On most platforms today, the algorithm, not the editor, decides who your audience is.
Unit C: How do we research and monitor?
By the end of today
What you will walk out knowing
01
Explain the social graph vs interest graph shift
Describe the three phases of algorithmic curation, from chronological feeds through social-graph ranking to interest-graph personalisation, and explain why the shift matters for audiences and creators.
Reveal
02
Explain how recommendation systems work
Describe the signals algorithms use, the feedback loops they create, and why around 80 percent of Netflix viewing is algorithm-selected rather than actively searched.
Reveal
03
Articulate algorithmic visibility as gatekeeping
Explain why a creator or publisher with no algorithmic distribution can have excellent content that reaches almost no one, and why this reshapes what audience research must now measure.
Reveal
04
Evaluate the engagement critique
State the argument that algorithms optimised for engagement amplify emotionally triggering content, and assess what this means for journalism, creators, and the audiences who consume their output.
Reveal
A Bangladeshi creator posts an excellent video. Nobody watches it. Was the content the problem, or was the algorithm?
Today we open the black box: not to reverse-engineer it, but to understand what it does to audiences, and what it does to the people trying to reach them.
The algorithmic turn
When data stopped informing and started deciding
Netflix: around 80 percent of what subscribers watch is selected by the recommendation system, not through active search. A title's commercial fate depends on whether the algorithm recommends it.
YouTube: the recommendation engine determines what video plays next. The algorithm's selection of "up next" is what drives the majority of total viewing hours on the platform.
TikTok: the For You feed is the most aggressive algorithmic curation in contemporary media. It predicts what each user will engage with next based on extraordinarily detailed behavioural signals.
The result: the algorithm has become a gatekeeper whose preferences materially affect what content is consumed and therefore what content can succeed commercially.
Before the algorithmic era, an editor decided what to publish and a programmer decided what to commission. What changed was that algorithms began making the decisions themselves, at a scale and speed no human editorial process could match.
Click each phase
Click each phase
Three phases of algorithmic curation
Phase 1 · ~2004-2010
Chronological feeds
The original model. Content displayed in reverse-chronological order from accounts you followed. Simple, transparent, but a poor allocator of attention. Most content was never seen. Audiences were bounded by when they happened to open the app.
Click to expand
Phase 2 · ~2010-2018
Social graph ranking
Platforms moved to ranking content from followed accounts by predicted engagement. Your audience experience was bounded by who you chose to follow. Follower count was the primary currency of reach. Growing an audience meant earning follows.
Click to expand
Phase 3 · 2018-present
Interest graph curation
Pioneered by TikTok, now adopted across platforms. The algorithm serves content based on what users engage with, not who they follow. A creator with zero followers can reach millions. A creator with a million followers can reach almost no one if their content underperforms. Follower count is decoupled from reach.
Click to expand
Core distinction
Social graph vs interest graph
Social graph
"Who you follow"
Your network of explicit connections: follows, friends, subscriptions. The platform shows you content from your graph, ranked by predicted interest.
Facebook (original), Twitter, early Instagram
Follower count = reach potential
Growing an audience means earning follows over time
Your experience is bounded by your network
Interest graph
"What you engage with"
The platform infers your interests from behaviour: watch time, skip rate, replay, saves, shares. Shows you content matching those signals regardless of who made it.
TikTok For You, YouTube recommendations, Reels
Behaviour signals = reach potential
A new creator can go viral with the right content
An existing audience offers no protection if content underperforms
Inside the black box
What signals algorithms actually use
Completion rate: how far through a video or article you go. The single strongest positive signal on most short-form video platforms.
Re-engagement: replaying, saving, sharing, or commenting. Each signals to the algorithm that this content was worth more than passive consumption.
Early engagement velocity: how quickly the content accumulates engagement signals in its first hour. A slow start is hard to recover from.
Negative signals: skipping, swiping past, hiding, or reporting. TikTok treats a long pause before a skip as informative as a completion.
Contextual signals: time of day, device type, current location, what was watched immediately before. The same piece of content can perform differently depending on when it appears.
On YouTube Bangladesh: thumbnail click-through rate is tracked to the second. A 10-second pause before click-away tells the algorithm your thumbnail overpromised.
Case study · Netflix
Micro-genres and the taste machine
Human coders at scale: Netflix employs people who watch content and classify it across dozens of variables: plot structure, mood, character traits, era, visual style, pacing. The result is tens of thousands of micro-genres.
The House of Cards decision: reportedly informed by data showing audiences who liked political dramas also liked Kevin Spacey and David Fincher. Three separate interest clusters pointed at the same investment.
The innovation problem: a system trained on what audiences have liked will recommend more of the same. Genuinely novel content does not fit existing patterns and may be at a structural disadvantage. Napoleon Dynamite famously broke Netflix's recommendation engine because nothing in the data predicted it.
The Bangladeshi parallel: Hoichoi classifies its drama catalogue into micro-segments to train its recommendation layer. The algorithm builds audience taste profiles independent of what the audience explicitly requests.
Algorithm sandbox
Algorithm sandbox
Toggle the signals, watch the feed change
Click signals to turn them on or off. The mock feed below re-ranks based on what the algorithm would prioritise.
Watch time
HIGH
Share rate
LOW
Comment rate
MEDIUM
Early velocity
HIGH
Save / bookmark
LOW
Skip rate (low=good)
LOW
Mock feed rank · 3 signals active
Strategic consequence
Algorithmic visibility as platform dependency
The 2017 Facebook pivot: when Facebook deprioritised publisher content in favour of friends and family posts, referral traffic to news sites fell by an estimated fifty percent within weeks. Many publishers found their businesses fundamentally damaged.
Rented land: any digital audience built on third-party algorithmic distribution is operating on land that can be repossessed at any time, without notice, and without compensation.
Creator vulnerability: a creator who builds a million followers can lose algorithmic visibility overnight if the platform updates its priorities. The followers remain but the reach disappears.
Bangladesh acutely exposed: Bangladeshi publishers depend on Facebook for a higher share of their digital traffic than publishers in almost any other major market. The strategic priority is building direct audience channels that the algorithm cannot mediate.
Sorting game
Sorting game
Signal or consequence?
Sort each item into whether it is an input signal the algorithm reads, or a consequence that results from its decisions.
Items to sort
Critical perspective
The engagement optimisation problem
The most serious critique of algorithmic curation is that the systems may work too well at maximising engagement, with adverse consequences for users and society.
01
The emotional amplification problem
Algorithms optimised for engagement tend to amplify content that triggers strong emotional responses: outrage, fear, indignation, tribal solidarity. This content outperforms calm, balanced content. The feed becomes an engine of emotional arousal, not information.
Reveal
02
The filter bubble debate
If algorithms serve you content matching past behaviour, they may progressively narrow your information environment. The academic debate is contested: research shows some filter bubble effects, but audiences also actively seek diverse content.
Reveal
03
The innovation discount
Recommendation systems trained on past patterns tend to surface more of what already worked. Genuinely novel content has no historical signal pattern and may receive less distribution. This creates a structural bias against truly new forms.
Reveal
04
The journalism crisis
For news organisations, algorithmic distribution creates a specific tension: the content most likely to receive strong engagement signals is not always the content most important for a democratic public sphere.
Reveal
Branching scenario · Creator strategy
Branching scenario
You have 500 subscribers and a great idea
You are a Bangladeshi street food creator on YouTube. You have 500 subscribers and have been making videos for six months. Your last three videos each got around 800 views. You have an idea for a new video about a legendary halim seller in Puran Dhaka. The algorithm is your gatekeeper.
Quality first. The video is your best work. 620 views in week one. The algorithm's early velocity signal is weak. It distributes the video narrowly. By month two it has 2,400 views, growing slowly from search. You are building an archive but not breaking through. The algorithm did not notice.
Signal optimisation. The shorter runtime improves completion rate. The 9 PM Friday post catches audience at peak time. Early velocity is the strongest of your career. The algorithm distributes to non-subscribers. 14,000 views in week one. You gained 340 subscribers. You did not make your best video, but you found your audience.
Cross-platform seeding. The Facebook teaser generates 1,200 reactions and 340 shares. When the YouTube video drops, early engagement is high. Algorithm reads strong early velocity and distributes it to interest-matched non-subscribers. 22,000 views. You gain 800 subscribers. Both platforms worked for you, but you now depend on both.
Knowledge check
Knowledge check
Under the interest-graph model, what is the primary determinant of a creator's reach on a platform like TikTok?
C. The interest graph decouples reach from follower count. A creator with zero followers can reach millions if their content generates strong early engagement signals. Conversely, a creator with a large following can see minimal reach if their latest content underperforms.
Netflix's recommendation system selects around 80 percent of what subscribers watch, meaning most viewing is driven by the algorithm rather than active search.True. Netflix's own data confirms approximately 80 percent of viewing is recommendation-driven. This is why the recommendation system's decisions materially determine a title's commercial fate.
For Bangladeshi publishers, the 2017 Facebook News Feed change illustrated the risk of building an audience strategy on rented algorithmic land that can be repossessed without notice.True. When Facebook deprioritised publisher content in 2017, referral traffic to news sites fell sharply. Bangladeshi publishers, who depend on Facebook for an unusually high share of digital traffic, face compounded exposure to this risk.
Your turn · 15–20 min
Activity: Reverse-Engineer the Feed
Step 1
Pick one app
Choose one platform: YouTube, Facebook, TikTok, or Spotify. Open it fresh. Do not search for anything.
Step 2
Log 10 consecutive recommendations
Write down the first 10 items the platform surfaces to you. Note format, topic, creator, approximate view count, and whether you have seen this creator before.
Step 3
Infer the three signals
Looking at the 10 items, what three signals do you think the algorithm weighted most heavily? What does this feed say the algorithm thinks you are?
Today in five lines
Recap
01Algorithmic curation has moved through three phases: chronological, social graph, interest graph. The interest graph is now dominant.
02The interest graph decouples reach from follower count. Behaviour signals, not subscriber numbers, determine distribution.
03Around 80 percent of Netflix viewing is algorithm-selected. The recommendation engine is the new gatekeeper.
04Platform-dependent audiences are rented land. One algorithm change can halve traffic overnight, as Bangladeshi publishers have experienced.
05Algorithms optimised for engagement may amplify outrage and fear over calm, important content, creating a structural tension for journalism and public-interest media.
Next class: Why Monitor at All? We move from how organisations know their audiences to why that knowledge has economic and strategic value.
Digital Audience · Lecture 12
The algorithm does not find your audience. It builds one.
This week’s reading
Reading list
PrimaryHill & Lunt (eds.) — Ch. 9, “Platformisation and Personalisation: The Making of ‘Contingent’ Online Audiences”
SupplementaryHill & Lunt (eds.) — Ch. 11, “Relationship Status of Journalists with Their Audiences on Social Media”