Likes, Views, Saves: Which Signal Matters Most to the Algorithm?

Not all interactions are worth the same to a recommendation system, and the ranking between them follows a principle simple enough to be useful: the more an action costs the viewer, the more it tells the system.

A like costs a thumb movement. A share costs reputation — you are attaching your name to something in front of people you know. The system weighs them accordingly, and that single idea predicts most of the hierarchy.

The hierarchy

Approximate, since no platform publishes weightings — Meta’s Transparency Center, TikTok’s recommendation documentation and YouTube’s How YouTube Works describe the inputs without quantifying them. What follows is the consensus reading of stated inputs plus observable behaviour, labelled as inference because that is what it is.

Why saves are the interesting one

A save is a statement that the content has future value. On a platform optimising for time spent, that is close to a direct prediction of the thing being optimised for — and it is unusually honest, because nobody saves something to be polite.

Saves are also the most improvable signal, which is what makes this actionable rather than merely descriptive. Likes are largely a function of whether people enjoyed something, which is hard to engineer. Saves are a function of whether the content will be needed again, and that is a format decision:

• Reference material — lists, comparisons, specifications.
• Instructions somebody will follow later rather than now.
• Resources: tools, sources, places.
• Anything with a number in it that people will want to check.

The corresponding trap: content that is saved but not watched. On video surfaces, a save with a two-second view is a weak overall signal, because completion is weighted heavily and you failed it.

Watch time, where video is involved

On Reels, TikTok and YouTube, watch time is usually the dominant input, and it has two components that behave differently.

Completion rate — the proportion who reached the end. Short content has a structural advantage here, which is the real reason very short videos over-perform: not that brevity is inherently better, but that completion is easier.

Rewatches — extremely strong, because they are unambiguous. A viewer who watches something twice was not being polite.

The practical consequence is that the first two seconds are the whole game. A viewer who scrolls past at 0.4 seconds registers as a failure on the metric that matters most, and nothing later in the video can compensate for a signal that was never generated.

Comments, and why they are less valuable than they look

Comments require effort and are weighted well. But the weighting appears to consider quality, not just count — length, whether they are replies to each other, whether the author responded.

Which is why comment pods and “comment below!” tactics under-deliver. Fifty comments saying “🔥” are cheap to produce and read as cheap. A genuine exchange under a post is a different signal.

If you want comments, ask something with a cheap answer that people have an opinion about. “Which of these two?” outperforms “what do you think?” by a wide margin, because the second one requires composing something.

Likes: the signal everyone optimises for and nobody should

Likes are the most visible interaction and the least informative. They cost the viewer nothing, they are given reflexively, and they are the cheapest thing to purchase — which are all the same fact from different angles.

They still count for something. They are simply the weakest of the countable signals, and the amount of attention they receive is a legacy of them having once been the only public number.

The connection to buying engagement

This hierarchy explains why purchased engagement does so little, and it is worth spelling out.

What is cheap to buy — likes, views, follows — maps precisely onto what is weakly weighted. What is heavily weighted — completion, rewatches, saves, shares to real people, substantive comments — is either impossible to fake or expensive enough that nobody sells it at scale.

This is not a coincidence. Signals are weighted by how hard they are to produce insincerely, because that is what makes them informative. A system that weighted likes heavily would be trivially gamed, so it does not.

The consequence: buying the cheap signals moves numbers on a screen while the inputs that decide distribution stay flat, and your ratios get worse because the audience denominator grew. What an SMM panel is covers what is actually on offer, and organic vs paid growth compares the two honestly.

What to optimise, in order

1. The first two seconds. Every other signal is downstream of not being scrolled past.
2. Completion, on any video surface. Cut everything that is not needed.
3. Saves. Ask directly, and make content that will be needed again.
4. Shares. Hardest to engineer. Content gets shared when sharing it makes the sharer look good.
5. Comments, via questions with cheap answers.
6. Likes. They will happen. Do not build around them.

Measuring it

Most platforms expose saves, shares and completion in their own insights, and those are the numbers to watch — not the like count on the front of the post.

Track saves and shares as rates against reach rather than as raw counts, for the same reason engagement rate is a ratio: a raw number moves with distribution and tells you nothing about the content. Engagement rate covers the formulas, and profile visits and Explore reach covers what happens after someone taps through.

The principle underneath this hierarchy — that signals are weighted by how hard they are to fake — runs through our guide to social media growth.