A design writer at a well-known trade publication recently made the kind of observation that only lands once someone says it out loud: AI-generated websites are starting to look like each other. Not in the loose sense of a shared aesthetic era, but in the tight sense of shared tells. Slightly off color palettes. Oversized serifs. A polished-but-generic feel. If the estimate that roughly a third of new web pages are now AI-assisted is even directionally right, the internet is quietly converging on a house style that no one deliberately chose.

So here is the question I keep circling: if AI floods the web with near-identical output, what does the market start paying a premium for?

My read is that it pays for managed variance. Not more content. Not better content in the abstract. Difference that has been deliberately selected, anchored, and preserved under speed pressure. That is a market signal, and it points at a product category most teams are not building yet.

The convergence is not just aesthetic

The reporting on AI-produced websites is careful about what it claims. Designers say the outputs are recognizable the way AI writing is recognizable, marked by familiar tics. That is a surface observation, and it might correct itself as tools mature and taste diversifies. Worth noting, not worth panicking about.

The deeper claim is the one that matters commercially. As AI-generated pages fill the web, they become training data for the next generation of models. Researchers have already shown that models trained heavily on their own output steadily produce less varied results. That is a bounded finding, not a prophecy of collapse. But it points to a real economic force: the marginal cost of creating a website, a landing page, a launch narrative, or a brand system is falling toward zero, and the marginal likeness of what comes out is climbing.

When the marginal cost of any input collapses, that input stops being where advantage lives. In creation, we are close to that point. The scarce thing is no longer output. It is difference that survives the pipeline.

What Semafor Intelligence quietly demonstrates

The counterexample is easier to see when it is concrete. Semafor just launched Semafor Intelligence, an AI-enabled editorial insight product built on the transcripts and video from its World Economy convening. The mechanics are worth pausing on, because they describe a product primitive, not an editorial gimmick.

According to Semafor's own description of the build, the team used AI to extract about 4,900 distinct claims from more than 300 speakers. Each claim is anchored to the exact quote and moment in the source transcripts. The system then organizes those anchored claims into themes and, crucially, maps where leader consensus is forming, where it is fracturing, and where the prevailing view might be dangerously wrong. One of the surfaced tensions is the widening gap between real-economy CEOs, who describe themselves as battle-tested, and Wall Street financiers, one of whom, Peter Orszag, described the moment as a "Road Runner" one in which the impact of what is happening is not yet manifest.

Read that mechanically. The product is not a summary. Summaries flatten. This system does something structurally different: it preserves disagreement, holds each claim to its source, and treats the space between speakers as the signal.

That is what curation looks like when it becomes infrastructure rather than a copy edit.

Variance management as a product primitive

Call it managed variance. The system takes in a large, messy corpus and outputs not a single narrative but a structured map of where narratives diverge, with provenance intact. It refuses to collapse difference into consensus for the reader's convenience.

Most AI content pipelines do the opposite. They ingest variety and emit uniformity. That is, at heart, what a generic model completion is: the statistical middle of everything it has seen. Useful for many tasks. Corrosive when repeated at civilizational scale on brand, editorial, and design surfaces.

A curation-first product treats variety as an asset that must be actively preserved. In practice that means at least three moving parts:

  • Claim-level extraction with provenance. Not paragraph summaries. Discrete assertions bound to a speaker, a document, a timestamp.
  • Structured disagreement. Themes with named contrasts, not a blended "consensus view." Difference has to show up on the surface, not get averaged away in the middle layer.
  • Distinct output structures. The final artifact should not look like every other AI artifact. If the last step in the pipeline is a generic model writing a generic paragraph, you have leaked all the variance you just spent effort preserving.

None of this is exotic. It is closer to information architecture than to prompt engineering. But it is not what most marketing, editorial, or brand teams are budgeting for when they buy "AI content."

The market opening

If you accept that surface convergence is happening and that training-data feedback loops will nudge future models toward even less varied defaults, then the strategic picture is not complicated.

When creation is cheap and abundant, disciplined selection becomes the moat. The teams that will be visibly different in two years are not the ones with the largest content velocity. They are the ones whose systems refuse to let sameness be the default output. That refusal has to be engineered. It will not happen by taste alone, because taste at the individual level is exactly what gets averaged out when work moves through a generic pipeline.

This is where the market signal sits. Products like Semafor Intelligence are early evidence that a category is forming around curation as infrastructure: claim graphs, provenance-anchored analysis, structured disagreement engines, brand-specific variance layers that sit between the generic model and the published surface. I would expect this to spread from editorial into brand systems, category analysis, competitive intelligence, and eventually into the design tools themselves, where a "do not look like everyone else's AI site" constraint becomes a real setting, not a wish.

The adjacent implication for anyone running an AI content operation is worth sitting with. Volume metrics will keep looking healthy while the outputs quietly converge with the outputs of every competitor using the same tools on the same models. That is a slow, invisible discount on brand equity. It shows up in the data late, if at all. Related archive reading here: on the shift from surface presence to structured evidence in The Citation Footprint, and on the governance version of the same problem in Stop Asking If It's AI. Ask If It's Additive.

What I would actually watch

The thing I would watch is not whether AI content grows. It obviously will. The thing I would watch is which teams start treating variance as a line item.

Who is funding claim-level extraction and provenance for their own knowledge base. Who is designing output structures that a generic model cannot casually reproduce. Who is measuring not "how much did we publish" but "how distinct is our surface from the category median." That last metric barely exists as a discipline today. It will.

My read is simple. The teams that win the next stretch will not be the fastest generators. They will be the ones who prevent sameness from becoming the default business outcome, because they built the system that actively selects what difference gets to survive. In a market where creation costs are collapsing, that selection is the product.

The rest is filler with better fonts.

Sources