How to get content ideas people are already searching for

In this article I will explain how to find out what people want to read before you write it. The method is borrowed from Paul Graham's essay on startup ideas, and I will apply it to a short technical article I published in 2022 and to two articles on this site.

To begin with, the problem.

Most pages on the internet are never found through search. Ahrefs looked at the roughly 14 billion pages in its index and found that 96.55% of them get no traffic from Google at all. When Ahrefs first ran the same study in 2018, the figure was 90.88%, so the share has grown since. The first reason the study gives is the obvious one: "The topic has no search demand." Somebody wrote an answer to a question that nobody was asking, and as the study puts it, "If nobody is searching for your topic, you won't get any search traffic", however well the page ranks.

Paul Graham describes the same mistake in a different trade. In How to Get Startup Ideas he writes that "by far the most common mistake startups make is to solve problems no one has. I made it myself." In 1995 he started a company to put art galleries online, and spent six months on it before noticing that galleries did not want to be online. His explanation is the useful part: "I invented a model of the world that didn't correspond to reality, and worked from that."

That sentence describes most content calendars I have seen. A team sits in a room, invents a model of what its readers want, and writes for that model. Thus, an article that nobody searched for is the content version of a startup nobody wanted, and it fails for the same reason.

What does reverse-engineering demand mean?

Reverse-engineering demand means starting from evidence that people already want something, such as the words they type into a search box or the questions they ask you, and working backwards to what you should write or build.

In simple words, you find the question first and write the answer second.

The usual process runs the other way. You start with what you want to say, which is usually a topic, and then go looking for a keyword that roughly fits it. That is brainstorming with a search tool attached. Reverse-engineering starts from the reader's words and lets them decide the topic, the title and often the structure.

It is important to note that this is not keyword stuffing, and it is not writing for robots. The evidence of demand is a record of real people asking real questions, and writing the answer they asked for is writing for people. Google's own guidance asks site owners whether they have "an existing or intended audience" who would find the content useful if they came to it directly.

Why made-up content ideas fail

Paul Graham has a name for startup ideas that come from thinking rather than noticing. At Y Combinator they call them "made-up" or "sitcom" ideas, because they sound like something a TV writer would invent for a character: plausible, and wrong. He says the danger is that this method "yields bad ideas that sound plausible enough to fool you into working on them." The content version has three causes.

Topics are not questions

People rarely search for a topic. They search for the problem in front of them, in the words they would use to describe it to a colleague. When I typed "multiple java versions mac" into Google on 8 October 2026, it suggested ten completions, including "install multiple java versions mac", "manage multiple java versions mac" and "switch between multiple java versions mac". Those are different jobs, because installing happens once and switching happens every day, and a reader who needs to switch does not want three screens about installing.

Thus, an article called "Java on macOS" touches all of those jobs and finishes none of them, while an article about switching versions can finish one.

You cannot see demand from inside your own head

The art gallery story is the clearest example I know. Paul Graham was not careless. He was attached to a model of the world, and he had already spent time on the software, so the model was expensive to give up. Writers do the same thing with an outline. Once 2,000 words exist, the question of whether anyone wanted them gets quietly dropped.

Plausible ideas fool you

"The future of AI in logistics" sounds like something people want to read. Plenty of people might find it interesting. Interest and demand are different things, however, and an idea that sounds reasonable in a planning meeting gives you no evidence either way. The only test that works is to look for the reader's own words, and a made up topic usually has none.

Where does demand leave evidence?

Demand leaves a trail. These are the places I look, in the order I trust them.

The questions people ask you. This is where Paul Graham's essays come from. Do Things That Don't Scale opens with: "One of the most common types of advice we give at Y Combinator is to do things that don't scale." In other words, the essay is advice he had already given many times, written down once. If you have sales calls, support tickets, client emails or DMs, you already have a list of demand that no keyword tool can see.

Google autocomplete. Google explains that its predictions reflect real searches that have been made on Google, which is why it is better evidence than a brainstorm. It does not tell you how many people searched, only that people did, and in what words.

Search Console. Once a page exists, Google Search Console shows the queries it was shown for. A query that earns impressions and no clicks is demand you half answered. It is about the cheapest content idea there is, because the reader has already told you what was missing.

Forums. Reddit and Hacker News threads show the question in full, with the context the searcher never types. Moreover, typing "how to find content ideas" into Google suggests "how to find content ideas reddit", so searchers already know where the real questions are.

Someone who did this before it had a name. Patrick McKenzie sold Bingo Card Creator, a small piece of software for teachers, and published his numbers every year. Teachers search for bingo cards on particular topics, so he built a system that turned a word list into a downloadable set of cards and a page to go with it, and paid a freelancer to write new activities. In SEO for Software Companies, in 2010, he wrote that "roughly half my sales and three quarters of my profits come as a result of organic SEO", and that the roughly $3,000 he had paid for that content had generated "well over $20,000 in sales" in the previous year alone. The detail I find most useful is a smaller one. Of about 900 activity pages, he counted 132 that had produced a sale that year. Even pages built from real demand mostly do nothing on their own, and the sales come from a minority of them.

The idea in one sentence

The way to get content ideas is the way Paul Graham says to get startup ideas: look for problems people already have, preferably ones you have had yourself, and check the evidence that they are searching for the answer before you write it.

This is his sentence adapted, on purpose. I think his method transfers almost unchanged, and the only addition search makes is that the evidence is now public and free.

The method, worked on a real article

In 2022 I published a short technical article on how I manage multiple Java JDK versions on a Mac. It came from a problem I had. One client project, built with React Native, needed Java JDK 8 to compile for Android, while another needed JDK 14 or above. When I installed JDK 14 for the second project, the first one started throwing a Java error. In my case, Java 8 was stored under the jdk1.8.0_202 folder, and the fix was making each project use the right one. That is the kind of idea Paul Graham calls organic, because nobody had to invent it.

Here is how I would check that idea today, and how I check new ones.

1. Write down the problem in the words you used when you had it.

Use the sentence you would have typed at 11pm, rather than the name of the topic. In my case it was probably something close to "switch java version mac".

2. Type it into Google and record every suggestion.

The ten completions for "multiple java versions mac" covered install, manage, use, maintain and switch. Go ahead and add a letter after the phrase as well, "multiple java versions mac a", then "b", and so on, because each letter shows a different set.

3. Group the suggestions by job, then pick one job.

Install and switch are the two jobs here. An article that does both is fine, but the title and the first section should belong to one of them. The job you pick becomes the headline. I wrote more about that step in How to write headlines that get clicked.

4. Read the top three results for the job and note what they leave out.

Do not count their words. Look for the axis on which all of them are weak: no real paths, no versions, no explanation of why the wrong JDK gets picked up. That gap is your article's reason to exist.

5. Write the answer with the real details.

Use the real folder names, versions and commands. A reader who arrived from "switch between multiple java versions mac" wants to copy something and get on with the day.

6. After publishing, read the queries in Search Console.

Google Search Console reports impressions, clicks, CTR and average position for each query. The queries you did not expect are the next articles.

7. Do the first twenty by hand.

This is "do things that don't scale" applied to research. The Airbnb founders, Paul Graham writes, went "door to door in New York, recruiting new users and helping existing ones improve their listings." The content equivalent is typing twenty questions into Google yourself and reading the results, before you buy a tool to do it for you. You will learn the reader's vocabulary in an afternoon, and no export gives you that.

The same check works on articles already published. Typing "ai transformation fails" into Google returns a single suggestion, "why ai transformation fails", and one of the articles on this site is Why AI transformation fails after the demo. Typing "how to become an ai native" returns "how to become an ai native engineer" first, which is inside the title of How to become an AI-native software engineer without getting slower. Those titles match the words people type. If they had not, the method above is how I would have found out.

Essays that create demand

There is a second thing in Paul Graham's work that the method above cannot reach, and it is worth understanding before you dismiss anything without search volume.

Some of his essays created a query rather than answering one. Typing "founder mode" into Google today suggests "founder mode meme", "founder mode cap", "founder mode bootcamp" and "founder mode lyrics". He named a thing that had no name, and a thing with no name has no searches until somebody names it. Typing "default alive" suggests "default alive calculator". An essay about a startup's runway produced people searching for a tool to work it out, and in that case the essay itself links to one, a calculator Trevor Blackwell made.

Even the format shows up. "paul graham essays" is completed with "pdf", "epub" and "txt". Readers are asking for something his site does not offer.

This is Paul Graham's other line from the same essay, "live in the future, then build what's missing", turned into a search habit. Autocomplete tells you what people already want. Once in a while it also points at something people want that may not exist yet. Whether a good one already exists is a five minute check, as the default alive calculator shows, and that check is the whole method.

What to do, in order

1. Write down the last ten questions someone asked you. This is the highest leverage step, and most people skip it because it does not feel like research.
2. Run each through autocomplete and keep the reader's exact words.
3. Group by job and pick one job per article. One article that finishes one job beats one that touches five.
4. Read the top three results and write down what all of them leave out.
5. Publish, then read Search Console every few weeks for queries you did not plan for.
6. Only then buy a keyword tool, to choose between ideas you already know are real.

Habits that produce articles nobody finds

Starting in a keyword tool. It gives you volumes for words you already thought of. It cannot give you the question a customer asked you on Tuesday.

Chasing the biggest number. A head term like "startup" is a single word with no question attached. The searcher's intent is unclear and the results are owned by sites with decades of links.

Writing about the tool instead of the problem. People search for the thing that broke, and most of them do not yet know the name of the thing that fixes it.

Trend pieces. "The state of X in 2026" is the purest form of a sitcom idea. It sounds like content and almost nobody types it.

What this does not solve

I would rather be straight about the limits.

Autocomplete shows that a phrase is searched. It does not show how often. Google's help page says predictions also take account of language, location, trending interest and your own past searches, and Google filters some predictions out under its policies. Treat autocomplete as evidence of the words, not of the size of the market.

Demand data looks backwards. It can only show you questions people already have words for. Before Paul Graham published "Founder Mode", that phrase had no demand to find, and this method would have told you not to write the essay. Thus, the method is good at telling you what to write next and bad at telling you what to write that is new. Both kinds of writing are worth doing. Only one of them can be researched in advance.

Demand is also not the same as being able to rank. The Ahrefs study's second reason is that "the page has no backlinks", and its third is that the page does not match what searchers want. A new site answering a real question can still lose to an old site answering it worse, at least for a while.

The 96.55% is an estimate built from Ahrefs' own keyword data, not measured traffic, and Ahrefs says its index is "somewhat biased towards the 'quality side of the web.'" If anything, that suggests the real share across the whole web is higher. I have not found a study that separates careful articles from everything else.

It is also worth being honest about Patrick McKenzie's method. Template pages written by a freelancer worked in 2010. Google's current guidance lists content "mass-produced by or outsourced to a large number of creators" as a warning sign. The lesson that transfers is the matching of pages to real queries. The production method would not survive today unchanged.

Lastly, I cannot yet show you traffic numbers from this site, because it has not been running long enough under Search Console to give numbers worth publishing. When it has, I will share them, including the articles that did not work.

Frequently asked questions

How do I find content ideas that people actually search for? Start from questions you have been asked or problems you have had, type the core of each into Google, and write down the autocomplete suggestions. Group the suggestions by the job the reader is trying to do, then read the top three results for the best group and write the answer they leave out.

What does reverse-engineering demand mean? It means starting from evidence that people already want something, such as the words they type into a search box or the questions they ask you, and working backwards to what you should write or build. You find the question first and write the answer second.

Why do most blog posts get no traffic from Google? An Ahrefs study of roughly 14 billion pages found that 96.55% get no traffic from Google. The three reasons it gives are that the topic has no search demand, the page has no backlinks, and the page does not match what searchers want.

Is Google autocomplete a reliable source of content ideas? It is reliable evidence that people type a phrase, because Google says predictions reflect real searches. It is not evidence of how many people type it. Use autocomplete to find the wording and the jobs, then use Search Console or a keyword tool for volume.

What did Paul Graham say about getting ideas? In How to Get Startup Ideas, he wrote that the way to get startup ideas is not to try to think of them but to look for problems, preferably ones you have yourself. He also wrote that the most common mistake startups make is solving problems no one has, which is the same mistake most content makes.

Do I need a paid keyword tool to find content ideas? Not to start. Your own conversations, Google autocomplete and Search Console are free and closer to the reader's real words. A paid tool helps later, when you need volumes and difficulty scores to choose between ideas you already know are real.

Start with the question you were asked this week

Write down one question a customer, colleague or client asked you this week, type it into Google, and read what comes back. If the suggestions use different words from yours, write the article in theirs.