Research · 6 min read
The agent is the new buyer
Your customers stopped searching. They started asking. And increasingly, the thing doing the asking, comparing, and clicking “buy” is not a person on a results page — it's an AI agent acting on their behalf.
The numbers moved faster than anyone planned for. Adobe measured AI-referred retail traffic up 693% year over year — and that traffic converts 42% better than non-AI traffic, because it arrives pre-sold. An agent doesn't browse. It compares in private, resolves the shortlist, and shows its human two options instead of two hundred.
Why this breaks the old playbook
Everything brands optimized for the last twenty years — page speed, SEO, ad creative, conversion-rate tweaks — optimizes an interface the agent never sees. The agent reads structured context: what the product is, what it claims, what it costs, whether it's in stock, how it compares. Product pages written for Google don't carry that information in a form an agent can use.
That creates a blunt new rule: if an AI can't read your brand, you don't exist in the shortlist. Not penalized — absent.
What to do about it
Treat agents as a channel, not a trend. That means a shelf agents can transact against (structured products, live price and stock), a path to purchase that lands in your own checkout, and attribution honest enough to survive scrutiny. Brands that put real supply on the agentic shelf now are accumulating sales signal — and ranking — that later entrants will have to buy their way past.
Nile is the backend that does this for you: connect your store, and your products go live across agentic channels in minutes.
Explainer · 5 min read
ACP vs. UCP: what brands actually need to know
Two protocols now define how agents buy. OpenAI and Stripe's ACP (Agentic Commerce Protocol) standardizes the transaction: how an agent carries a cart, a payment, a confirmation. Google's UCP (Universal Commerce Protocol) — co-built with Shopify, Etsy, Target, Wayfair, and Walmart — standardizes the product: a compact, semantic schema an agent can reason over.
They're often framed as rivals. For a brand, that framing is mostly noise: one governs checkout, the other governs the shelf, and your products will need to speak both — plus MCP for tool access and Agent Skills for distribution into the assistants people already use.
The real differences
Backing: ACP rides ChatGPT's audience and Stripe's payment rails; UCP has the broadest retail coalition ever assembled behind a spec. Shape: ACP is a flow, UCP is a format — a handful of semantic fields replacing the keyword-stuffed feeds of the last era. Consequence: with so few fields, ranking stops being about keyword tricks and starts being about whether your data is genuinely good.
What a brand should do
Nothing, ideally — by hand. Maintaining separate integrations per rail, per platform, forever, is a roadmap tax that compounds with every new channel. The sane architecture is one canonical brand context, compiled into whichever rail each channel speaks. That's the layer Nile operates — which is also why we're neutral on who wins. Whoever it is, your catalog is already there.
Opinion · 5 min read
Why GEO can't fix a 3,000-SKU catalog
Generative engine optimization — getting your content cited by AI answers — is real, and for publishers it matters. For commerce, it answers the wrong question. Getting mentioned is not getting bought.
The scale problem
GEO operates on prose: pages, posts, citations. A working catalog is thousands of SKUs, each with variants, pricing, stock that changes hourly, and category-specific attributes an agent needs to compare against alternatives. No amount of content polish encodes that — and hand-writing agent-grade context for 3,000 products is not an editorial task. It's data engineering.
The shelf problem
When an agent shortlists products, it isn't quoting articles — it's querying supply. It wants structured product data it can trust, a price it can verify, and a purchase it can execute. Content optimization produces none of those. Worse, each agent ranks that supply differently: what surfaces in ChatGPT is not what surfaces in Gemini or Rufus. The only way to know how you rank is to measure each channel — with real agent sessions — and tune what you feed it.
The economics problem
GEO is paid on retainer whether or not it sells anything, and visibility without a transaction path can't be attributed anyway. Flip the model: put actual supply on the shelf, attribute actual sales, and pay only on what sold. Visibility then arrives free — as a consequence of selling. That's not an optimization layer. It's a backend.
