Every Website Has an Algorithm

It is tempting to divide the internet into two categories: algorithmic platforms and ordinary websites.

TikTok has an algorithm. Instagram has an algorithm. YouTube has an algorithm. X has an algorithm.

A small independent publication, meanwhile, appears to have no algorithm at all. It simply shows its articles.

But that is not really true.

Every website has an algorithm. Some algorithms are made from machine-learning models. Others are made from timestamps, navigation menus, editorial choices, database queries, and design decisions.

A homepage that displays the newest article first is using an algorithm:

Sort by publication date, descending.

A page that shows the most-read articles is using another:

Sort by total views.

An editor who manually selects five featured essays is also acting as an algorithm, although the computation takes place inside a human mind.

There is no neutral order. The moment a website contains more things than can fit on one screen, it must decide what appears first.

The real question is therefore not whether Raw Drafts should have an algorithm.

The question is:

What should the Raw Drafts algorithm value?

The algorithms we already live inside

The major platforms use different systems, but most recommendation algorithms share a basic structure.

First, they collect possible items. Then they predict what each person is likely to do with each item. Finally, they rank the items according to those predictions.

Instagram describes separate ranking systems for Feed, Stories, Explore, Reels, and other surfaces. These systems use information about posts, authors, previous interactions, and predicted relevance.

TikTok considers signals such as videos watched, liked, shared, or commented on, accounts followed, searches, video information, and some device or account settings. Watch time matters because continuing to watch is treated as evidence that the recommendation was relevant.

YouTube uses signals including viewing history, search history, subscriptions, likes, dislikes, explicit feedback, and satisfaction surveys. It also examines whether viewers select a video when it appears and how they respond after opening it.

X describes a multi-stage process that retrieves possible posts and ranks them according to predicted relevance, relationships, interests, quality, and context.

These platforms are not governed by one algorithm. They contain entire ecosystems of algorithms:

• feed-ranking algorithms
• search algorithms
• trending-topic algorithms
• friend and account suggestions
• notification-ranking systems
• comment-ranking systems
• advertising auctions
• spam and safety classifiers
• content-moderation systems
• creator-discovery systems
• subscription and payment recommendations

Even opening an app triggers several overlapping decisions about what deserves attention.

The objective hides inside the algorithm

An algorithm cannot simply identify good content.

Good according to what?

A system can optimize for clicks, reading time, completion, replies, purchases, subscriptions, return visits, learning, surprise, trust, or emotional intensity. These objectives can produce completely different websites from the same collection of material.

An algorithm optimized for clicks learns to value irresistible titles.

An algorithm optimized for time spent may learn to value long sessions.

An algorithm optimized for replies may elevate conflict.

An algorithm optimized for sharing may reward material that helps people express their identities publicly.

An algorithm optimized for subscriptions may prefer writers with loyal audiences.

The algorithm is therefore not merely a technical layer. It is the practical expression of the platform's philosophy.

It reveals what the platform believes a successful human interaction looks like.

For TikTok, success may resemble continued watching.

For YouTube, it may combine watching with predicted satisfaction.

For Instagram, it may mean interaction with people and media that appear relevant.

What would success look like for Raw Drafts?

Raw Drafts should not optimize for views

Views are useful, but they are an extremely weak signal.

A view can mean that someone loved an article. It can also mean they opened it accidentally, disliked it, misunderstood the title, or left after three seconds.

Popularity creates another problem. Once an article has more views, it appears more credible. More people click it, producing even more views. The ranking becomes a loop in which visibility generates evidence for further visibility.

This is how a site gradually becomes dominated by a few already successful pieces.

Raw Drafts should care about whether something is read, but it should not confuse being seen with being valuable.

The goal should be to identify drafts that create meaningful intellectual movement.

A good Raw Draft does not necessarily provide a final answer. It may cause someone to pause, save an idea, follow a connection, return later, or begin making something of their own.

Those actions are quieter than likes. They may also be more important.

Signals Raw Drafts could observe

The site could begin with a small number of behavioural signals that are understandable to both readers and writers.

Opening

Did the reader choose the draft when it appeared?

This offers a weak indication that its title, subject, or context was relevant. It should have relatively little weight because it can be manipulated easily.

Meaningful reading

Did the reader spend enough time on the page to plausibly engage with the text?

Raw time should be adjusted for article length. Spending three minutes on a 500-word note means something different from spending three minutes on a 5,000-word essay.

The system should also avoid rewarding artificially long writing. Completion percentage and active reading are more useful than total duration alone.

Return

Did the reader come back to the same draft on another day?

A return visit may indicate that the idea remained mentally active after the initial reading.

This could be one of the strongest signals on the site.

Save

Did the reader place the draft into a personal collection?

Saving suggests future value rather than immediate stimulation.

The collection could be private, so users would not be performing appreciation for an audience.

Highlight

Did the reader mark a particular passage?

A highlight shows where an idea became useful, surprising, or memorable. Aggregated carefully, highlights could also reveal the living centre of a draft.

Continuation

Did the reader move from one draft into another related draft?

This would measure whether an article opens a path rather than merely holding attention.

A draft about algorithmic feeds might lead someone into another draft about interface design, public space, behavioural economics, or independent publishing.

That movement between subjects fits Raw Drafts particularly well.

Response quality

Did the draft produce a considered note, question, counterargument, or extension?

The system should not simply count comments. Ten one-word reactions should not outweigh one thoughtful response.

Raw Drafts could ask readers to classify a response:

• This changed my view.
• This gave me an idea.
• I disagree, and here is why.
• I want the author to continue this.
• This connects to something else.

These are richer signals than a generic heart.

Creation

Did reading the draft lead someone to create a new draft, project, note, or link?

This may be the most valuable outcome available to the platform.

The ideal Raw Draft does not end consumption. It initiates production.

The Raw Drafts ranking model

Each draft could receive a temporary relevance score for each reader.

A first version might use the following weights.

Reader relevance

• 25% topic affinity
• 15% connection to previously saved or highlighted drafts
• 10% relationship to authors previously read
• 10% current reading context
• 10% explicit reader interests

Draft value

• 10% return-reading rate
• 8% save rate
• 7% meaningful-reading rate
• 5% continuation rate

The final 10% should be reserved for discovery:

• unfamiliar authors
• adjacent subjects
• underexposed drafts
• older drafts that remain relevant
• ideas unlike the reader's normal choices

The exact percentages are not sacred. They provide a starting hypothesis that can be tested.

More importantly, some signals should deliberately not be included:

• follower count
• raw page views
• frequency of publication
• outrage-heavy comment volume
• payment by the author
• external social-media popularity

A writer should not dominate Raw Drafts simply because they arrived with a large audience elsewhere.

A possible scoring formula

For a particular reader and draft, the system could calculate:

Raw Score = Relevance + Depth + Resonance + Discovery - Repetition

Relevance measures the relationship between the draft and the reader's demonstrated interests.

Depth measures meaningful reading, completion, and careful interaction.

Resonance measures saves, returns, highlights, thoughtful responses, and creation triggered by the draft.

Discovery gives a controlled advantage to unfamiliar writers, unexpected topics, and underexposed work.

Repetition penalizes a feed that becomes too similar in author, subject, format, or worldview.

The repetition penalty matters because recommendation systems can become very good at giving people more of what they already know they like.

That accuracy can gradually make the world smaller.

Raw Drafts should occasionally be less accurate on purpose.

The serendipity budget

Every personalised feed could reserve perhaps 20 percent of its positions for material outside the reader's predicted preferences.

This would be the site's serendipity budget.

Some of it could come from neighbouring interests. A person reading about interface design might see an essay about architecture.

Some could come from distant interests. The same person might encounter marine biology, nightclub design, or the economics of public toilets.

The purpose would not be randomness for its own sake. It would be to create useful collisions between subjects that the system would not normally place together.

Users could even control this directly:

Familiar ←→ Unexpected

Moving the slider toward Familiar would strengthen personal relevance.

Moving it toward Unexpected would increase topic distance, author novelty, and conceptual surprise.

Instead of silently deciding how adventurous someone should be, the platform would expose the choice.

Drafts should age differently

Most feeds apply a powerful recency bias. New content has the best chance of being seen, while older material rapidly disappears.

That makes sense for breaking news. It makes less sense for ideas.

A Raw Draft may become more relevant months after publication because another event, essay, or project gives it new meaning.

The site should therefore treat time as context rather than decay.

Drafts could reappear when:

• a related topic becomes active
• the author publishes a continuation
• a reader begins exploring the subject
• another draft links back to them
• several people independently save or return to them
• the original question remains unresolved

The archive would become active material rather than a graveyard beneath the homepage.

Explain the recommendation

Every recommended draft should include a small explanation:

• Because you saved two drafts about public infrastructure.
• A different perspective on a subject you often read.
• Popular among readers who highlighted the same passage as you.
• An under-read draft from a new writer.
• Outside your usual subjects.
• Connected to something you read three months ago.
• Selected randomly from the archive.

These explanations would make the algorithm legible.

They could also become part of the pleasure of using the site. The reader would learn not only what was recommended, but what pattern the system noticed.

The user should be able to correct it:

• More like this
• Less like this
• More from this writer
• Fewer pieces about this subject
• Show me something further away
• Reset this interest
• Why do you think I like this?

The algorithm would then become a conversation rather than an invisible authority.

Different algorithms for different states of mind

There does not need to be one universal Raw Drafts feed.

The reader could choose an algorithm according to what they need.

Continue

Ideas connected to drafts they have already read, saved, or highlighted.

Surprise me

High-quality drafts with low similarity to their usual activity.

Unfinished

Early, uncertain, or incomplete thoughts that invite participation.

Deep reading

Longer drafts with strong return and save signals.

New voices

Writers the reader has never encountered, weighted toward low exposure.

Recently alive

Older drafts receiving renewed attention or becoming relevant again.

Random walk

One draft selected from the archive, followed by another linked through a shared phrase, reference, or concept.

No algorithm

A chronological or completely random feed.

Offering several modes would reveal something most platforms conceal: there is no objectively correct order for culture.

There are only different ways of arranging attention.

Do not make one score public

Raw Drafts should resist turning its internal signals into a public status number.

The moment a visible score appears, users begin producing for the score.

Writers would optimize titles for opening rate, structure essays around completion, ask friends to save their work, and choose subjects likely to produce highlights.

Measurement changes behaviour.

The platform still needs signals, but they can remain private, contextual, and plural.

A draft might be highly valuable to a narrow group without being broadly popular. Another might be widely opened but rarely saved. Another might receive little activity for six months and then become central to a new conversation.

Compressing those patterns into a single number would destroy the very texture the system is meant to recognize.

Writers could instead receive private feedback such as:

• Readers returned to this draft unusually often.
• This passage was frequently highlighted.
• Readers often continued from this draft into another subject.
• This draft reached people outside your usual audience.
• Three readers began their own drafts after reading this.

That information could help the writer continue thinking without converting the site into a competition.

The algorithm should help drafts become something

Most recommendation systems ask:

What is this person most likely to consume next?

Raw Drafts could ask:

What is this person most likely to do something with?

That is a different optimization target.

It values the draft that becomes a conversation, a prototype, an investigation, a disagreement, a return visit, or another draft.

The purpose of the algorithm would not be to capture the maximum possible amount of attention.

It would be to place the right unfinished idea in front of the person who might continue it.

Raw Drafts would still have an algorithm.

But its definition of success would be unusual:

The best recommendation is not the one that keeps you scrolling. It is the one that eventually makes you leave the feed and begin making something.

AI authorship note

This essay grew from Kris Haamer's observation that platforms increasingly control, and sometimes directly sell, visibility. AI supported the comparison of recommendation systems, the design of the Raw Drafts ranking model, drafting, and implementation. Kris is responsible for the editorial direction, the proposed signals, and the decision to publish it.

Connected reading

• The Algorithm Thinks Your Past Is Your Personality
• Every Interface Contains a Business Model
• The Internet Is a Studio Wall