The Sea of Sameness: How AI Is Flattening Every Brand

The sea of sameness used to be a creative director's complaint. Now it is a measurable property of your category, with a mechanism behind it: every brand drafts with the same 3 models, the models pull toward the same statistical average, and the category converges until a classifier cannot tell one company's homepage from another's. The flattening is not a taste problem. It is arithmetic, and it compounds every quarter.

The mechanism, not the mood

A language model is a compression of its training data, and generation pulls toward the dense middle of that compression. Ask it for a tagline and you get the modal tagline. Ask 12 competitors' teams to "polish this copy" and you get 12 versions of the same polish, because the model's idea of polish is singular. The UW, CMU, and Allen Institute "Artificial Hivemind" study (NeurIPS 2025) measured exactly this: models repeat themselves and each other, converging on near-identical phrasings across supposedly independent runs, and turning up the temperature parameter did not bring the diversity back.

The result reads fluent everywhere and belongs nowhere. The homogenisation runs deeper than text: when every strategy deck, positioning doc, and FAQ is drafted from the same prior, the categories start to think alike, not just sound alike. The words are the visible symptom of converged decisions.

How the flattening compounds

The sameness loop has 4 turns, and each one tightens it:

1. Brands draft with the model. Output lands near the machine average by construction.

2. The average content gets published at volume. AI-assisted articles went from roughly 2% of new pages in 2020 to over half by 2025, so the average is now most of what exists.

3. Models retrain on the published average. The middle gets denser. The pull gets stronger.

4. Buyers stop distinguishing, so engines do the choosing. And an engine facing 12 interchangeable brands names the 1 it can resolve, which is rarely the 1 that converged hardest.

The stakes concentrate accordingly. Hexagon measured 3% of brands capturing 71% of AI recommendations¹, and roughly 85% of what AI says about any brand comes from third-party sources² rather than the brand's own site. Converged brands do not just blur into each other on their own pages. They blur in the training data, in the citations, and finally in the answer, where the blur becomes a competitor's name.

The test most brands fail

Strip the logo from your homepage and your top competitor's. Shuffle the paragraphs. Ask a colleague, or a classifier, to sort them back. The logo test is the cheapest distinctiveness instrument that exists, and most of a category fails it, because the model has already averaged their voices into one dialect. We ran a version of this across a full category: our State of Sameness scan of Dubai real estate found a market of brands describing different buildings in identical sentences.

Failing the test has a precise consequence. A model that cannot tell you apart cannot recommend you, for the same reason it gives generic answers to generic prompts: nothing in the input forces it off the average. Your category's sea of sameness is, to the engine, 1 undifferentiated blob with 12 URLs.

Resisting the pull, measurably

The third row is a practice, not a vibe. At Ivanooo, Firoz Azees measures a brand's distance from the machine's own draft of the same brief: 2 vanilla model drafts sit about 0.06 apart, and genuinely distinct writing runs 3 to 5 times further out. That distance is the anti-sameness number. It cannot be prompted into existence, it survives the logo test by construction, and it is the property Distinctiveness Engineering exists to widen. The sea rises either way. The question is whether your brand reads as water. Find out where you sit: get recommended by AI.