Category: Stock Markets
- Total units to be filled: 8
- Created: Past 48h
- Headline: 250 char. max
- Description: 1000 char. max
- Native ads: Accepted
- Partner Content: Accepted
- Max ads / widget: 2
- Optimised for: CTR
- User Agent: Mozilla/5.0 (iPhone; CPU iPhone OS 9_1) AppleWebKit/601.1.46
- User's segments: !91288622618,338363704!99999
- Geo location: London, UK
Please note that most requirements above are extracted from the widget’s settings. The requests also include user’s device capabilities, the geo-data, and the segments the user falls under.
2. Adding Active Demand
Active Demand is the sum of all platform content that would “demand” visibility into the news feed. The platform will now start to look up for all the active cards and match them with the opportunity presented in the Content Request above.
Cards are automatically nested under strategies - or campaigns that represent these stories in the auctioning process - on a mission to drive traffic to their respective articles. So these campaigns would “speak” on behalf of the Cards in the selection process - passing along their own targeting criteria.
So if for example, one campaign would only be targeting London, UK users and the Content Request would issue a ticket out of Southampton, UK - the strategy would not return any content for consideration. However, it’s more common for native ad campaigns to carry additional targeting, while content would run untargeted.
Generally, every editor has 5 active types of strategies that represent cards in auction in different ways - from a very aggressive breaking news mode to an evergreen mode for older content. Read more about cards life stages here. Platform demand also includes external content as well as native ads that can be allowed in the auction - depending on the widget settings.
By the end of this selection process, the platform would have built the list of all stories that fit one particular widget.
3. Placement Auction
The auctioning process can start with thousands of potential stories and establish a number of winners to fit with the few stories that one widget can display.
The auctioning process will now rely on 2 main variables:
- click probability or recency
- bid amount
Click probability score is a metric that each story gets to describe their potential ROI when delivered under similar circumstances. Such circumstances include: time of day, device type and geo-location. Nonetheless, our systems can dynamically flag any attributes or pairings that users who previously clicked that story have in common.
For example, if a certain evergreen story (the topic is not relevant) has proved successful when delivered after 7 pm in Perth, Australia, our systems will aggressively try to replicate this pattern and give this story a high click probability between 7 and 9 pm. Similarly, if one story has proven successful when recommended on one specific page, this story will be consistently awarded a high click probability when entering auctions in that particular URL.
Such patterns are commonly identified by our platform, especially on content that triggers emotional response. Travel and shopping stories, for example, tend to have significantly higher click through rates between 7 and 9 pm. However, we don’t encourage publishers to consider hourly settings or frequency caps based on such insights - given our platform’s ability to match such insights with secondary parameters, like location or device types.
Depending on the settings of the widget requesting the content, recency can overcome click probability. If one widget requires latest stories first, cards with the newest timestamp wins. This is a good way for the newest cards to build up CTR data for later stages when they are competing in widgets that prioritise click probability.
Most content will bid the same amount of the platform’s virtual currency - allowing the click probability and recency decide the auction winner. For example, if two stories have a $0.2 bid but different click probabilities, the highest scoring one wins the auction and deliver in page.
However, though, publishers can define multiple bid levels and set higher bid amounts for content that would not win naturally in a given context. Higher bids are temporarily applied to Boosted cards. Read more about cards here.
Bid amount stands as a differentiator between native campaigns that are competing for the ad-slots available in a widget. Most native campaigns would practice variable bidding though, placing higher or lower bids, depending on their internal prioritisation and overall accomplishments.
In this stage, the platform looks up whether any of the auction winners have been previously displayed earlier on the page or are subject to a noise-reduction policy. Noise reduction prevents cards from being recommended repetitively or shown after having been clicked by the user. In this case, the platform instead picks the 2nd or 3rd winning cards of the auction and repeats the de-duplication process.
5. Content Delivery
The platform will deliver the assets associated with the winning cards into their respective placements and record the impression.