The Double Flattening: Why AI Is Erasing Your Brand From Both Sides

There's a question I keep coming back to every time I scroll through a product category on Amazon or Instagram: when did everything start looking exactly the same?

Not bad. Not lazy. Just... identical. The soft diffused lighting. The warm, slightly desaturated colour palette. The lifestyle imagery with that aspirational-but-geographically-nowhere aesthetic. Swap the logo on a skincare brand's Instagram feed with a supplement company's and nobody would notice. Swap that one with a pet food brand's and you'd still be fine. The details change. The visual grammar does not.

It turns out this isn't a vibe. It's a research finding. Last week, I presented at Shopable in The Park, on how consumer behaviour is changing in the era of agentic commerce — and why brand building, the discipline that performance marketing spent a decade trying to kill, is becoming essential again. The timing was deliberate, because a growing body of peer-reviewed research is now confirming what ecommerce operators have been sensing for months but struggling to articulate.

The most rigorous version of this argument comes from a piece published by The State of Brand, drawing on a NeurIPS 2025 Best Paper, a PNAS Nexus study, and a meta-analysis spanning over 130 studies. Their thesis — "The Great Flattening" — was written for a B2B marketing audience. But the implications for ecommerce are considerably worse, because ecommerce brands face a second flattening that the original analysis never considered. They're being compressed from two directions simultaneously — and most haven't realised the pressure is coming from both sides.

The First Flattening: You Sound Like Everyone Else (And So Does Everyone Else)

A paper awarded Best Paper at NeurIPS 2025 — one of the most prestigious conferences in machine learning — tested over 70 large language models using 26,000 real-world queries. When researchers asked different models to write product descriptions for iPhone cases, two systems built on opposite sides of the planet (DeepSeek in China, GPT-4o in San Francisco) produced outputs with 81 per cent measured similarity. Not the same model wearing different hats. Completely independent systems arriving at the same words.

Here's where it gets genuinely fascinating. A separate PNAS Nexus study confirmed that cranking up the randomness dial on these models doesn't produce more creative outputs. It produces incoherent ones. There is, apparently, no setting between "sounds like everyone else" and "sounds like a fever dream." The convergence isn't a workflow bug. It's architectural.

For ecommerce operators, this describes something you can already see happening across every touchpoint. Product descriptions within a category converge on the same sentence structures, the same benefit-first framing, the same rhythmic patterns. Email subject lines follow the same cadences. Ad creative deploys the same hooks. Not because marketers got collectively lazy — because the tools they're using are architecturally disposed toward the statistical centre of everything ever written. Researchers are calling it an "Artificial Hivemind," which sounds dramatic until you open three competitor listings side by side and realise it's just... accurate.

And consumers are noticing. Klaviyo's 2026 AI Consumer Trends Report surveyed 8,000 consumers across eight countries and found that when people spot AI-generated content in brand marketing, they're four times more likely to trust the brand less than more. Half of consumers can now correctly identify AI-generated content, and when they do, 52 per cent disengage. Even among people who actively use and enjoy AI tools, 39 per cent said they'd trust a brand less for using AI-generated content.

The audience isn't being fooled. It's developing antibodies.

The Second Flattening: The Machines Can't See What Makes You Different

So let's say you're one of the disciplined brands. You've kept a genuinely differentiated voice. Original photography. A visual identity that doesn't look like it was batch-rendered at 2am. Congratulations — your human audience might actually notice.

Here's the thing: the new discovery layer doesn't care. This was the central argument I made at Shopable.

AI shopping agents — Amazon's Alexa for Shopping, Perplexity's Comet, Google's information agents, ChatGPT's shopping features — evaluate products on structured data, specifications, price, review sentiment, and availability. They parse. They do not feel. A clever brand tagline is invisible to them. Emotional lifestyle imagery doesn't register. The keyword-stuffed bullet points we all spent years optimising are being replaced by conversational queries that agents answer by pulling from structured product data and review corpora.

Research published in Harvard Business Review confirmed this directly: traditional marketing techniques do not work on AI shopping agents. The signals that have historically driven purchase decisions in human buyers — brand storytelling, aspiration, identity — are not inputs that agents process. An agent evaluating running shoes cares about cushioning type, weight, drop height, and what reviewers said about durability on wet surfaces. It does not care about your agency's campaign or which influencer wore them. (Though the influencer's feelings may be hurt, the agent is unmoved.)

So brands are being compressed from two directions simultaneously. From above: their own AI-generated content sounds like everyone else's. From below: the agent mediating the purchase can't process whatever differentiation survives. The brand is getting squeezed from both ends, and most haven't realised the pressure is coming from two places at once.

The Agentic Spectrum (Or: Why Your Category Determines Your Panic Level)

The double flattening doesn't hit every category equally, which is either reassuring or terrifying depending on what you sell.

On one end sit specification-driven, comparison-heavy categories: electronics, appliances, commodity health and beauty, office supplies. These are the categories where agents excel — where a purchase decision can be reduced to objective parameters, compared across alternatives, and optimised on price. If your product competes on features and price alone, and your AI-generated content is indistinguishable from your competitors', there's no remaining layer of differentiation for either humans or machines to latch onto. That's not a branding problem. That's an existential one.

On the other end sit categories where purchase decisions are shaped by identity, aspiration, and emotional connection: fashion, fragrance, luxury, artisanal food. An AI agent isn't going to mediate the decision to buy a particular perfume or a hand-thrown ceramic bowl — at least not yet. Human emotion is the moat.

But more runway is not immunity. If a fashion brand's social presence looks like every other fashion brand's social presence because the same AI tools produced the imagery and the copy, the emotional distinctiveness that should be its competitive advantage is being undermined at the exact moment it matters most.

Data from Amsive makes the stakes concrete. Branded queries — searches where consumers ask for a specific brand by name — earn roughly 18 per cent higher click-through rates under Google's AI Overviews. Generic queries see a 34 to 46 per cent decrease. When an agent mediates discovery, the brands that survive the compression are the ones consumers already know and search for by name. Everyone else gets sorted into an undifferentiated shortlist assembled on spec alone.

The Compounding Problem

Here's what should concern anyone thinking about this on a 12- to 24-month horizon: the Great Flattening doesn't stabilise. It compounds.

Research on "model collapse" shows that as AI models train on increasingly AI-generated content — which now constitutes a growing share of what's published online — their outputs become more homogeneous with each generation. The models ingest their own outputs. The variance narrows. Over 74 per cent of new web pages already contain detectable AI-generated content, and that proportion is rising.

Apply this to ecommerce specifically. Product listings, A+ content, brand stores, email campaigns, social imagery — all increasingly AI-generated, all feeding back into the training data that powers the next generation of shopping agents. The Alexa for Shopping of 2027 will have been trained on the AI-generated listings of 2025 and 2026. Each cycle tightens the convergence. It's an ouroboros of mediocrity — the AI eating its own tail and somehow producing blander output each time it comes back around.

Building For Both Registers

The strategic implication is that ecommerce brands now need to build identity that works on two registers simultaneously: emotional resonance for human buyers, and machine-parseable structured data for the agents mediating discovery.

Structured data is the new storefront. Complete, accurate, machine-readable specifications aren't a differentiator — they're table stakes. Not "lightweight and durable" but the actual weight in grams, the material composition, the test standards met. Without this, agents can't evaluate your product and you're invisible in the discovery layer.

Answer intent, not keywords. Consumers using AI shopping assistants don't type two-word keyword strings. They ask questions: "best running shoes for flat feet under £120" or "is this pickleball paddle good for beginners." Amazon's Alexa for Shopping data shows shoppers who engage with the assistant are 2.74 times more likely to purchase than those who don't, and conversion rates climb with conversational depth. Content that answers specific, contextual questions performs. Content optimised for keyword density doesn't.

Make your brand the answer, not just an option. The only sustainable defence against agent-mediated commoditisation is brand recognition strong enough that consumers search for you by name, bypassing the agent's shortlist entirely. The Amsive branded query data already shows this conferring a concrete advantage.

And here's where the research delivers its most uncomfortable finding: you can't get there by using AI differently. The tools converge toward the statistical centre of all language and imagery ever produced. That convergence is architectural. The point of view has to exist before the AI touches it.

The Bottom Line

Brand building — the discipline that performance marketing spent a decade trying to make obsolete — is becoming essential again. Not as a nostalgic sentiment. As a structural argument about where competitive advantage now sits.

The brands that navigate the double flattening are the ones doing two things at once: investing in structured data infrastructure that makes them visible to agents, while simultaneously building authentic, recognisable brand identity that makes humans search for them by name. The first is an engineering problem. The second is a creative one. Neither substitutes for the other, and AI, used carelessly, undermines both.

The tools that promised to make your brand more efficient are, left unchecked, making it more invisible. Which is quite the value proposition when you think about it.

P.S. If you're wondering whether your own brand content has been flattened, try this: pull your last five email subject lines and your top competitor's last five. If you can't reliably tell them apart, you have your answer.

P.P.S. The AI that wrote your competitor's product descriptions was trained on the same data as the AI that wrote yours. They are, in a very real sense, the same copywriter. At least human copywriters had the decency to plagiarise each other one at a time.

• Microsoft launches Copilot Cowork worldwide, an agentic system for long-running tasks billed by usage in credits. It runs on Anthropic's Opus 4.8 and Sonnet 4.6 yet claims to undercut Claude Cowork on cost by 30-40%.

• WPP Media partners with Givsly to target programmatic buys using synthetic audiences built on charity donation data, crossing a line agencies avoided. Early beauty and fashion tests show a 2% lift in video completion rates.

• A 20,000-response study finds ChatGPT's product recommendations shift 80.2% when search is enabled versus disabled. Cited sources now drive what gets recommended, so brands win by appearing in the articles ChatGPT already cites.

• TikTok bans AI-generated voices and recordings in shopping livestreams, requiring real-time human or sign language. The rule reverses its own AI avatar push after brands flooded streams with repetitive automated messages.

• A Kapwing report finds 59% of videos served to new TikTok accounts are AI slop, triple YouTube's rate. Kids content is worst at 57%, and despite opt-out tools the default feed still leans heavily AI.

• Bing Webmaster Tools rolls out Citation Share, Intents, Topics, and Compare in its AI Performance dashboard. Citation Share reveals the percentage of AI citations your site captures per grounding query, turning raw counts into relative reach.

🎯 Dart - A tool to manage tasks and workflows using AI assistance.

🧬 Delphi is a cloning tool designed for influencers, creators, experts, and coaches to create an AI version of themselves, allowing for the scaling of expertise and availability.

🌐 Wonderish - Describe what you want to build. Wonderish makes beautiful websites, landing pages, funnels and more.

🖥️ Matrices - AI native spreadsheet that fills itself.

🤖 Fora - Artificial intelligence designed for the C-suite.

😍It was great, all I needed

🙂Good, not great

🙄It sucked