Nobody outside these companies knows the ranking functions. Not agencies, not the people selling algorithm courses, and not us. What exists is a set of published statements, a much larger set of things practitioners observe, and a very large set of confident nonsense. This page separates them, and links to a per-platform breakdown for each.
The short version: every major feed does the same four things. It assembles a pool of eligible content, reads signals about you and the content, predicts what you are likely to do, and sorts by that prediction. The platforms differ in what they predict and how much they admit.
The four steps every feed follows
Meta describes this most explicitly, in its Transparency Center, as inventory, signals, predictions, and ranking by score. LinkedIn describes the same shape with different words, splitting it into retrieval and ranking: “retrieval determines which posts reach the ranking stage, ranking determines what a member actually sees.” TikTok and YouTube describe the outcome rather than the pipeline, but the structure is not in dispute.
| Step | What happens | Why it matters to you |
|---|---|---|
| Inventory | Eligible content is gathered; policy-violating and ineligible content is excluded | Being ineligible costs you everything, and it is documented, not secret |
| Signals | Thousands of features about you, the post, and the poster | Most are about the viewer, not about you |
| Predictions | Models estimate what you will do: watch, comment, share, hide, report | Negative predictions are inputs too |
| Ranking | Scores are combined and sorted, with diversity constraints | There is no single number you can optimise |
Two consequences follow, and they explain most of the confusion. First, ranking is per viewer, so “the algorithm” is not showing your post to an audience, it is deciding independently for each person. Second, because predictions include negative actions, a post can lose distribution without breaking any rule.
What each platform actually confirms
The level of documentation varies enormously, and knowing which platform tells you what is more useful than any tactic.
| Platform | How much is documented | The clearest first-party statement |
|---|---|---|
| TikTok | Most | Follower count and prior video performance are stated as not direct factors; finishing a longer video is called a strong indicator |
| High, on process | Four-stage pipeline, 100+ prediction models, and a published list of demoted categories | |
| Medium | “Instagram doesn’t have a singular algorithm”; separate systems for Feed, Stories, Explore and Reels | |
| High, on architecture | Semantic retrieval plus a sequential transformer over 1,000+ past interactions, published March 2026 | |
| YouTube | Medium | Home ranks on how well a video “interested and satisfied similar viewers” |
| Medium, unusually practical | Duplicate saving can get you flagged as spam; keyword-rich metadata gets more distribution | |
| Threads | Least | No dedicated ranking explainer exists |
Note: Statements here were verified against each platform’s own documentation (Meta Transparency Center, about.instagram.com, TikTok Newsroom and Transparency Center, LinkedIn Engineering, YouTube Help, Pinterest Business help) as of September 2026. Platforms change these without notice.
What changed most recently
Confirmed: LinkedIn published a full feed rebuild on 12 March 2026, moving retrieval to an LLM dual-encoder and ranking to a sequential transformer. That is the most substantive first-party algorithm disclosure of the year from any major platform.
Confirmed: every large feed now mixes recommended content from accounts you do not follow into what used to be a following-only surface. This is a product decision, not a ranking subtlety, and it is the reason follower count predicts reach less well than it used to.
Widely reported, not confirmed: specific shares of recommended versus followed content, on any platform. Numbers circulate. None of them are published.
Four myths worth killing
Shadowbans. No major platform confirms a mechanism by that name. Several publish something better: an actual list of what they demote and why. Meta’s demotion categories cover clickbait, engagement bait, fact-checked misinformation and likely policy violations. Pinterest says duplicate saving can get you spam-flagged. TikTok and YouTube publish recommendation eligibility rules. If your reach dropped, one of these documented causes is far more likely than a secret penalty, and unlike a shadowban you can check it. Our glossary covers what a shadowban actually is.
Engagement pods. Unfounded everywhere, and the reason is the same on every platform: ranking is per viewer. A group that engages with everything teaches the model that the group engages with everything, which carries almost no information about who else should see your post. On Meta and YouTube it also runs into explicit policy.
“The algorithm hates links.” The most durable myth in social media and the least supported. Not one platform publishes a demotion for containing a URL. Meta demotes clickbait links, which is about deception. Pinterest requires the link to work. The real effect, where it exists, is mechanical: people who leave do not engage, and engagement is what gets predicted. Write a post that stands alone and put the link where it reads naturally.
Posting time is decisive. It is not, and it matters less every year, because fewer feeds are chronological. It matters most on Threads, where For You skews recent, and least on Pinterest and YouTube, where content accumulates views for months. Test windows rather than trusting a chart: our best time to post on social media guide and the posting time tool give you a starting point in your own timezone.
What actually moves distribution, in priority order
- Being eligible. Read the demotion and recommendation guidelines for your platforms once. Most people never do.
- The first two seconds, or the first two lines. Every system on this list is conditioned on someone not scrolling past.
- Format fit for the surface. A Reel, a carousel, a Short and a Pin are different products, not the same idea resized.
- Actions that cost the viewer something. Sends, shares, saves and substantive comments. These are weighted above passive taps everywhere anyone has said anything.
- Topical clarity. Every one of these systems works by finding people similar to those who already liked your work. Make that group identifiable.
- Consistency. Not a ranking signal on any published list. It is how you accumulate enough attempts for anything above to matter.
What to do
Pick two platforms rather than seven. Read their published guidelines. Decide, per post, which surface you are making for. Judge your results across twenty posts rather than one, because per-viewer ranking makes single-post variance enormous. Ignore anyone quoting a percentage they cannot source.
Doing this across platforms without doubling the work
The practical problem with all of the above is that “format fit per platform” and “post consistently” pull against each other. BulkPublish schedules to 15 platforms from one calendar with per-platform text and media, so a piece of content can be adapted per surface once and queued, rather than reformatted by hand every time. The scheduling feature page covers the workflow, and pricing covers the plans.
The parts you control
You do not control the ranking function. You control three things, on every platform, in every year: the format you publish in, the hook that decides whether anyone stays past the first moment, and whether you keep going long enough for the numbers to mean something. Every honest piece of algorithm advice is a restatement of those three. Everything else is a story about a black box that the storyteller has not seen either.