Link header for the pivot. The human-visible badge was irrelevant. Structured catalog summaries cut the listing payload by 93%.The question
Everyone is talking about agentic commerce — ChatGPT handles on the order of 50M shopping queries a day, and AI-referred traffic to retail sites grew roughly 8× year over year in early 2026. But almost nobody has published primary data on the most basic question: when an AI agent lands on your store with a task, what does it actually do?
So we tested it. Blind agents (a small GPT-class model, no system-prompt hints, no URLs beyond the homepage), a concrete task — find product X on this store and verify its price — against real WooCommerce stores in production. Then we varied one thing at a time.
The setup
- Baseline — a production store (~1,000 SKUs) with no agent infrastructure beyond a
robots.txtmention. - Full signals — a store running our open-source plugin exposing a structured catalog API, plus two discovery signals: an HTML comment in
<head>and an HTTPLink rel="kalicart-agent"response header pointing to a machine-readable discovery document. - Signals, badge off — same as 2, with the human-visible "agent-ready" badge disabled.
What happened
| Cell | Runs | Outcome |
|---|---|---|
| Baseline (no signals) | 1 | Failed discovery — 27 fetches, robots.txt seen and ignored, answered from HTML fragments |
| Full signals | 3 | 3/3 — found discovery doc, switched to structured API, verified price & stock |
| Signals, badge off | 3 | 3/3 — identical behavior; badge never mentioned |
Baseline: the agent scraped category HTML like a 2015 crawler, burned its budget on theme markup, and produced an answer from fragments. It saw the robots.txt hint — and ignored it.
With signals: in one run the attribution was surgical — the agent's own reasoning cited the Link header as the reason it pivoted to the structured API. One HTTP response header, the cheapest possible intervention, flipped the outcome.
Badge off: agents never needed it. The badge is for humans. If you're adding "AI-ready" stickers to your storefront for the agents' sake, save the pixels.
The economics nobody mentions
We also measured the cost of a cold visit. Without structure, the cold start on the baseline store weighed about 134 KB of fetched content — and the homepage HTML alone was 90 KB, 67% of the entire budget, before a single product appeared. With discovery in place, the same task starts at roughly 44 KB.
Inside the catalog itself: serving listing results as structured summaries instead of full records dropped one page of results from 40,382 to 4,686 bytes — −93%.
Tokens are money and latency. A store that is expensive to read is a store agents will read badly — or not at all.
Why this matters for WooCommerce specifically
Shopify solved this wholesale: Shopify Catalog auto-enrolled over a million merchants into ChatGPT's shopping surface with zero merchant action. WooCommerce — powering more stores than any other platform — has no native equivalent today. Every Woo merchant is, right now, the baseline cell of this experiment.
That gap is why we built KaliCart Bridge (open source, on WordPress.org): structured catalog API, MCP endpoint, the discovery signals tested above, an OpenAI-spec product feed generator, and telemetry showing which agents already visit your store. Install-and-forget — because the clearest lesson of the experiment is that defaults win.
Honest limits
Small n (three runs per cell plus one baseline), one model family, two stores, one task type. This is a flashlight, not a floodlight. We are instrumenting more stores and will publish 30 days of live agent-traffic telemetry next — including which crawlers (ClaudeBot and OAI-SearchBot are already in our logs) actually read which surfaces.
The one-line takeaway: agents don't reward beauty, they reward legibility. And legibility, today, is one HTTP header away.