AI Made Output Cheap. Judgment Is the Expensive Part.

The cost of producing a competent article fell to roughly zero. Everybody noticed at the same time.

That is the whole problem.

For about eighteen months, being fast was an edge. A team that could ship 100 pages a month beat a team that could ship ten. Then the tooling got good enough that the second team could ship 100 too, on a subscription that costs less than a single freelance invoice. The arbitrage closed.

What is left is the part that was always hard and is now the only thing left to compete on: deciding what should exist, catching what is wrong, and knowing what a customer actually needs.

Volume stopped being a moat the moment it stopped being hard

Here is the test I run on any content plan that lands on my desk.

Could a competent competitor produce this exact asset in an afternoon, using tools they already pay for, without talking to a single customer?

If yes, it is not a moat. It is a cost. You will publish it, they will publish their version two weeks later, and you will both have spent money to arrive at parity.

The uncomfortable version of this: a lot of content programs that looked like growth engines in 2023 were really just arbitrage on production cost. When the cost went to zero, the engine went with it. Nothing was wrong with the strategy. The strategy just had a shelf life nobody priced in.

Three jobs that did not get cheaper

1. Deciding what deserves to exist

The brief is now the product. A model will write whatever you point it at, enthusiastically, at length, whether or not the page should exist.

Nobody is going to tell you that the page you are about to commission will cannibalize two pages you already have, or that the search behind it is served by a free tool rather than an article, or that the query has no commercial intent and never will. That judgment comes from knowing the market and looking at your own data. It does not come out of a prompt.

Most of the value in a content program is now decided before a single word is generated.

2. Catching what is wrong

Verification cost scales with volume. Production cost does not. That gap is where content programs quietly break.

When you shipped ten pages a month, a human read all ten. At 100 pages a month, nobody reads all 100. So errors ship. Not dramatic hallucinations, usually. Small things: a pricing tier that changed last quarter, a feature described the way it worked two releases ago, a statistic that was never real but reads plausibly, an integration you no longer support.

Every one of those is a trust cost with a customer and an accuracy cost with the systems deciding whether to cite you. And the more you publish, the less likely anyone is to catch it.

3. Knowing what the customer actually needs

This one has no shortcut at all.

The reason a support team can write a better FAQ than any model is that they know which question gets asked at 2am by a person who is about to churn. The reason a solutions engineer can write a better comparison page is that they have lost the deal to that competitor and know the real objection, not the one on the feature matrix.

That knowledge lives in sales calls, support tickets, churn interviews, and your own product telemetry. It is proprietary by definition. Nobody else can generate it, because it is not in the training data.

What I would build instead

Fewer pages, each with something in it that could only come from you.

• Original numbers. Your own aggregate data, anonymized. Benchmark figures from your customer base. Pricing you actually charge. Nobody can copy a number they do not have.
• Real artifacts. Screenshots of the actual product doing the actual thing. Configuration files. Query examples. The output of the real tool, not a description of it.
• Named experience. A person who did the work, saying what happened, including what did not work. Generic advice is infinitely reproducible. A specific failure is not.
• Tools instead of articles. A calculator, a checker, a template that does the job beats a page explaining how to do the job. It is also much harder to summarize you out of existence.

None of this is anti-AI, and none of it is a new discipline either. Most of what gets sold as AI SEO is ordinary SEO in a more expensive jacket. I use these tools every day, for drafting, for restructuring, for catching my own inconsistencies, for grinding through the parts of an audit that are genuinely mechanical. They are very good at the middle of the process.

They are not good at the start of it, where you decide what is worth doing, and they are not good at the end of it, where you decide whether what came out is true.

The reframe

The question stopped being "how much can we publish."

It became "what do we know that nobody else can generate." Then: "what is the smallest number of pages that communicates it well." Then: "who is accountable for checking that every one of them is still true six months from now."

Those are strategy and operations questions. They were always the expensive part. It is just that production used to be expensive too, which made it easy to hide behind.

It is not hiding anything anymore.