Doctor SpinThe PR BlogDigital TransformationWhat You Need to Know About TikTok's Algorithm
Digital PRDigital Transformation

What You Need to Know About TikTok’s Algorithm

How iterative testing is outperforming account authority.

Jerry Silfwer
by Jerry Silfwer

Cover photo: @jerrysilfwer

Table of contents

TikTok’s algorithm does­n’t work the way most people think.

TikTok, an app for mak­ing and shar­ing short videos with oth­ers, is one of the world’s most pop­u­lar social media plat­forms. But how does its algorithm work? 

This art­icle will exam­ine TikTok’s algorithm and explore why it’s so suc­cess­ful for vir­al videos.

Here we go:

The Success of TikTok

Since its cre­ation in 2017, TikTok has become one of the world’s most pop­u­lar social media plat­forms. The formerly Musical​.ly app has over 750 mil­lion users and is espe­cially pop­u­lar with teens and young adults. 1TikTok will have 750 mil­lion monthly users world­wide in 2022, accord­ing to Insider Intelligence’s fore­cast. It is now the third-largest social net­work of the Big Five world­wide (Facebook, Instagram, … Continue read­ing

The app, owned by Chinese com­pany ByteDance, is a video-shar­ing plat­form where users can share short videos of them­selves with oth­ers on the app. 

TikTok’s algorithm is con­stantly chan­ging and evolving to bet­ter match users with videos they will enjoy.

How Social Media Has Changed

Algorithms: The Silent Switch

Not that long ago, social media algorithms would deliv­er organ­ic reach accord­ing to a dis­tri­bu­tion that looked like this:

The Silent Switch - Doctor Spin - The PR Blog.001
Social media algorithms before the silent switch (click to enlarge).

Today, social media algorithms deliv­er organ­ic reach more like this:

The Silent Switch - Doctor Spin - The PR Blog.002
Social media algorithms after the silent switch (click to enlarge).

It’s the Silent Switch where social net­works have demoted the pub­lish­er­’s author­ity and repu­ta­tion and pro­moted single con­tent per­form­ance instead.

This algorithmic change has likely had pro­found and severe media logic amplifications:

Classic Media Logic Effects

Classic media logic is hypo­thes­ised to influ­ence the news media in the fol­low­ing ways: 2Nord, L., & Strömbäck, J. (2002, January). Tio dagar som skakade världen. En stud­ie av medi­ernas beskrivningar av ter­ror­at­tack­erna mot USA och kri­get i Afghanistan hösten 2001. ResearchGate; … Continue read­ing

  • Aggravation. As a res­ult of media logic, the news media will exag­ger­ate events, con­cepts, and ideas to make them seem more severe or dan­ger­ous than they are.
  • Simplification. As a res­ult of media logic, the news media will dumb down events, con­cepts, and ideas to make them seem more under­stand­able than they are.
  • Polarisation. Because of media logic, the news media por­trays events, con­cepts, and ideas as more conflicting/​provocative than they are.
  • Intensification. As a res­ult of media logic, the news media will sen­sa­tion­al­ise events, con­cepts, and ideas to make them more irres­ist­ible than they are.
  • Concreteness. Because of media logic, news media will report events, con­cepts, and ideas more straight­for­wardly than they are.
  • Personalisation. Due to media logic, the news media will over­em­phas­ise the role of named indi­vidu­als in con­junc­tion with events, con­cepts, and ideas.
  • Stereotypisation. Because of media logic, the news media frames events, con­cepts, and ideas as more aligned with con­ven­tion­al perceptions/​opinions than they are.

Social Media Logic Effects

  • Programmability. Social media logic enables and encour­ages users to cre­ate and manip­u­late con­tent, lead­ing to a tailored por­tray­al of events, con­cepts, and ideas that might not fully rep­res­ent reality.
  • Popularity. Driven by social media logic, con­tent that gains ini­tial pop­ular­ity can dis­pro­por­tion­ately influ­ence pub­lic per­cep­tion, regard­less of accur­acy or completeness.
  • Connectivity. Social medi­a’s inter­con­nec­ted nature, rein­forced by social media logic, facil­it­ates the rap­id spread of inform­a­tion, often without suf­fi­cient veri­fic­a­tion, lead­ing to a dis­tor­ted under­stand­ing of events and ideas.
  • Datafication. The social media logic of con­vert­ing inter­ac­tions into data points emphas­ises quan­ti­fi­able aspects of events, con­cepts, and ideas, poten­tially over­look­ing their qual­it­at­ive nuances.

Our job as PR pro­fes­sion­als is to help organ­isa­tions nav­ig­ate the media land­scape and com­mu­nic­ate more effi­ciently, espe­cially dur­ing times of change.

Learn more: The Silent Switch: How Algorithms Have Changed

Misconceptions About TikTok’s Algorithm

Many people have mis­con­cep­tions about how TikTok’s algorithm works. They believe the algorithm is based solely on tim­ing and engagement. 

However, this is not the case. ByteDance is basing the algorithm on an iter­at­ive pro­cess that tests dif­fer­ent audi­ence sub­sets to see which videos are more likely to be popular. 

Wait. What?

Sure, the TikTok algorithm is, to an extent, designed to sur­face the most rel­ev­ant con­tent for users. It con­siders vari­ous factors, such as how often a user has inter­ac­ted with a par­tic­u­lar cre­at­or or video, the time of day, and the user­’s location.

But there’s more to the TikTok algorithm. Much more.

How People Think About TikTok’s Algorithm

Here’s how most people think that the TikTok algorithm works:

TikTok’s algorithm tracks your activ­ity on the app. Based on this data, it uses advanced machine learn­ing to build a psy­cho­lo­gic­al pro­file on you, the user, to serve you con­tent spe­cific­ally to max­im­ise your engagement.

If the TikTok algorithm builds psy­cho­lo­gic­al pro­files on each user, here’s what that type of AI com­put­ing would entail:

Whenever your “per­son­al” algorithm is fired off, your smart­phone app must call TikTok’s cent­ral­ised machine-learn­ing net­work. Then, the cent­ral net­work must pro­cess all vari­ables, decide on an out­come, and deliv­er the res­ult to your smart­phone.

From a busi­ness per­spect­ive, the pro­cess comes with two sig­ni­fic­ant draw­backs:

1. AI serv­er calls are expens­ive. 3“Sufficient algorithm per­form­ance is one of the major cost factors to con­sider because accur­ate pre­dic­tions require sev­er­al rounds of tun­ing ses­sions, which raises the cost of imple­ment­ing AI … Continue read­ing
2. Individual human beha­viour is still an AI chal­lenge. 4“Human beings are most often an integ­rated part of com­plex sys­tems. In order to describe such a sys­tem with appro­pri­ate accur­acy it is neces­sary to mod­el the human com­pon­ents with the same accur­acy … Continue read­ing

Two great examples of cur­rent com­mer­cially viable AI cap­ab­il­it­ies are Siri and Alexa. Your spoken words are uploaded to the cloud, ana­lysed using arti­fi­cial intel­li­gence, and the out­come is then delivered back to your device. Siri and Alexa are excit­ing tech­no­lo­gies but have quite lim­ited capabilities.

What if there’s a way to design a social media algorithm that’s a) faster, b) cheap­er, and c) more accurate?

How the TikTok Algorithm Works

Here’s how a faster, cheap­er, and more accur­ate social media algorithm operates:

Social Media Algorithms - Silent Shift - Doctor Spin - The PR Blog
How algorithms iter­ate to max­im­ise engage­ment and con­tent quality.
  • The algorithm pre-builds a data­base of users based on their beha­vi­our­al data.
  • When a cre­at­or posts con­tent, that con­tent is shown to the smal­lest pos­sible (but still stat­ist­ic­ally sig­ni­fic­ant) sub­set of pre-defined user types.
  • If the con­tent is suc­cess­ful with the small test group, the con­tent gets pushed to a slightly lar­ger test group.
  • If the con­tent is still thriv­ing, it is passed on to a some­what lar­ger test group. And so on.

TikTok, the video-shar­ing app, has been shad­ow­ban­ning users who post con­tent incon­sist­ent with its guidelines. Their videos do not appear in the app’s search res­ults, and their fol­low­ers do not see their posts. TikTok has not clearly explained why this is hap­pen­ing, but the com­pany is likely try­ing to avoid leg­al trouble.

Consequences for Users and Creators

You might won­der what this means for the users and con­tent creators.

To bet­ter explain, we must estab­lish a baseline, i.e. have some­thing to com­pare the out­comes of today’s social media algorithms.

Here’s how social media algorithms used to work:

Historically, social net­works used algorithms to match past user beha­viour with past con­tent cre­at­or content.

In the early days of social media, social net­works did­n’t have the tech­nic­al cap­ab­il­ity to run live beha­viours against live con­tent. So, they had to rely on his­tor­ic­al data accu­mu­lated over time on users and con­tent creators.

In sum­mary, we have dis­cussed three gen­er­al types of algorithms:

Matching based on his­tor­ic­al data = This is how most social media algorithms oper­ated in the past.

Live serv­er-side pro­cessing based on indi­vidu­al inputs = This is how many people (and the news media) think social media algorithms oper­ate.

Iterative test­ing on pre-built audi­ences = This is how most of today’s social media algorithms work.

Hopefully, we will soon see more live serv­er-side pro­cessing based on indi­vidu­al inputs. However, noth­ing indic­ates that this is how social media algorithms oper­ate today.

So, what does this mean? For users and con­tent cre­at­ors, the primary out­come is still the same:

Consistency = With his­tor­ic­al match­ing and iter­at­ive test­ing, con­sist­ency is the most prom­in­ent approach to get­ting stable and some­what pre­dict­able res­ults. Live serv­er-side pro­cessing would allow users and con­tent cre­at­ors to exper­i­ment (and reach rel­ev­ant audi­ences dynam­ic­ally), but we’re not there yet.

But most oth­er algorithmic out­comes are dif­fer­ent today:

Authority = Having a large fol­low­ing mat­ters less today because iter­at­ive test­ing will push single con­tent based on their live scores, not the con­tent cre­at­or’s mer­its. It pro­motes attract­ive thumb­nails, click­bait head­lines, second-by-second user reten­tion, etc. — and less on cre­at­or clout.

Timing and decay = Iterative tests can pro­duce a slow burn, allow­ing con­tent to go vir­al more slowly. Historical match­ing requires con­tent cre­at­ors and users to be act­ive with­in the same time inter­vals for a piece of con­tent to pre­serve its momentum.

Competition = With iter­at­ive test­ing, there are few­er short­cuts. To a great­er extent than before, every con­tent com­petes with every oth­er part. And with the trend of show­ing users more con­tent from non-fol­lowed cre­at­ors, com­bined with increas­ingly more avail­able con­tent over­all, com­pet­i­tion will only get more challenging.

Iterative Testing Pushed to the Limit

A reas­on­able asser­tion at this point is to con­clude that social media algorithms favour eas­ily digest­ible con­tent that has been dumbed down to a bare minimum. 

Today’s social media algorithms seem to favour fant­ast­ic­al and atten­tion-grabbing con­tent — almost to the point of becom­ing par­od­ies of themselves.

Well, yes.
This is sort of true.
But not quite.

TikTok is the lead­ing example of push­ing the bound­ar­ies of iter­at­ive testing.

The TikTok algorithm favours con­tent requir­ing very little of the user­’s men­tal effort. And this is, obvi­ously, suc­cess­ful for pro­du­cing enter­tain­ing (and addict­ing) user experiences.

However, Google Search is on the oth­er side of this spec­trum, with a heav­ier lean towards his­tor­ic­al match­ing. Building author­ity through traffic, links, and ana­lyt­ics data for SEO suc­cess still matters.

But this is why social net­works like Facebook, YouTube, and Instagram struggle with their algorithms. They’re posi­tioned between TikTok and Google Search.

On the one hand, they want to emu­late TikTok’s massive algorithmic suc­cess by serving eas­ily digest­ible and fant­ast­ic­al video shorts. On the oth­er hand, they also need to pre­serve their relevance:

What good is a friend app without friends? (Facebook)
What good is a long-form video app without long-form videos? (YouTube)
What good is a photo app without pho­tos? (Instagram)

How To Improve Your TikTok Content

You’re prob­ably read­ing this blog post to under­stand how to nego­ti­ate the TikTok algorithm suc­cess­fully as a con­tent creator.

Great!

When you have a deep­er under­stand­ing of how TikTok’s algorithm works, it becomes easi­er to fig­ure out the right formula:

Stay con­sist­ent and only make minor iter­at­ive tweaks to your videos based on your res­ults. Perhaps your thumb­nails aren’t grabbing atten­tion? Make tweaks and post ten videos using bet­ter thumb­nails. Did it work? If not, try anoth­er thumb­nail approach for ten videos. Otherwise, move on to the next tweak.

Repeat.

Maybe you’re now won­der­ing where to start.
I would­n’t worry too much.

  • Thumbnails can con­tinu­ously be improved.
  • Headlines and cap­tions can con­tinu­ously be improved.
  • User reten­tion through­out your videos can con­stantly be improved.
  • Storytelling twists and turns can con­tinu­ously be improved.
  • Production qual­ity can con­tinu­ously be improved.
  • On-cam­era per­form­ances can con­tinu­ously be improved.

See my point? The con­scious act of iter­at­ively improv­ing your con­tent out­put is more para­mount than your choice of where exactly to start.

Signature - Jerry Silfwer - Doctor Spin

Thank you. Please sup­port my blog by shar­ing art­icles with oth­er com­mu­nic­a­tions- and mar­ket­ing pro­fes­sion­als. Please also con­sider my PR ser­vices or speak­ing engage­ments.

PR Resource: The 7 Graphs of Algorithms

Spin Academy | Online PR Courses

Types of Algorithm Graphs

Search engines, social net­works, and online ser­vices typ­ic­ally have a wealth of user data to optim­ise the user experience.

Here are examples of dif­fer­ent types of graphs that social media algorithms use to shape desired behaviours:

  • Social graph. The media com­pany can access your friend list and push their con­tent (or favoured con­tent) into your feed.
  • Interest graph. The media com­pany can access your interests (top­ics, per­sons of interest, dif­fi­culty level, format pref­er­ences, on-plat­form-spe­cif­ic beha­viours, etc.) from your usage history.
  • Predictive graph. The media com­pany can access all graphs from users not con­nec­ted to you but with whom you share a stat­ist­ic­al like­ness and show their pre­ferred con­tent to you.
  • Prescriptive graph. The media com­pany can push con­tent into your feed to manip­u­late your over­all emo­tion­al exper­i­ence when using the platform.
  • Trend graph. The media com­pany can push con­tent into your feed based on what seems to be trend­ing on the platform.
  • Contextual graph. The media com­pany can access con­tex­tu­al data like loc­a­tion, weath­er, cal­en­dar events, affil­i­ations, world events, and loc­al events.
  • Commercial graph. The media com­pany can access data on how you and oth­ers like you inter­act with com­mer­cial content.

The dif­fer­ent graphs are typ­ic­ally weighted dif­fer­ently. For instance, some media com­pan­ies allow a fair degree of social graph con­tent, while oth­ers offer almost none. Changes are con­stantly being enforced, and the silent switch might be the most not­able example of a media com­pany shift­ing away from the social graph. 5Silfwer, J. (2021, December 7). The Silent Switch — A Stealthy Death for the Social Graph. Doctor Spin | The PR Blog. https://​doc​tor​spin​.net/​s​i​l​e​n​t​-​s​w​i​t​ch/

The media com­pany can lever­age these graphs using two main approaches:

  • Matching. The media com­pany can use vari­ous graphs to gen­er­ate your social feed. Depending on the com­plex­ity of the ana­lys­is, this approach is slow and expens­ive with react­ive (unpre­dict­able) results.
  • Profiling. The media com­pany can use vari­ous graphs to place you in stat­ist­ic­al sub­groups, allow­ing con­tent to iter­ate to the right audi­ence. This approach is fast and cheap with pro­act­ive (pre­dict­able) results.

Today, pro­fil­ing seems to be the dom­in­ant approach amongst media companies.

Learn more: The 7 Graphs of Algorithms: You’re Not Unknown

💡 Subscribe and get a free ebook on how to get bet­ter PR.

Logo - Spin Academy - Online PR Courses

PR Resource: The Electronic Age

Enter: The Electronic Age

Human cul­ture is often described based on our access to pro­duc­tion tech­no­lo­gies (e.g., the Stone Age, the Bronze Age, and the Iron Age).

According to Marshall McLuhan and the Toronto School of Communication Theory, a bet­ter ana­lys­is would be to view soci­et­al devel­op­ment based on the prom­in­ence of emer­ging com­mu­nic­a­tions technologies.

Marshall McLuhan - Cambridge University - Digital-First
Marshall McLuhan at Cambridge University, circa 1940.

McLuhan’s Four Epochs

McLuhan sug­gests divid­ing human civil­isa­tion into four epochs:

  • Oral Tribe Culture. Handwriting marks the begin­ning of the end of the Oral Tribe Culture. The Oral Tribe Culture per­sists but without its former prominence.
  • Manuscript Culture. Printing marks the begin­ning of the end of the Manuscript Culture, which per­sists but without its former prominence.
  • Gutenberg Galaxy. Electricity marks the begin­ning of the end of the Gutenberg Galaxy. The Gutenberg Galaxy per­sists but without its former prominence.
  • Electronic Age. Today, we reside in the Electronic Age. Possibly, we haven’t exper­i­enced the begin­ning of this age’s decline yet.

“The Gutenberg Galaxy is a land­mark book that intro­duced the concept of the glob­al vil­lage and estab­lished Marshall McLuhan as the ori­gin­al ‘media guru’, with more than 200,000 cop­ies in print.”
Source: Modern Language Review 6McLuhan, M. (1963). The Gutenberg galaxy: the mak­ing of typo­graph­ic man. Modern Language Review, 58, 542. https://​doi​.org/​1​0​.​2​3​0​7​/​3​7​1​9​923

The Electronic Age according to Marshall McLuhan.
“The Electronic Age,” accord­ing to Marshall McLuhan.

As a PR pro­fes­sion­al and lin­guist, I sub­scribe to the concept of the Electronic Age. I firmly believe soci­ety is unlikely to revert to the Gutenberg Galaxy.

  • Like the rest of soci­ety, the pub­lic rela­tions industry must go digit­al-first, too.

Read also: The Electronic Age and the End of the Gutenberg Galaxy

Annotations
Annotations
1 TikTok will have 750 mil­lion monthly users world­wide in 2022, accord­ing to Insider Intelligence’s fore­cast. It is now the third-largest social net­work of the Big Five world­wide (Facebook, Instagram, TikTok, Snapchat, and Twitter).
2 Nord, L., & Strömbäck, J. (2002, January). Tio dagar som skakade världen. En stud­ie av medi­ernas beskrivningar av ter­ror­at­tack­erna mot USA och kri­get i Afghanistan hösten 2001. ResearchGate; Styrelsen för psyko­lo­giskt förs­var. https://​www​.researchg​ate​.net/​p​u​b​l​i​c​a​t​i​o​n​/​2​7​1​0​1​4​6​2​4​_​T​i​o​_​d​a​g​a​r​_​s​o​m​_​s​k​a​k​a​d​e​_​v​a​r​l​d​e​n​_​E​n​_​s​t​u​d​i​e​_​a​v​_​m​e​d​i​e​r​n​a​s​_​b​e​s​k​r​i​v​n​i​n​g​a​r​_​a​v​_​t​e​r​r​o​r​a​t​t​a​c​k​e​r​n​a​_​m​o​t​_​U​S​A​_​o​c​h​_​k​r​i​g​e​t​_​i​_​A​f​g​h​a​n​i​s​t​a​n​_​h​o​s​t​e​n​_​2​001
3 “Sufficient algorithm per­form­ance is one of the major cost factors to con­sider because accur­ate pre­dic­tions require sev­er­al rounds of tun­ing ses­sions, which raises the cost of imple­ment­ing AI solu­tions. The high­er the accur­acy and effi­ciency of the AI pre­dic­tions, the more the cost of imple­ment­ing these solu­tions.” Source: Analytics Insight
4 “Human beings are most often an integ­rated part of com­plex sys­tems. In order to describe such a sys­tem with appro­pri­ate accur­acy it is neces­sary to mod­el the human com­pon­ents with the same accur­acy as the tech­nic­al com­pon­ents. Human factors must be included and mod­elled with the same degree of pre­ci­sion as the sys­tem’s mech­an­ic­al parts.” Source: Human Factors in Complex Systems The Modelling of Human Behaviour
5 Silfwer, J. (2021, December 7). The Silent Switch — A Stealthy Death for the Social Graph. Doctor Spin | The PR Blog. https://​doc​tor​spin​.net/​s​i​l​e​n​t​-​s​w​i​t​ch/
6 McLuhan, M. (1963). The Gutenberg galaxy: the mak­ing of typo­graph­ic man. Modern Language Review, 58, 542. https://​doi​.org/​1​0​.​2​3​0​7​/​3​7​1​9​923
Jerry Silfwer
Jerry Silfwerhttps://doctorspin.net/
Jerry Silfwer, alias Doctor Spin, is an awarded senior adviser specialising in public relations and digital strategy. Currently CEO at Spin Factory and KIX Communication Index. Before that, he worked at Kaufmann, Whispr Group, Springtime PR, and Spotlight PR. Based in Stockholm, Sweden.

The Cover Photo

The cover photo isn't related to public relations obviously; it's just a photo of mine. Think of it as a 'decorative diversion', a subtle reminder that it's good to have hobbies outside work.

The cover photo has

Lake in winter landscape.
Lake in winter landscape.

.

Subscribe to SpinCTRL—it’s 100% free!

Join 2,550+ fellow PR lovers and subscribe to Jerry’s free newsletter on communication and psychology.
What will you get?

> PR commentary on current events.
> Subscriber-only VIP content.
> My personal PR slides for .key and .ppt.
> Discounts on upcoming PR courses.
> Ebook on getting better PR ideas.
Subscribe to SpinCTRL today by clicking SUBSCRIBE and get your first free send-out instantly.

Latest Posts
Similar Posts
Most Popular
Doctor Spin | The PR Blog
The PR Blog
Contact: jerry@doctorspin.net
Facebook Instagram Linkedin TikTok Twitter
Web and design by Doctor Spin