I appear to be building a content farm.

That is not how I described ActuallyRandom.com when I bought the domain. I called it an outlet for the ideas that bounce around my ADHD brain, a place to experiment with AI-generated writing, SEO, advertising and affiliate links. It sounds friendlier when I put it that way.

Still, the basic plan is to publish articles, attract people from search engines and eventually make a little money from their visits. That is close enough to a content farm that pretending otherwise would be silly.

The important question is what kind of farm it becomes.

The old content-farm model treated articles as cheap containers for keywords and advertisements. The 2026 version can be even cheaper because a language model can produce a plausible draft in seconds. It can also be much worse. Give an AI a spreadsheet containing 10,000 questions and it will confidently fill the internet with 10,000 answers, including answers to questions nobody asked and facts that do not exist.

My experiment has a different constraint: AI can help with every article, but I still have to care about the question, check the claims and be willing to put my name on the result. That makes it slower than a proper spam operation. I suspect it also gives it a chance of working.

A very short history of SEO

Search engine optimization has always been a negotiation between publishers trying to be found and search engines trying not to be manipulated.

The 1990s: put the words on the page

Early search engines had limited signals. If a page mentioned a search term in its title, description and body, it had a decent chance of being considered relevant. This encouraged wonderfully direct tactics such as repeating the same phrase until the page became painful to read, stuffing keywords into metadata and hiding extra text by making it the same colour as the background.

The web was smaller, search was more literal and it was easier to mistake repetition for relevance.

Google changed the balance by looking at links as well as words. The 1998 paper describing the original Google system explains how it used the structure of the web to improve results. A link from one page to another could act a little like a citation or vote.

Naturally, an industry then emerged to manufacture those votes.

SEO became more professional during the 2000s. Sites improved their structure, wrote clearer titles and made pages easier for search engines to crawl. Those are still sensible practices.

The less attractive side was an arms race for links and scale. Publishers bought links, swapped them, hid them in templates and built networks of thin sites that linked to one another. Large content operations commissioned enormous numbers of inexpensive articles, each aimed at a slightly different search phrase.

This was the classic content-farm period. A search for how to fix a leaking tap might lead to an article written by somebody who had never held a spanner. The article did not need to solve the problem particularly well. It needed to rank, display ads and be cheap enough that a small number of visits could make it profitable.

The 2010s: Google gets better at recognising the trick

Google’s 2011 quality update, later widely known as Panda, targeted low-value pages, copied material and sites that were not very useful. Google said the initial change affected 11.8 percent of queries and was intended to reward original research, reporting and analysis. Its announcement is still worth reading because it sounds remarkably current.

Penguin followed in 2012 and went after manipulative link practices. Mobile usability became important as phones replaced desktop computers for much of the web. Search also became less dependent on exact keyword matching and better at understanding what a person meant.

By 2019, Google’s BERT language system was examining words in the context of the words around them. Google described it as one of the biggest advances in Search and said it would initially help interpret one in ten English-language searches in the United States. The practical lesson was simple: pages could be written in natural language for natural questions. There was less need to repeat an awkward phrase because a keyword tool said that was exactly what people typed.

The early 2020s: helpful content and first-hand experience

Google’s 2022 helpful content update made the target explicit. It wanted pages that left visitors satisfied and warned against extensive automation, shallow summaries, arbitrary word counts and publishing across many unrelated topics in the hope that something would rank.

The wording has evolved, but the current people-first content guidance still asks whether a page adds original information, demonstrates first-hand knowledge and gives a visitor enough information to achieve their goal. It also asks whether the site has a clear purpose.

That last question is awkward for a site called Actually Random. Randomness is the purpose, but it is not exactly a niche.

2026: search engines answer the question themselves

Search results are no longer just lists of blue links. Featured snippets, knowledge panels, AI Overviews and conversational search can answer a question before the user reaches a publisher’s website.

This changes the economics of informational content. A basic definition or a fact that can be expressed in one sentence may earn an impression without earning a visit. Pages need to offer a reason to click: an experience, a test, an opinion, better evidence, useful detail or simply a voice the reader enjoys.

There is now a small industry selling “generative engine optimization,” usually shortened to GEO. Some of its advice is sensible, but Google says there are no special technical requirements for appearing in its AI features. The same fundamentals apply: a page must be indexable, understandable, useful and connected to the rest of the site.

In other words, we have spent 30 years inventing new names for making a page worth finding.

AI content is not automatically spam

Google does not ban a page because an AI helped write it. Its guidance on AI-generated content says appropriate automation is allowed. The problem is using automation primarily to manipulate rankings.

That distinction matters. AI can be used to:

  • find promising research leads;
  • organise notes and compare sources;
  • turn a rough idea into a workable outline;
  • challenge an assumption;
  • draft a section that a human then checks and rewrites;
  • find gaps, contradictions and clumsy sentences;
  • create metadata and structured data from a finished article.

It can also be used to generate 5,000 pages overnight, each paraphrasing the same three results already on Google. The software is not the deciding factor. The purpose and the result are.

Google’s spam policy calls the second approach scaled content abuse. The definition covers large amounts of unoriginal, low-value content created to manipulate rankings, whether it was produced by AI, human writers or some other form of automation.

So the safe number of AI-assisted articles is not 10, 100 or 10,000. There is no magic publishing rate. Ten useless pages are still useless. A thousand genuinely valuable database pages may be completely reasonable. Scale makes quality control harder, but scale itself is not the offence.

The economics have flipped

For the old content farms, writing was a major cost. If an article earned $3 over its lifetime but cost $2 to commission, the model could work at enormous scale.

AI pushes the first-draft cost close to zero. That sounds like an amazing business opportunity until you notice that everybody else has access to the same machine. The supply of acceptable-sounding prose is now effectively unlimited.

The scarce parts are elsewhere:

  • choosing a question that is actually interesting;
  • knowing which sources deserve trust;
  • noticing when a plausible answer is wrong;
  • having an experience that cannot be copied from a search result;
  • taking a useful photograph or measurement;
  • developing a recognisable point of view;
  • earning enough trust that somebody returns directly.

This is why my article about bats became more interesting after the original fact fell apart. “Bats are bad at echolocation but heal quickly” was a neat claim, but I could not find good evidence for it. The better story was about how an extremely plausible fact travelled from a Russian TikTok account, through my wife, into my notes and almost onto a website.

AI was useful in researching and writing that article. Human context was the article.

My rules for an honest content farm

I am not pretending that Actually Random is an academic journal. It is a personal experiment that may eventually carry ads and affiliate links. These are the rules that stop that experiment turning into sludge.

1. Start with a real question

Keyword research can tell me that people search for something. It cannot make me interested in the answer.

The best subjects so far have come from ordinary life: a strange fact from TikTok, eating Thai food while Adrien Brody appeared on television, or trying to choose a comment system for a Hugo blog. The personal origin gives the article a reason to exist before search traffic enters the picture.

I can still check Google Trends, Search Console and keyword tools later. They help me phrase the question and understand what information people need. They should not be the only source of the idea.

I tried it the other way round once, mining three years of my own ChatGPT history for keywords with real search demand. It half worked, and the half that failed is more interesting.

The idea was that Google’s autocomplete is a free demand oracle. Feed it a phrase, and if it completes that phrase, people are typing it. If it returns nothing, essentially nobody is. That premise held up: two thirds of everything I sent to the endpoint came back empty. What I missed is that my own script refused to accept the empty answer. When a conversation title returned nothing, it quietly chopped the title down to three words and asked again, then filed the result as though the original question had demand. “Scout card game tokens” became “card game tokens”. “4% rule flexibility” became “rule flexibility”. Of the 1,039 hits I was pleased with at the time, only 448 came from the question I actually asked. The rest were answers to a question the script invented after the real one failed.

The corrected picture is duller and more useful. Roughly one conversation in five has a title that Google recognises as a real search. Whether I cared about the subject makes no difference to that rate. Conversations where I argued with the model for forty turns score the same as the ones I abandoned after two. What does predict it is the number of words in the title, which is a fact about ChatGPT’s automatic titler rather than anything about me or about search.

So keyword research is a good editor and a terrible muse. It can tell me which of my questions other people also ask. It cannot tell me which ones are worth answering, and it will cheerfully lie about the first thing if I let it.

2. Add something that was not in the search results

Summarising the first five results is easy for an AI, which means it is easy for everyone. A useful article needs at least one ingredient that cannot be produced by rearranging those pages.

That might be:

  • a personal decision and the reasoning behind it;
  • an original photograph or screenshot;
  • a small experiment;
  • a calculation using public data;
  • a correction to a popular misconception;
  • a comparison based on actually trying the products;
  • an interview or email response;
  • a collection of sources that nobody else has connected clearly.

This does not have to be groundbreaking research. In the Hugo comments article, the useful details were that I had removed Disqus from another blog after it began placing ads above the comments, and that when I installed Cusdis here it turned out to be broken in two ways: its iframe never sizes itself, and its notification emails never arrived, so every comment would have sat unapproved and unseen. I took the comment box off the site again.

None of that is in the search results, because it only happens to you if you actually install the thing. A generic list of twelve commenting platforms cannot tell you that one of them silently swallows your readers’ comments.

3. Research individual claims

Asking an AI to “research bats” is too vague. A better process is to list the claims the article might make and find support for each one.

Primary sources are best when they exist: research papers, official documentation, legislation, company policies and direct interviews. Secondary sources are useful for context and for identifying disagreements. TikTok, Reddit and forum posts can reveal interesting questions, but popularity is not evidence.

Links should lead readers to the evidence, not merely decorate the article. If the source does not support the sentence beside it, the link is worse than useless because it creates fake confidence.

4. Never publish the first draft

AI drafts tend to be smooth, complete and slightly dead. They repeat the premise, arrange everything into tidy groups and use confident transitions to step over missing evidence.

My editing pass checks facts first. Then I remove repetition, generic advice, fake quotations and any paragraph that exists only to make the article longer. Finally, I read it as writing. Does it sound like something I would say? Is there a joke where I would make one? Did the AI turn a simple observation into a corporate strategy document?

Every article gets read before it is published. That seems like an embarrassingly low standard until you look at the volume of unreviewed AI text appearing online.

5. Be clear about who made it and how

Google recommends thinking about the “who, how and why” behind a page. Readers should be able to identify the author, understand the production method when it matters and see that the purpose is to help rather than manipulate them.

Actually Random should therefore have a proper About page, real author information and a plain explanation that AI is part of the workflow. I do not need to begin every paragraph with a warning that a robot was nearby. A site-wide editorial note and article-level disclosure where appropriate should be enough.

The important part is that I remain responsible. “The AI made it up” is an explanation, not an excuse.

6. Give a random site some structure

A narrowly focused site has an obvious advantage. Articles naturally relate to each other, internal links make sense and readers know what they will get.

Actually Random has chosen the opposite strategy. The name gives me permission to write about bats, actors, Hugo and whatever happens next. That is fun, but it can still be organised.

Categories, tags and small topic hubs can connect related posts. A Hugo article should link to the other Hugo experiments. A misinformation article can link to later fact checks. Descriptive internal links help readers and crawlers discover those relationships without pretending the whole site is about one subject.

The editorial focus is not a topic. It is curiosity, verification and showing the process.

7. Get the boring technical details right

Technical SEO cannot rescue an empty article, but technical mistakes can hide a good one.

For a small Hugo site, the checklist is short:

  • one stable URL for each article;
  • a descriptive title and sensible meta description;
  • a canonical URL;
  • a generated XML sitemap;
  • a robots.txt file that does not accidentally block the site;
  • readable HTML with a logical heading structure;
  • useful image alt text;
  • internal links to related content;
  • Article structured data that matches what visitors can see;
  • fast pages without intrusive pop-ups or large layout shifts;
  • Google Search Console and Bing Webmaster Tools configured.

Hugo and GitHub Pages handle much of this nicely. The harder discipline is not changing URLs or deleting articles casually once other pages have linked to them.

Page speed matters, but it is not a substitute for relevance. Google’s current Core Web Vitals measure loading, responsiveness and visual stability. Passing them is good for readers. Shaving another 20 milliseconds from a blank page does not turn it into a useful search result.

8. Measure what people actually do

The first few months will probably involve tiny numbers. That is fine. A new domain has no audience, history or links, and search engines do not owe it traffic.

The useful measurements are:

  • which pages are indexed;
  • which queries generate impressions;
  • whether titles earn clicks when a page appears;
  • which articles attract links or comments;
  • whether visitors read another page;
  • whether affiliate clicks or ad revenue justify the interruption;
  • which corrections or updates improve an article.

Rankings are observations, not a daily score for my self-worth. Search results vary by location, device, language and time. Search Console trends are more useful than repeatedly Googling my own title and becoming offended.

9. Update for a reason

Changing the year in a title does not make an article fresh. Neither does editing the publication date while leaving old prices, broken links and outdated claims untouched.

An update should fix or add something: a new source, a changed product, a better explanation, an actual test or a decision I made after publishing. If two weak pages compete for the same question, combining them is usually better than creating a third.

Old articles are an asset only if they remain trustworthy.

Monetization without ruining the experiment

There is nothing immoral about wanting a website to pay for itself. Hosting, domains, research tools and time all cost money. The trouble begins when the monetization becomes more important than the reader.

Display advertising

Display ads need traffic. On a tiny site, they usually produce tiny revenue while adding scripts, visual clutter and privacy obligations. It makes little sense to make the first 200 visitors fight through a wall of ads in exchange for eleven cents.

Google’s publisher policies do not allow ads on low-value pages, copied pages without added value, or pages with more promotion than publisher content. They also require a privacy policy explaining relevant data collection and cookie use.

My plan is to wait until there is enough traffic to learn something from advertising, then introduce it carefully. Revenue per thousand sessions matters, but so do page speed, return visits and whether I still like looking at my own site.

Affiliate income works best when a page helps somebody make a genuine decision. A first-hand review, comparison or tutorial can earn a commission and still be the best answer to a question. A random Amazon link pasted beside an unrelated fact is unlikely to help anyone.

Paid relationships also need to be obvious. The FTC’s endorsement guidance says the disclosure should be clear, conspicuous and close to the recommendation. Merely writing “affiliate link” may not explain that I receive money from purchases.

Amazon has its own required identification. Its current operating agreement requires the statement: “As an Amazon Associate I earn from qualifying purchases.” Its help page also says individual affiliate links need a clear disclosure placed where readers will notice it.

For search engines, paid and affiliate links should use rel="sponsored". Google prefers that value for advertisements and paid placements, although nofollow remains acceptable.

The simplest rule is to recommend only something I would mention without a commission. If the commission changes the conclusion, the review is an advertisement pretending to be advice.

Things I am not going to do

There are plenty of ways to make the project grow faster on a spreadsheet and worse everywhere else.

I am not going to:

  • publish thousands of untouched AI drafts;
  • copy competitors and swap their words for synonyms;
  • buy an expired authoritative domain and fill it with unrelated affiliate pages;
  • invent personal experience with products I have never used;
  • write medical, legal or financial advice as if I were qualified;
  • buy links or fill comment sections with my URLs;
  • create nearly identical pages for every city, model number or phrasing of a question;
  • use a fake expert byline;
  • hide corrections because an incorrect page is receiving traffic;
  • promise an answer where the honest answer is “nobody knows.”

Several of these tactics may work temporarily. That does not make them a business. It makes them a bet that the search engine, advertising network, affiliate program and readers will all remain fooled for longer than I need them to.

I would rather own a small odd website than operate a large disposable one.

A practical publishing workflow

The workflow I am testing is simple enough to repeat without automating away the parts that matter.

  1. Write down a question when it occurs to me.
  2. Search for the obvious answer and identify what is missing, doubtful or interesting.
  3. Gather primary sources and record the specific claims they support.
  4. Ask AI to help outline the article around the evidence and my personal context.
  5. Generate a draft, then cut anything generic or unsupported.
  6. Check every factual claim, link, name, number and date.
  7. Add useful metadata, internal links and a clear description.
  8. Read the whole article before publishing it.
  9. Submit or inspect it in Search Console, then leave it alone long enough to collect real data.
  10. Update it when new evidence, reader feedback or my own experience justifies an update.

The AI can make steps four and five dramatically faster. It can assist with most of the others. It cannot decide whether I genuinely understand the article or whether it deserves to exist.

So, can you still build a content farm in 2026?

Absolutely. The tools are cheap, publishing is easy and there are more ways than ever to turn attention into money.

Building a successful one is harder.

The old advantage was access to cheap words. That advantage has disappeared because everyone now has access to nearly free words. Search engines are explicitly hunting scaled, unoriginal content, while their own AI systems are absorbing many of the simple questions that once produced reliable clicks.

What remains valuable is inconveniently human: curiosity, taste, experience, accountability and the patience to find out whether a good story is true.

That is the farm I want to build. AI can drive the tractor. I still have to decide what is worth growing, pull out the weeds and make sure nobody gets poisoned.

Sources and further reading