Home | Search

2026-01-29

Ads versus recommendations

Disclaimer

Recommendations are expected positive value for you.
Ads are expected negative value for you (you're only watching them because they're interspersed with recommendations that are positive value).
The first N users who see some content, see it as an ad. This teaches the recommendation algo who wants to see the content. Everyone else afterwards who sees it, sees it as a recommendation.
If AI recommendation algos improve, N will reduce. Platforms won't need as many test users to see some content, before they can accurately guess who wants to see what.
If there is increased competition between various AI recommendation algos and the content hosts, N will reduce. It will become harder for any one platform to monopolise all the content and then forcibly show you ads you don't want to see.

Ads

Why are they watching it - interspersed with some other content that the user expects is positive value for them
Usefulness for user - for most users, the information in the ad is of zero value. reading the ad is net negative value for them because of lost attention. for a few users, the information in the ad is of positive value.
Profitability for the platform - it's possible that adding ads to your platform reduces the net value you provide to users via information, but increases the profit you gain

Recs

Why are they watching it - they expect positive value from it
Usefulness for user - for few users, the information in the rec is of zero value. reading the rec is net negative value for them because of lost attention. for many users, the information in the rec is of positive value.
Profitability for the platform - if content on your platform is correctly recommended to many new users, then your platform itself is correctly recommended to many new users

Make recs not ads

Improving targettedness of recommendations can actively grow your platform. Improving targettedness of ads will only actively kill your platform less, until it hits the SNR threshold where it becomes a recommendation not an ad.

Ideally all ads are recs

Ideally when an ad is created, the first N users to look at the ad are all test subjects. It's possible most of them don't get value by looking at the ad.
However, once this is done, you now have a targeted user profile for who is going to want to watch this ad. You can now make targeted recommendation of this content on people matching this user profile.
The whole challenge of a recommendation algo is converting ads into recs using as small N as possible. In an ideal world, N is zero, but gpt-5.2 isn't that good yet (I have tried).

How to reach everyone

If you personally have some reason you want to reach anyone (maybe you want to be a successful politician or religious leader or philosopher or similar), you have two approaches.
- One approach is to make the content so good that people (or AI) recommend the content enough, because everyone wants to see it and gets actual value from me.
- Another approach is to make some other content or app very good (may or may not be related), and then put your desired content as an ad in it.
Sandwiches
- Ideally you want to create content so good it becomes popular via organic recommendations.
- This however can be hard as most people by default only value content of a few types such as entertainment.
- In this case, you can end up with a strategy where you sandwich some content that the user wants to see (rec) with the content you want to push on the user (ad).
- You can trust the social media discovery algo to create sandwiches, or you can yourself create a sandwich by including both types of content in the same video/app/etc.

Everyone agrees that organic reach is better than ads, whether you're making software or games or videos or blogs or whatever.
- However, some people think ads can help in the latter stages of increasing reach once organic reach has already been solved for.
- Some other people think ads are basically entirely useless.
I'm increasingly starting to join the camp that says ads are useless.
I want actual data, l've seen too many people's guesses with no data backing them.

In a theoretical maximally competitive market, all ads should be useless.

Given a piece of content, AI should be able to accurately guess whether you will like it or not. If it makes bad guesses, you will switch platforms.

In practice, making accurate guesses is hard and the AI needs to show it to a few test users first.

If the AI guesses that you won’t like it, it should not show that content to you. If it shows you, you should switch platforms.

In practice, the AI might sometimes show you the content anyway, because it knows there’s other content you want to see on the same platform. Also, you can’t just switch platforms because all the data is locked into that specific platform.

Okay wait, so data lock-in is the only reason ads work at all? (Assume we had really good AI recommeders for now.)

The platform that owns the data knows they can shove the users with X amount of content they don't want to see as long as it is sandwiched with Y amount of content they want to see.
Then the platform can use this to either shove their own content on the user, or sell this right to advertisers.
This still doesn't explain why advertisers find it profitable to run these ads. Why does there exist any price point above zero, at which advertisers find it worth it to pay for ads?
- Funny model that might (or might not) have a kernel of truth: Assume Alice Airlines is selling air conditioners in Norway. Assume literally nobody in Norway wants to purchase air conditioners. Assume 1 in 1000 people in Norway is a Moron who will buy anything that ads on their TV/phone/etc show them, regardless of whether they actually need it. Assume Alice Airlines sells air conditioners at $100 profit margin per unit. It now makes sense for Alice Airlines to pay $0.10 / user on untargeted ads (targeted at all users in Norway). It could make sense for Alice Airlines to pay $1 / user on targeted ads, if the platform could narrow down some 100 out of these 1000 users that are likely to contain most of the Morons on their platform. It could make sense for Alice Airlines to pay $100 / user on targeted ads, if the platform had sufficiently good AI that could identify Morons directly.

corollaries

Corollary - Suppose we could scrape entire social media and run our own recommendation algo on that. If the algo is sufficiently good (as per whatever users think is good), then that entire social media platform will die. Making scraping easier makes it easier to kill social media platforms, who are the only ones who can sell advertising.
- Many social media platforms are more-or-less scraped already. Reddit, discord, substack are heavily scraped. Insta, facebook, whatsapp etc seem more successful in blocking scrapers.
- Bluesky scraped data explicitly includes lots of info on who liked what.
- If all I need for making a good recommendation algo is stale data of which users tend to like which posts, then I already have that data.
- What if I literally just implement this? I now need 3 mandatory layers - comparing other user's likes, embedding search using dumb model and inference using smart model and smart prompt. I can actually just implement this entire thing as a bluesky feed.
Corollary - A truly open source social media platform should also encourage the first N users to 'like' any post to make their likes public, as this like data is useful for future recommendation algos. This platform should not try to monopolise this data.

TO DO

read more on each of the bluesky feeds that are popular as of 2026, and how much do they rely on other people's likes, versus AI reading the content directly

Enter email or phone number to subscribe. You will receive atmost one update per month

Enter comment