--- title: "The Next Buyer Is an Agent. Most Stores Are Invisible to It." description: "A growing share of shopping now runs through an assistant that reads data, not pages. A store built only for human eyes doesn't lose that sale. It was never considered." canonical: "https://www.pollyester.com/blog/the-next-buyer-is-an-agent" date: "2026-05-19" author: "Pouya Nafisi" --- [Pollyester](https://www.pollyester.com/index.md) › [Ideas](https://www.pollyester.com/blog.md) › The Next Buyer Is an Agent. Most Stores Are Invisible to It. # The Next Buyer Is an Agent. Most Stores Are Invisible to It. ![Abstract glyph and silk hero. a wide current of turquoise and silver silk filaments flows across a near-black field, sliding past dark sealed matte cubes it cannot enter, and threads cleanly through the notch of a single forged matte lightning bolt, the one form built to be read](https://www.pollyester.com/blog/the-next-buyer-is-an-agent/banner.dark.webp) *A growing share of shopping now runs through an assistant that reads data, not pages. A store built only for human eyes doesn't lose that sale. It was never considered.* A new kind of buyer showed up in the traffic logs this year. It doesn't scroll, it doesn't watch the hero video, and it never reads the homepage. It's an assistant a customer asked to find the best option, and it shops by reading data: your products, your prices, your stock, your reviews. Then it hands its person a shortlist of two or three. If your store can't be read that way, you aren't losing those comparisons. You were never in them. That's the shift worth taking seriously, and it changes what a store even is. ## The shift is real. So is the noise. The numbers stopped being theoretical this year. AI-referred visits to US retail sites grew 805% year over year on Black Friday 2025 (Adobe Analytics, via MetaRouter). What those visitors do matters more than how many there are. A year earlier they converted worse than everyone else. By early 2026 they converted 42% better, stayed 48% longer, and spent 37% more per visit (Adobe). And ask shoppers where this goes: 40% expect to use AI to compare products by 2030, and a third say they'd hand the purchase decision over entirely (FoodNavigator). Now the other half, because both halves are true. Measured against all of ecommerce, this channel is still a rounding error. Anyone selling you panic is selling. It's a small channel on a steep curve, which means the brands that get readable early collect the highest-intent traffic in commerce while it's still cheap to win. > The channel is small. The direction isn't. ## What the agent actually reads When an assistant picks between two brands, it doesn't weigh the brand story. It weighs availability, price, quality signals, whether you're the primary seller, and whether checkout is wired up (OpenAI). All of that is data, and most stores can't hand it over. Adobe looked at retail sites during the same traffic surge and found most of them still aren't readable by a machine at all (Adobe). The cost of that gap is already familiar from the human side. 42% of shoppers abandon a purchase over missing product information, and poor data quality costs the average business around $15M a year (Mirakl, via MetaRouter). An agent is that same behavior with the patience removed. A vague product page makes a human squint. It makes an agent move on. There's no partial credit. You're on the shortlist or you're not. ## The default answer was built for a different buyer The reflex when a new channel shows up is to add something: a bigger platform license, another retainer, a new dashboard to watch it all. That stack was designed for a buyer who browses. It renders pages a human loves and an agent can't see, and the run-rate climbs either way. MIT found 95% of enterprise AI pilots delivered no measurable P&L impact (MIT, via Fortune), and the pattern underneath was consistent: the money went to visible add-ons while the product data and plumbing stayed a mess. The other trap is betting on one assistant. OpenAI shipped Instant Checkout inside ChatGPT in September 2025 and was scaling it back by March 2026, after only about a dozen of Shopify's millions of merchants had gone live (Digital Commerce 360). The buy button moved, then moved again. Any brand that rebuilt around a single assistant's checkout spent months building on sand. So the move isn't more platform, and it isn't picking the winning assistant. It's making the store itself the thing any of them can read. > A vague product page makes a human squint. It makes an agent move on. ## What an agent-ready core actually is [Building the agent-ready core](https://www.pollyester.com/services/build-the-agent-ready-core.md) is a smaller build than the word "replatform" suggests. Four properties, in operator terms. **One product record that stays true.** Title, price, stock, and the attributes a buyer filters on, complete and current to the minute, in one place. The model rewards the merchant whose "in stock" means in stock, and it remembers the one whose doesn't. **Built to answer machines, not just browsers.** The core can hand its data to any software that asks, through the standards agents already use. MCP is the current one for reading a catalog, and Shopify now stands one up for every store on the platform (Shopify), with a cluster of checkout protocols forming behind it — though the protocol matters far less than [what an agent-ready catalog actually needs](https://www.pollyester.com/blog/mcp-acp-ucp-agent-ready-catalog.md) flowing through it. Wire the core to the standard once, and when the next protocol ships, connecting it is a settings change instead of a rebuild. **Intelligence that runs on your own data.** The answers an agent gets about you should come from your reviews, your stock, your order history, not from a vendor's black box that learns from your customers and keeps the learning. That record is also the one input that compounds — [the data is the moat, the model is a rental](https://www.pollyester.com/blog/data-is-the-moat-model-is-a-rental.md). Models keep getting cheaper; what they read doesn't. **A run-rate a P&L can carry.** For a $12M brand, our own build-vs-rent model puts running this kind of stack at roughly a few hundred dollars a month. The rented equivalent, platform fees plus the retainers to operate it, runs into the tens of thousands a month. And the intelligence inside the core keeps getting cheaper on its own; the blended price of the models it runs on keeps falling fast year over year. The core is the rare piece of infrastructure whose bill trends down. ## Get readable before the traffic arrives None of this work is stranded if agents take longer to arrive than the curve suggests. The same clean, machine-readable record lifts the channels you already run, Google, Amazon, and paid, because every one of them is a machine reading your data too. The move is to [fix the data a machine reads](https://www.pollyester.com/blog/stop-doing-geo-fix-the-data-underneath.md) once and let every channel draw on it. The agent channel is the newest reader, not the first. We've watched channels change shape before, and the pattern repeats: the traffic shows up first at the stores that were ready for it. The buyer is becoming an agent. It reads before it buys, and it can only buy what it can read. Build the store that answers. Newer: [Where AI Learns About Your Category: Reddit, YouTube, and the Testers](https://www.pollyester.com/blog/where-ai-learns-about-your-category.md) Older: [Build a Review Flow That Captures the Words the Model Reads](https://www.pollyester.com/blog/review-flow-words-the-model-reads.md) --- ## Pollyester — site navigation Primary: [Services](https://www.pollyester.com/services.md) · [Work](https://www.pollyester.com/work.md) · [Partners](https://www.pollyester.com/partners.md) · [About](https://www.pollyester.com/about.md) · [Ideas](https://www.pollyester.com/blog.md) Company: [About](https://www.pollyester.com/about.md) · [Careers](https://www.pollyester.com/careers.md) · [Contact](https://www.pollyester.com/contact.md) · [How a project runs](https://www.pollyester.com/process.md) · [Engagement](https://www.pollyester.com/engagement.md) Services: [Get found](https://www.pollyester.com/services/get-found.md) · [Get chosen](https://www.pollyester.com/services/get-chosen.md) · [Grow the customer](https://www.pollyester.com/services/grow-the-customer.md) · [Earn advocacy](https://www.pollyester.com/services/earn-advocacy.md) · [Pick, pack & ship](https://www.pollyester.com/services/pick-pack-ship.md) · [Order orchestration](https://www.pollyester.com/services/order-orchestration.md) · [Cost optimization](https://www.pollyester.com/services/cost-optimization.md) · [Build the agent-ready core](https://www.pollyester.com/services/build-the-agent-ready-core.md) · [Brand & identity](https://www.pollyester.com/services/brand-identity.md) · [Creative & art direction](https://www.pollyester.com/services/creative-art-direction.md) · [Content & campaigns](https://www.pollyester.com/services/content-campaigns.md) · [Video & media](https://www.pollyester.com/services/video-media.md) · [Storefront & product](https://www.pollyester.com/services/storefront-product.md) · [Experiential & activations](https://www.pollyester.com/services/experiential-activations.md) · [Events & retail](https://www.pollyester.com/services/events-retail.md) · [Design & build](https://www.pollyester.com/services/design-build.md) Legal: [Privacy](https://www.pollyester.com/privacy.md) · [Terms](https://www.pollyester.com/terms.md) Full index: [llms.txt](https://www.pollyester.com/llms.txt) — the markdown map of the site.