tldr : Great content only reaches the right audience if the recommendation system can reliably understand both the content and the audience well enough to connect them.
When we make a really good video, we expect that algorithm would pick it up and show it to people who will like it and we get the results we want
For that first it would need to parameterize your video then compare it with the data set of people it has and then decide if it’s worth it to push it to matching profiles
And algorithm is not a god. It’s not like it can infer everything about people from their digital footprint or data points they leave behind.
So the clear constraints in this system are
- if the people you want this video to reach to are even on this platform.
- Is the behavior that you are trying to target (of those people) even inferable by the algorithm?
- Are the parameters of your video inferable by the algorithm?
- Is there even a match between the parameters of your video and the people it has
- If there is a match is it worth pushing your video to them or should it just push some safe content like cat videos
A lot of times we expect magic from it that we will upload some obscure video and expect it to deliver it to the people we want as so that we can get the fame or the economic benefits. But would it be even possible for algorithm to find your tribe?
You have to see what kind of data points your icp might be leaving on the internet so that you can target them
because sometimes the behavior we might think is really obvious to notice might not be for the algorithm
like imagine you make a video where you talk about how great the color green is
there definitely are people who like green
but what if they never displayed any behavior that helps the algorithm to infer that they like the color green
how would it recommend your video to them
Now same can happen when it comes to the business video made or the sports match analysis you did or that obscure tech magazine you were talking about
Successful distribution happens, when your video can be understood by the system properly, your target audience is on the platform and is displaying behavior that can help the algorithm conclude that they will like your video. And finally, it is cool enough that it can take the risk to show you video instead of something else.
A video cannot reach its audience merely because that audience would enjoy it. The platform must be able to identify both the video and the audience using observable signals, then confidently match them.
Therefore, theoretical TAM is not enough. You need algorithmically legible TAM.
I think this can establish the same difference as knowing what an AI model can do and then prompting accordingly versus just saying, Claude make no mistakes, make a billion dollar business
I think a piece of content on social media is almost like a prompt to an LLM
you are giving it a set of data and you’re expecting that it should reach the right people
so you need the right prompt which considers the limitations of the system itself, but also delivers results you want
A social media algorithm is also a generative algorithm if you want to bend the definition, that you give it a piece of content and context and it will generate a list of accounts it should show the video to at this moment
basically make video for an audience that is highly available on that platform you are posting in a language they understand and know hopefully that language is also popping off in their culture and show face and have cool audio
if you are making content it is to mostly get fame and earn money. so first thing is you have to choose between two options: do you want clients via this social media fame or do you want brands? because both are different things. you don’t see a brand sponsoring service companies because they are good at it or whatever. so if you want clients, you’ve got to have a show of competence and knowledge while if you want brand sponsorships, you have to show personality and your own brand so it depends on the objective.
I would say it’s easier to go for brands because the cost of entry is way lower. you don’t need way too high of expertise even a beginner showing his journey can get brand deals but not clients or maybe he can but you get the vibe.
also I just realized that it must be f*cking insane to run the algorithm for billions of videos every second
so that means that this algorithm is run N number of times only per video right or am i tripping
and every video might have a different N but it’s not like this algorithm is being run on the whole database every moment
there must be intervals and there must be a set on which it is being run and on which it is not being run otherwise it is too computationally expensive
When a user opens the app, the platform needs to generate the next batch of videos for that user. It does not inspect every video on the platform. First, it uses everything it already knows about the user to pull a smaller set of plausible videos.
User opens app → system retrieves plausible videos → filters them → ranks them → shows the next N videos → user reacts → system updates its understanding
Recommendation usually comes from two sources:
What the system knows directly about you, such as your past behaviour.
What it can infer from patterns across other users, such as people like you also liking something else.
Suppose you love the color green, but you have never searched for green objects, watched green-themed videos longer, liked them, followed creators associated with them, or behaved differently around them. The system has no explicit “this person likes green” signal.
Thus, it has to be information which the system already contains that drives this algorithm
What it does not know
It cannot use to recommend
Therefore you are cooked if you use things in your videos that algorithm does not know yet or has no ability to infer
So, how do we come to know what the algorithm already knows? Well, since inferred information is driving recommendation, you can just look at videos being recommended to you and understand that okay, these are all the parameters that algorithm understands. Usually it’s something like clear audio, face being visible in video, some hook or suspense going on, whatever.
You might notice clusters built around:
the person or type of face shown;
spoken language, accent, tone, and audio quality;
topics, named entities, countries, products, celebrities, or cultural phrases;
visual settings such as offices, gyms, bedrooms, streets, studios, or cars;
formats such as commentary, tutorials, reactions, transformations, stories, or interviews;
emotional structures such as suspense, outrage, aspiration, fear, curiosity, or validation;
production signals such as captions, pace, framing, music, lighting, and visual clarity.
The important thing is not merely that successful videos possess these qualities. It is that the platform appears able to repeatedly group and distribute content using them.Observe repeated recommendations → identify recurring attributes → infer the audience cluster being constructed → create content using attributes the system already routes reliably
This is also why trends are valuable. A trend is partially a publicly visible proof that the system understands a particular audio, format, phrase, aesthetic, or subject well enough to route it repeatedly.
The end message is not that here go copy trains or copy people, I just want to say that look at the algorithm, you will understand what it understands, what are the invariants in all the videos that reach your for you page and incorporate those invariants inside your videos instead of blatantly copying anything.