Why is writing for the algorithm bad?

Why is writing for the algorithm bad?

Writing for the algorithm becomes bad when the system you are trying to please starts making editorial decisions that should belong to the reader and writer. Optimizing a useful article so people can find it is sensible. Choosing what to say, how to say it, or whether to say it primarily because you expect an algorithm to reward it is different. The result is often writing designed around measurable reactions rather than the value it delivers.

The important distinction is not optimization versus no optimization. It is who gets the first vote.

Algorithms measure proxies, not quality itself

A recommendation algorithm cannot sit down after reading your article and say, “That changed how I understand the problem.”

It has to work with measurable signals.

On TikTok, for example, recommendations use signals including viewing behavior, likes, shares, comments, searches, and other interactions. TikTok has also described watch time as one signal of viewer interest. (TikTok Newsroom)

Those signals tell a platform something about behavior. They do not directly measure whether a piece of content was accurate, original, useful, or worth remembering.

That distinction changes how writers behave when they optimize for the metric itself. A surprising claim attracts clicks. An angry claim attracts comments. Withholding an answer gives people a reason to keep watching. Simplifying a difficult subject makes it easier to consume.

None of those techniques is inherently bad. The problem begins when the measurable response becomes the goal rather than evidence about how the audience received the work.

Then the proxy changes the work.

A headline gets more dramatic. An argument becomes more certain than the evidence warrants. The introduction delays the answer to preserve suspense. A complicated subject becomes “five things you need to know.”

Each decision is small. Together, they produce something different from what the writer might have created by starting with a simpler question: What would make this worth the reader’s time?

Chasing successful formats creates sameness

Imagine a hypothetical creator who notices that posts beginning “Nobody tells you this about…” consistently outperform their other openings. They use the format again. Other creators notice similar patterns and imitate them.

Nobody issued a rule requiring that opening. The creators inferred a rule from performance.

That produces a feedback loop:

writers imitate successful content → algorithms receive more of that content → audiences respond to it → writers receive more evidence that the format works → more writers imitate it.

This example is hypothetical, but the broader problem of recommendation systems becoming repetitive is not. TikTok itself has acknowledged that recommendation systems risk creating increasingly homogeneous streams and has described efforts to introduce variety and interrupt repetitive recommendation patterns. (TikTok Newsroom)

Writers face a related pressure. When large numbers of creators study the same visible signals and imitate what appears to perform well, their choices converge.

The irony is that optimizing for visibility can erase the distinctiveness that made someone worth discovering.

Search optimization is not the enemy

This argument becomes silly if taken too far.

If you write an excellent article using terminology nobody searches for, give it an opaque title, and make the page difficult for search engines to understand, you have not performed some noble act of artistic independence. You have made useful information harder to find.

Google explicitly recommends using language people might search for in titles, headings, links, and other descriptive parts of a page. Its SEO guidance also says optimization is useful when applied to people-first content. (Google Search Central)

That suggests a useful division of labor:

Write for the person. Optimize for discovery.

Answer the question well first. Then make the title descriptive. Use recognizable language. Organize the article clearly. Add useful internal links. Make sure search engines can crawl it.

This distinction matters even more with AI-generated material. Google’s guidance for generative AI search emphasizes unique, useful, non-commodity content rather than producing variations for every conceivable search query. Its spam policies also identify producing large amounts of low-value material primarily to manipulate rankings as scaled content abuse, regardless of whether AI created it. (Google Search Central)

So “write for the algorithm” is not even particularly good advice for search engines.

Ask what survives without the algorithm

There is a simple editorial test worth using:

If the algorithm disappeared tomorrow, would this still be worth publishing?

Not: Would it receive the same traffic?

Would it still answer a real question? Would you still stand behind the claim? Would a reader still learn something? Does it contain an observation, explanation, example, experience, or argument that justified creating another piece of content?

If yes, optimize it.

If no, optimization may be the only reason the content exists.

Algorithms are intermediaries. Readers are the destination. Treat the intermediary as the audience and you risk becoming extremely good at attracting attention to work that gives people increasingly little reason to pay attention.

📚Bookmarked for You

The Sirens’ Call by Chris Hayes: An examination of attention as a scarce resource and the systems competing to capture it, useful for understanding the incentives behind algorithm-driven media.

Filterworld by Kyle Chayka: Explores how algorithmic recommendation shapes culture and taste, including the pressure toward predictability and sameness.

The Chaos Machine by Max Fisher: Investigates how social platforms and their engagement-driven systems influence what gets amplified and how people respond to it.

🧬 QuestionStrings to Practice: From Reach to Worth

Use this sequence when performance metrics start determining what you create.

“What does my audience actually need from this?” → “Which parts of that need can the algorithm measure?” → “What valuable parts can’t it measure?” → “What would I change if usefulness, rather than reach, were the primary goal?” → “How can I optimize that version without weakening it?”

The goal is not to ignore distribution. It is to decide what deserves distributing before optimizing how far it travels.

QuestionClass publishes a new question to think about each day at QuestionClass.