For about fifteen years, a particular kind of content was a reliable business. You picked a question your buyers were asking, wrote a clear 1,500 word answer, optimized it, and waited for search traffic to find it. The article was not special. Forty other sites had a version of it. But it was competent, it was yours, and competent plus yours was enough to rank and to be read.
That trade is over. Not slowing down, not under pressure. Over. The reason is simple enough to state in one sentence: a language model now writes that competent 1,500 word answer in seconds, for a fraction of a cent, and so does every competitor's model. When the supply of competent prose becomes effectively infinite, competent prose stops being worth anything. The quality bar did not rise gently. It moved to a place most content operations were not built to reach.
This post is about where it moved to, and what a content team has to become to clear it.
How the floor fell out
The collapse has a date you can point at. In November 2024, according to an analysis by Graphite of 65,000 English-language article URLs sampled from Common Crawl, the number of AI-generated articles published on the web overtook the number written by humans. A year after ChatGPT's launch, in November 2023, AI articles had already reached 39 percent of what was published. The line then crossed 50 percent and has stayed roughly level since. Half the new web, give or take, is now machine output.
The economics behind that line are not subtle. An Ahrefs survey of 879 marketers put the average cost of an AI-generated blog post at about 131 dollars against 611 for a human-written one, a 4.7 times gap. That AI figure already folds in editing and oversight. A raw first draft from a 20 dollar monthly subscription costs almost nothing per article. When you can produce a thousand keyword-targeted pages in a weekend, you will, and many people did.
Here is the part that catches teams off guard. The flood did not just add noise around your content. It changed what your content is worth, because most content was, in plain terms, a summary of things already known. An explainer on how attribution works. A listicle of email tools. A definition piece. That work had value only because producing it took effort and most competitors would not bother. Remove the effort and you remove the value. You did not get worse at your job. The job stopped being scarce.
SEO professionals have a clean term for the content that lost its footing: commodity content. Content that covers a familiar topic in a familiar way, that any competent writer or any competent model could produce from a brief, that is interchangeable with the version next to it. If your page disappeared and a near-identical one filled the gap with no real loss to the reader, it was commodity content. And commodity content is exactly what the supply curve just buried.
What the platforms did about it
Two forces then made the situation acute rather than merely uncomfortable.
The first was Google. Its March 2026 core update made scaled content abuse the headline enforcement target, after a February Discover-only update that already pushed Discover toward in-depth, original content over clickbait. The policy language matters and is widely misread: Google does not prohibit AI-generated content. It targets content that gives users no real value, however it was made. The distinction is the whole story. Hand-written thin content gets the same treatment as machine-written thin content. AI sites just dominate the casualty list because AI made thin content trivial to mass-produce. Sites that had quietly accumulated rankings on volumes of generic AI pages saw 50 to 80 percent of their organic traffic vanish inside roughly two weeks. Affiliate sites were hit hardest of any category.
The same updates pushed harder on the first E in Google's E-E-A-T framework, experience. The quality rater guidelines ask whether the author has genuine first-hand or life experience of the topic. A product review from someone who used the product. A method described by someone who ran it. None of which a model has, because a model has read about the world rather than lived in it.
The second force is AI search itself, and this one is the more interesting reframe. When a buyer asks ChatGPT or Perplexity or Google's AI mode a question, the engine does not return ten links for the reader to judge. It writes one answer and decides which sources to cite. So the question your content has to survive is no longer "is this the best page for the query." It is, as one good description of GEO puts it, "what is the safest thing the engine can repeat without being wrong."
That sentence should reorganize how you think about content. An AI engine is a risk-minimizing machine. It will preferentially cite a verifiable, attributable claim over a vague restatement, because the verifiable claim is safer to repeat. Original data with a documented methodology has no distortion layer between the claim and its source. A generic summary has nothing the engine can point at. The Princeton-led GEO study, presented at the KDD 2024 conference, measured this directly across the modification strategies it tested: adding relevant statistics to a page lifted its visibility in generative engines by about 41 percent, adding citations to credible external sources by a comparable margin, and adding quotations from named sources by about 28 percent. Keyword stuffing did slightly worse than the baseline, and simply adding more words moved nothing.
Put the two forces together. Search engines are demoting commodity content, and AI engines are citing the content that carries verifiable, original claims. They are pushing in the same direction, toward the same kind of work.
What "original research as the new minimum" actually means
The word minimum is doing real work in that phrase, and it is worth being precise about it.
A few years ago, original research was a prestige move. You ran one big survey a year, published a report, earned a wave of backlinks and some press, and went back to the explainer mill. It was the showpiece on top of the ordinary content.
That arrangement is inverted now. The explainer mill produces things of no value, so it is not a floor to build on. The floor is gone. Original research, original data, first-hand testing, expert judgment, a genuine point of view: these are no longer the showpiece. They are the entry condition for content being worth publishing at all. If a piece contains nothing a model could not have generated on its own, the model already did, faster and cheaper, and the platforms will treat your version accordingly.
This is not a moral argument about authenticity. It is a supply argument. The market clears at the marginal cost of production, and the marginal cost of competent prose is now roughly zero. The only inputs that hold value are the ones with a real cost a model cannot pay: the cost of running a study, of testing a product for a month, of interviewing twenty practitioners, of having spent ten years doing the thing you are writing about. Those costs are exactly what makes the output scarce, and scarcity is the only thing the market still pays for.
The data is starting to confirm the shift in spending. Content Marketing Institute research, cited in industry round-ups, finds 86 percent of marketers planning to increase research budgets, with teams that publish original data reporting markedly higher conversion and stronger organic performance than teams that do not. The Content Marketing Institute's 2026 trends piece, drawing on 42 practitioners, lands repeatedly on the same point in different voices. GitHub's content lead Aaron Winston says it is no longer enough to answer a question, that you have to approach every topic with the rigor of a journalist. CMI's Robert Rose puts it as depth over speed. NetLine's David Fortino is blunter: 2026 is the year AI floods every channel with more of the same, fast and cheap and forgettable. The forgettable part is the warning. Forgettable content is now free, and free content does not earn attention.
Rebuilding a content operation around proprietary inputs
If the input is the scarce thing, the content operation has to be rebuilt around securing inputs rather than around producing words. A few concrete shifts follow from that.
Treat research as a pipeline, not an annual event. The teams adapting fastest are not running one survey a year. They are pulling a steady stream of proprietary signal: aggregated and anonymized product usage data, patterns from sales calls and support tickets, results from their own tests, quarterly pulse surveys of their customer base. Each piece of that becomes a claim no competitor and no model can reproduce, because the underlying data exists nowhere else.
Make every brief specify its original input before a word is written. A useful discipline is to refuse to commission a piece until someone can answer one question: what is in this that did not already exist on the internet. A number we generated. A test we ran. A practitioner we interviewed. An argument only someone with our experience would make. If the honest answer is nothing, the piece should not be written, because the model already wrote it.
Shift the writer's job from production to extraction. When drafting is cheap, the expensive and valuable work moves upstream and downstream of it. Upstream: getting the real input out of a subject matter expert who is busy and out of the proprietary data that is messy. Downstream: editorial judgment, fact-checking, the point of view. Surveys repeatedly find expert contribution is the genuine bottleneck in B2B content, not drafting. The team that solves the bottleneck of getting expertise onto the page, rather than the non-problem of generating text, is the team that wins.
Use AI for exactly what it is good at, which is not the part that matters. A model is excellent at structuring an argument, drafting from your raw inputs, adapting one piece into ten formats, and handling the mechanical labor around content. It is not a source of insight, because it can only recombine what already exists. The correct division is straightforward: the model handles the production, the humans own the input and the judgment. Teams that get this backward, that let the model supply the substance and treat humans as editors of machine output, are producing the exact commodity content the platforms are erasing.
For a marketing team, the agentic shift makes this division sharper rather than softer. As content agents take over more of drafting, optimization, routing, and reporting, the human contribution narrows to the part agents structurally cannot do: deciding what to study, running the study, and reading the result with judgment. The work that survives automation is the work that produces proprietary inputs. Building a content operation around that today is the same thing as building one that will still function in two years.
The honest version of the future
It would be easy to end on the comfortable note that human creativity always wins. That is not quite the claim, and the honest version is more demanding.
Most content will be machine-generated, and that is fine, because most content is mechanical and always was. The competent explainer was never the valuable part of content marketing. It was the cheap part that happened to be scarce. AI removed the scarcity and exposed how little the explainer was ever worth.
What is left is harder and smaller and more valuable. A model cannot run your survey. It cannot test the product for a month and report what broke. It cannot interview the customer or sit in the sales call. It cannot have an opinion that is genuinely yours, formed by experience it never had. Those are the inputs that earn a reader's attention and a citation in an AI answer, and they are the inputs no model can manufacture, which is precisely why they are now the minimum rather than the bonus.
The teams that will be fine are not the ones generating the most content. They are the ones that asked, before writing anything, what do we know that nobody else does, and built their whole operation around answering it well.
Council summary
This post argues that competent explainer content has lost its commercial value because language models produce it at near-zero marginal cost, so original research, first-hand testing, and genuine expertise are now the entry condition for publishing rather than a prestige add-on. The council verified the load-bearing figures: Graphite's analysis of 65,000 Common Crawl URLs dating the AI crossover to November 2024, the Ahrefs survey of 879 marketers showing a 4.7 times cost gap, the Princeton GEO study's lift from statistics and citations, Google's March 2026 scaled content abuse enforcement, and the 86 percent research-budget figure. We corrected the Graphite sample size, reattributed the Ahrefs cost data to its survey, scoped the February 2026 update correctly as Discover-only, and softened an unsourced production-speed claim. The reader takeaway is concrete: before commissioning any piece, demand its proprietary input, the number, test, interview, or argument that did not already exist online, and rebuild the operation around securing those inputs rather than generating words.
Comments