How to write headlines that get clicked: what 234 Paul Graham essays show

In this article I will explain what makes a reader click a headline. I will use the titles of all 234 essays on Paul Graham's site, the scores those essays received on Hacker News, a published analysis of thousands of headline experiments from a news site, and a mistake I made with the titles on this site.

To begin with, the problem.

Most people write the article first and the headline last, usually in the five minutes before publishing. That order is backwards, because the headline is the only part most people will ever read. David Ogilvy claimed that five times as many people read the headline as read the body copy. On a search results page the ratio is harsher. A reader sees ten titles, picks one, and never learns what was in the other nine.

Thus, the headline is where the click is won or lost, and the usual advice for winning it is to make it catchy: add a number, a strong adjective, a gap the reader has to click to close.

Paul Graham has ignored that advice for over twenty years, and his bio says his site "now gets around 25 million page views per year". The site does not even put its titles in text. When I fetched "How to Get Startup Ideas" on 8 October 2026, the title on the page was a GIF image, 206 by 18 pixels, with the words in its alt text. There is no meta description, no Open Graph tag and no H1, and the robots.txt points to a sitemap that returns a 404.

His titles do something most catchy headlines do not, and the data below shows what it is.

What makes a headline get clicked?

A headline gets clicked when the reader recognises his or her own question in it and believes the page behind it will answer that question. In simple words, a click is a match plus a promise. The match is the words: the reader has a phrase in mind, and the headline either contains it or does not. The promise is the reason to pick this result over the other nine.

A catchy headline is built for a different reader. It is designed to stop someone who was not looking for anything, which is the job of a newspaper front page or a social feed. A search headline is read by someone who was already looking. Thus, the two are tested against different people, and advice that works on one often fails on the other.

It is important to note that a click is also a small contract. If the page does not keep the promise, the reader goes back to the results, and the next headline from the same writer starts with less trust.

What 234 Paul Graham essay titles show

I took every title from the essay index on paulgraham.com, 234 of them on 8 October 2026, and measured them. Then I pulled every Hacker News submission of a paulgraham.com essay from the Hacker News search API, 1,570 stories in total, and kept the highest-scoring submission for each essay. 231 of the 234 essays had been submitted at least once.

The titles first. The median title is four words and 21 characters long. The longest is "How to Convert Between Wealth and Income Tax", at 52 characters. Twenty-three titles start with "How to", thirteen with "Why" and eighteen with "What". Only eight end in a question mark, which surprised me, because "add a question" is one of the most repeated pieces of headline advice. The usual source for it is Backlinko's 2019 click-through study, which reported that "title tags that contain a question have a 14.1% higher CTR". That figure still circulates. The current version of the same study says the opposite: "Titles that do and do not contain a question have similar CTRs." Thus, one of the most repeated pieces of headline advice has been contradicted by its own source, and the old number is still the one I see quoted most.

Then the scores. The "How to" titles are 10% of his essays, but they are 6 of his 20 highest-scoring essays on Hacker News, that is 30%. Their median best score is 356 points. The median for the other 208 essays that were submitted is 209 points.

These are the ten highest scores:

1. Having Kids, 1,992 points
2. Jessica Livingston, 1,740
3. The Lesson to Unlearn, 1,333
4. Founder Mode, 1,330
5. Do Things that Don't Scale, 1,275
6. How to Work Hard, 1,223
7. The Bus Ticket Theory of Genius, 1,164
8. How to Think for Yourself, 1,090
9. Frighteningly Ambitious Startup Ideas, 1,032
10. How to Do Great Work, 1,008

The other three "How to" titles in the top 20 are How to Get Startup Ideas (914), How to Write Usefully (821) and How to Earn a Billion Dollars (739).

The second thing his titles show only appears when you type them into Google. On the same day, Google's autocomplete finished "how to get startup ideas" with "how to get startup ideas paul graham", "how to get startup ideas paul graham pdf" and "how to get startup ideas pg". It finished "do things that don't scale" with "do things that don't scale paul graham", and "write simply" with "write simply paul graham". Typing "paul graham" on its own suggests "paul graham how to get startup ideas".

In other words, his titles became search queries, and then the queries picked up his name. I cannot think of a better result for a headline, because people have stopped searching for the topic and started searching for the page.

Why do plain headlines beat clever ones?

Four things are going on, and the first is the one most writers skip.

The reader types plain words

Go ahead and type "how to write headlines" into Google. On 8 October 2026 the suggestions were "how to write headlines for news", "how to write headlines in linkedin", "how to write headlines for google ads" and "how to write headlines copywriting", among others. Every one of them is plain, and none of them looks like the clever headline a copywriting course would teach you to write.

A headline that uses the reader's phrasing matches the query. A headline that paraphrases it, however well, makes the reader do a translation step before deciding, and most readers will not bother when nine other results are on screen. Paul Graham's "How to Get Startup Ideas" is, word for word, what a founder would type, and autocomplete shows that founders do type it.

Google rewrites the titles it does not believe

Your title is not guaranteed to appear at all. Google's documentation on title links says they are generated automatically from several sources, starting with the title element and the main heading on the page, and it asks you to "write descriptive and concise text" for the title. After Google changed this system in 2021, it said publicly that "title elements are now used around 87% of the time, rather than around 80% before."

The independent picture is harsher. In early 2022, Cyrus Shepard at Zyppy examined 80,959 title tags across 2,370 sites and found that Google had rewritten 61.6% of them, although he notes that many of the rewrites were minor. Length mattered a great deal. Titles over 70 characters were rewritten 99.9% of the time, and any title over 60 characters had "a greater than 76% chance of getting rewritten". The lowest rewrite rates, 39% to 42%, were for titles between 51 and 60 characters.

Clearly, a title that does not describe the page gets replaced with one that does, and the replacement is written by a system that has never met your reader. A plain, accurate title of a sensible length is the one most likely to survive.

Negative and curious headlines win the click and cost the reader

This is the part of the evidence that cuts the other way, and it should be stated plainly. Upworthy, a viral news site, A/B tested its headlines on live readers for years, and the Upworthy Research Archive records 32,487 of those experiments run between January 2013 and April 2015. Robertson and colleagues analysed 12,448 of the tests, covering 53,699 headlines, in a 2023 paper in Nature Human Behaviour. For a headline of average length, "each additional negative word increased the click-through rate by 2.3%", and each positive word lowered it by 1.0%. It is important to note that these are relative changes, so a 2.0% click-through rate becomes roughly 2.05%, not 4.3%.

Thus, negativity works, if the only number you care about is the next click. Hacker News, the forum Paul Graham started, goes the other way on purpose. Its guidelines ask submitters to "use the original title, unless it is misleading or linkbait; don't editorialize." They go further than most editors would. A gratuitous number is to be cropped, so "10 Ways To Do X" becomes "Ways To Do X", and submitters are asked not to make titles stand out with capitals, exclamation marks or "saying how great an article is". Google's own guidance on helpful content asks a similar question of every page: "Does the main heading or page title avoid exaggerating or being shocking in nature?"

The two positions are consistent once you separate the reader you have from the reader you want back. Upworthy measured a single click inside a feed. A search reader who clicks a misleading result and bounces has learnt something about the site, and I would expect that reader to skip the domain the next time it appears in the results.

A promise the page keeps compounds

Paul Graham's titles have no tricks in them because the essays do not need tricks. "How to Do Great Work" promises a large thing and then spends more than 10,000 words delivering it. Moreover, a reader who got what the title promised is the reader who shares it, submits it to Hacker News, and searches for it by name two years later. That is the "paul graham pdf" suffix in autocomplete. It is demand the title created by being right the first time.

The idea in one sentence

A headline earns the click when it uses the words the reader would type and makes a promise the page keeps, and Paul Graham can do that in four words only because his name supplies the rest of the promise.

Everything practical follows from the second half of that sentence. Most of us do not have the name, so the specific promise has to go into the title itself.

How I write a headline now

I will use this site as the worked example, including the part I got wrong.

In August 2026 I audited the rendered pages of this site rather than trusting the code, and found that every title was doubled. A metadata helper appended my name, and the root layout template appended it again, so the work page went out to Google as "Work - Radwan Altaf - Radwan Altaf". The title of the AI-native engineering article reached 97 characters. Nothing failed a build. The pages looked fine in the browser, because a browser tab does not show you much of a title. It was only visible in the HTML. Google's title link documentation asks you to "avoid repeated or boilerplate text" and to "brand your titles concisely", and I was doing the opposite on every page. Thus, the first step below exists because of that mistake.

1. Look at the title the way the search engine sees it.

View the source, or run curl against the page and read the title element. Do not trust the CMS field or the browser tab. The bug on this site came in with a rebuild, and the fix was small once I could see the output.

2. Write the headline before the article.

Writing it first forces you to decide the promise before you have 3,000 words you are attached to. If you cannot write the headline, you do not yet know what the article is for.

3. Find the reader's words with autocomplete.

Type the start of the question into Google and write down every suggestion. When I typed "ai transformation fails", Google offered exactly one completion: "why ai transformation fails". One of the older articles on this site is called Why AI transformation fails after the demo. Typing "how to become an ai native" returns "how to become an ai native engineer" as its first suggestion, which sits inside the title of How to become an AI-native software engineer without getting slower. The longer method for this step is in How to get content ideas people are already searching for.

4. Put those words first, and keep the whole title short.

My pre-publish checklist caps the title at 60 characters with the site name included. Google does not publish a hard limit, so this is a working rule rather than a law, but the Zyppy data above backs it: 51 to 60 characters was the length Google rewrote least often. The question form is worth using when it is the reader's own phrasing, and not as a trick.

5. Add the specific promise.

This is where you part company with Paul Graham. His four words work because a reader already trusts the name. Without that, add the specific thing the page delivers: a number, a named audience, or the outcome the reader wants. Ogilvy called this flagging your audience. In the AI-native article, "without getting slower" is the promise, and it points at the study the article is built on.

6. Write the meta description as the first line of the answer.

The description is the second half of the headline. Do not summarise the article in it. Answer the question in one sentence and stop.

7. Run the Hacker News test.

Ask whether the title would survive a moderator applying the HN guideline. If it would need to be rewritten as "misleading or linkbait", the article is probably not delivering what the title says, and the fix belongs in the article.

8. Read Search Console, and rewrite the worst pages first.

Google Search Console's Performance report gives impressions, clicks, CTR and average position for every query and page. Sort by impressions and look for pages that sit near the top of the results with a click-through rate well below their neighbours. Those pages already won the ranking and are losing at the headline. Rewrite those first, then compare the same length of time before and after.

What to work on, in order

1. The reader's words in the first half of the title. This is the highest leverage change and costs nothing.
2. One specific promise. A number or a named audience, chosen because it is true and checkable.
3. The description as an answer. One sentence that would satisfy the reader if he or she never clicked.
4. Pages that rank and are not clicked. Fix the headline where the ranking is already done.
5. Your name. This is the slow one. Paul Graham's titles work because two decades of essays kept their promises. That is the only way I know to earn a four-word headline.

Headline habits that cost clicks

Writing the title last. It gets five minutes and carries the whole article.

Puns and wordplay. They make sense after reading the article, which is the one thing a searcher has not done yet.

Hiding the answer to force the click. "You will not believe what happened next" is built for a feed. In search, the reader picks the result that looks like it already knows the answer.

Putting the keyword at the end. If the title is cut off, the reader's own words are the part that disappears.

Changing titles every week. Rankings move on their own, and a week of data on a small site is noise. Change one thing, then wait.

What this does not solve

I would rather be straight about the limits of this evidence than sell it as a law.

The Paul Graham data is one writer. His four-word titles work partly because of who he is, and the same title from a stranger would be a vague promise. They may not even reach the search results as written. In the Zyppy study, titles of 20 characters or fewer had a better than 50% chance of being rewritten by Google, and his median title is 21 characters.

Moreover, Hacker News points are not search clicks. They measure what one community upvoted, which is related to what it clicked but is not the same thing.

The "How to" effect also depends on the period. For submissions made from 2012 onwards, the gap is wide: a median of 737 points for "How to" titles against 291 for the rest. Before 2012 it reverses, at 78 against 147, although only ten "How to" essays fall in that period. Furthermore, "How to" essays may simply be better essays, because a practical question forces a practical answer. I cannot separate the title from the essay with this data, and nobody can without running the same essay under two titles.

The Upworthy study is about news headlines in a social feed between 2013 and 2015. It may not transfer to a search result, where the reader has already stated what he or she wants. Moreover, in 2024 the archive's authors published a correction saying that 22% of the experiments may not have been properly randomised, and a 2025 replication supported the negativity finding but found the effect of positive language did not hold when it was measured by meaning rather than by word lists.

Lastly, I have not shown you a before-and-after from my own site, because I do not have one worth showing yet. Support for Search Console verification was only added to this site on 2 September 2026. When the data is long enough to mean something, I will publish it, including the parts that make this article look wrong.

Frequently asked questions

What makes a headline get clicked? A headline gets clicked when the reader recognises the words of his or her own question in it and believes the page will answer that question. On a search results page that means the reader's own phrasing, placed early, plus a specific promise the page keeps.

Do "how to" headlines get more clicks? In Paul Graham's case, the essays whose titles start with "How to" are 10% of his 234 essays but 6 of his 20 highest-scoring essays on Hacker News. Their median best score is 356 points against 209 for the rest. That is one writer and one site, so treat it as a strong hint rather than a law.

Do negative headlines get more clicks? Often, yes. A 2023 study in Nature Human Behaviour analysed 12,448 of Upworthy's headline A/B tests and found that, for a headline of average length, each additional negative word raised the click-through rate by 2.3% in relative terms, and each positive word lowered it by 1.0%. The study measured single clicks in a news feed, so it says nothing about whether those readers trusted the site afterwards.

How long should a headline be for Google? Google does not publish a limit and says it truncates title links to fit the device. In Zyppy's 2022 study of 80,959 titles, titles of 51 to 60 characters were rewritten least often, any title over 60 characters had a greater than 76% chance of being rewritten, and titles over 70 were rewritten 99.9% of the time. I cap titles at 60 characters including the site name and put the reader's words first.

Does Google change my page title? Sometimes. Google's documentation says title links are generated automatically from several sources, including the title element, headings and anchor text. Its own figure, published in 2021, is that it uses the title element 87% of the time. A descriptive, accurate title is the one most likely to be kept.

Should I use clickbait headlines? Not if you want the reader back. A misleading headline wins one click and costs the next one. Hacker News, the forum Paul Graham started, asks submitters to use the original title unless it is misleading or linkbait.

Before you write the next one

Type the first four words of your next article's question into Google before you write anything else, and use the reader's phrasing as the start of the title. Then write the article that keeps the promise.