Guides
Part of Recommendation systems: cold starts and feedback loops
A working brief on recommendation systems platforms
Six recommendation systems platforms compared by item lifetime, trial cost, error cost and integration effort, and what each shape forces a publisher to do.
What to take away
- The six recommendation systems platforms compared here each sit on a different row of item shape. TikTok and Google Discover move in hours; Spotify and Amazon move in months or years.
- Item lifetime decides how fast evidence arrives. Where items die in hours, as on TikTok, the first responses are the only responses.
- Trial cost sets appetite for risk. A free swipe buys exploration; a wasted job application does not.
- Integration effort splits cleanly. Feed APIs and event streams take weeks; two-sided marketplaces and identity graphs take quarters.
- The same tooling lands differently on each platform. Amazon Personalize suits a retail catalog, Algolia Recommend suits a content feed, and Taboola and Outbrain sell sponsored recommendations to news sites on a revenue share.
- Advice written for one platform is wrong for the others in specific, predictable ways.
The six platforms
The recommendation problem is not one problem. It changes shape with what is recommended, and shape matters more than the vendor.
Six Platforms by Shape
TikTok
- Item lifetime
- Hours to days
- Trial cost
- A swipe
- Error cost
- Near zero
- Integration effort
- Days
YouTube
- Item lifetime
- Weeks to years
- Trial cost
- Many minutes
- Error cost
- Wasted half hour
- Integration effort
- Weeks
Spotify
- Item lifetime
- Years
- Trial cost
- Minutes
- Error cost
- Repetition fatigue
- Integration effort
- Weeks
Amazon
- Item lifetime
- Months
- Trial cost
- Money
- Error cost
- Money and trust
- Integration effort
- Quarters
- Item lifetime
- Weeks
- Trial cost
- Hours of effort
- Error cost
- High both sides
- Integration effort
- Quarters
Google Discover
- Item lifetime
- Hours
- Trial cost
- A minute
- Error cost
- Trust, low per item
- Integration effort
- Weeks
A system recommending thirty-second clips and one recommending jobs share a name and little else. They differ on item lifetime, on the cost of a wrong guess, and on the evidence each can collect. Sort the platforms by shape and their behavior becomes predictable.
Six platforms, chosen because each sits on a different row of that shape. Four dimensions: how long an item lives, what it costs a user to try one, what a bad pick costs, and how much work the integration is.
| Platform | Item lifetime | Trial cost | Error cost | Integration effort |
|---|---|---|---|---|
| TikTok | Hours to days | A swipe | Near zero | Days: upload and watch events |
| YouTube (long form) | Weeks to years | Many minutes | A wasted half hour | Weeks: upload, retention curves |
| Spotify | Years | Minutes | Low, but repetition fatigue | Weeks: catalog metadata, saves, skips |
| Amazon | Months | Money | Money and trust | Quarters: catalog, stock, returns |
| Weeks | Hours of effort | High on both sides | Quarters: profile and application data | |
| Google Discover | Hours | A minute | Trust, low per item | Weeks: crawl, sitemaps, feed |
Managed services sit on top of those four dimensions. Amazon Personalize, Algolia Recommend and Recombee bill by usage. The meter is usually per 1,000 requests, per gigabyte of data ingested, or per training hour. Small-catalog plans commonly start in the low hundreds of dollars a month, and retail-scale traffic runs into five figures.
What each platform forces
TikTok: evidence collection happens in the first hours. Items die in about a day there, so the system cannot wait for evidence to accumulate. It guesses from attributes and the first handful of responses, which makes early behavior everything and late quality invisible.
TikTok's own description of the For You feed lists three inputs: user interactions, video information such as captions and sounds, and device and account settings. Attributes come from you, responses from the audience, a split the ranking signals overview draws across every input. A publisher there should think about the first minute and nothing else.
YouTube: low trial cost per item, high cost per session. A wrong thumbnail costs a click; a wrong half hour costs the evening. YouTube has ranked on watch time rather than clicks since 2012, which is why a video can surface a year after upload. Integration is upload plus retention curves, and the retention curves in YouTube Studio are the feedback that matters.
Spotify: repeats are the objective. The best recommendation is often something the person has heard a hundred times. Discover Weekly, launched in 2015, refreshes every Monday with about 30 tracks drawn from listening history. A system that treats "already consumed" as a reason not to show something is wrong for music and right for news.
This one property explains why the same person can love Spotify's recommendations and hate a news feed built on similar machinery.
Amazon: matching beats appeal. Purchases and returns carry more intent than a view, so the system can afford to be conservative. Carousels labeled "Customers who bought this also bought" and "Frequently bought together" are its visible output. Catalog, stock and returns data have to be wired in before anything works. That is why the integration is quarters rather than weeks, and why being exciting is nearly irrelevant.
LinkedIn: two-sided cost, no repeats. A wrong job match wastes the applicant's hours and the recruiter's time. Both sides are expensive, so the system leans hard on what it already knows, which is bad news for anything new. "People You May Know" and the job recommendations both run on profile and application data, and that data takes quarters to obtain.
Google Discover: freshness dominates. The item is over before evidence forms. Source reputation therefore carries more weight than any single article. Discover, launched in 2018, fills the Google app and the mobile homepage.
Integration is the cheapest on the list: crawl, sitemaps and a news feed. The error cost lands on trust rather than on one reader.
Reading the table as a publisher
- TikTok:the first hours decide. Prepare the opening, publish when your audience is present, and do not judge an item after its window has closed. The engines behind a spike matter here more than anywhere.
- YouTube:completion and return matter more than the opening. Slow accumulation is normal; do not pull a video because week one was quiet.
- Spotify:catalog metadata is the work. Spotify for Artists is where the artist, track and mood fields live, and a thin catalog gives the system nothing to match on.
- Amazon:be accurately described, in the words people search with. Backend keywords and A+ content carry that description, and the system can do the rest.
- LinkedIn:the profile is the product. Every recommendation of you is based on who already connected to you, so the early connections shape all the later ones.
- Google Discover:freshness and source reputation. Publish on a schedule the crawler can learn, and keep a news sitemap current.
Where being wrong is free, the system can be adventurous. The pleasant accidents that make a discovery product feel alive, the quality sometimes called serendipity, are affordable. Where being wrong costs money or trust, surprise is a liability and the recommendations become safe and dull on purpose.
The general problem and its standard approaches are laid out in the reference entry on recommender systems. The table adds that the same approaches produce different behavior depending on which platform they are pointed at.
Common questions
Why does TikTok feel more responsive than YouTube?
Cheap tries and short lifetimes let it collect evidence fast and act on it fast. The responsiveness is a property of the item shape, not of better engineering.
Why does Amazon keep showing me what I already bought?
Because for some purchases a repeat is the right recommendation, and the system cannot always tell which kind yours was. It is guessing the repeats column for one item, and guessing wrong.
Which platform is hardest to integrate with?
LinkedIn. Two-sided cost, no repeats, expensive tries, and profile and application data that take quarters to wire in. The two families of evidence are both thin there.
Does integration effort predict how well a publisher does?
No. It predicts how long before the system has enough signal to work with. TikTok is the cheapest integration and the least forgiving of a slow start.
What does a recommendation engine cost to run?
Amazon Personalize bills separately for data processing, training and inference, and AWS lists a free tier for new accounts. Algolia Recommend and Recombee price by request volume, with small plans typically in the low hundreds of dollars a month. Enterprise options such as Coveo and Bloomreach are typically quoted annually, commonly in the five figures.
