AI Agents Are Buying Packaging: 90 Days of First-Party Order Data

Published July 31, 2026 · First-party data from packrift.com · We update this page monthly.

We run Packrift, a packaging supplies store with 13,000+ SKUs shipped from five US warehouses. In late April 2026 an order arrived that no marketing channel could explain: the buyer came from chatgpt.com. We started measuring. This page shares what happened over the following 90 days, because most of what is written about “agentic commerce” is projection, and we would rather publish receipts — even small ones.

Key findings (April 30 – July 30, 2026)

• AI assistants drove 59% of the revenue our analytics could attribute to a source — more than Google search or any other channel. (Against all store revenue including orders analytics could not attribute, the AI share is roughly 45%.)
• 20 of our 37 source-attributed orders came from AI-assistant referrals (19 ChatGPT, 1 Gemini).
• ChatGPT referrals have produced orders in every full month since the first one landed on April 30.
• Our largest single AI-referred order — a four-figure order — came from Google Gemini, not ChatGPT: one order, in July, from 11 sessions.
• Roughly half of ChatGPT-referred sessions add a product to cart, far above our search-traffic rate. The assistant already did the comparison shopping; the human arrives pre-qualified.
• In July we logged visits from more than ten distinct agent frameworks — Copilot, Perplexity, and automation agents built on LangChain, n8n, Zapier and similar — almost none of them buying yet. The crawl wave is ahead of the purchase wave.
• AI did not replace search — both grew. By July, Google organic search led our first-touch orders (7) ahead of ChatGPT (4).

Monthly data: ChatGPT and Gemini referrals

*Sessions and orders per GA4 session-source attribution. Our first AI-referred order landed April 30, 2026 per order-level attribution; in GA4’s monthly view the ChatGPT order stream begins in May. Perplexity and Copilot sent sessions in every recent month but no orders yet.

What AI-referred buying looks like

These buyers do not land on our homepage. They arrive directly on a specific product page for a specific spec — a 12 x 10 x 6 ECT-32 corrugated box, a 2-mil poly bag in one exact size, a freezer paper roll — and either buy it or leave. Session depth is shallow because the comparison already happened inside the assistant. Order values range from small sample packs to the four-figure Gemini order. Repeat orders have started — small numbers, but packaging is a consumable, and an assistant that bought the right box once has every reason to buy it again.

What we can verify about why it happens

We cannot see inside any assistant’s ranking system, so we limit this section to what is public and what our own data supports:

• Product data is the surface AI shops from. ChatGPT’s shopping results draw on merchant product feeds and catalog integrations (OpenAI’s commerce documentation describes its Shopify catalog integration as requiring no additional merchant setup), and Shopify’s own engineering material states that title matches outweigh description matches in product search.
• Complete, spec-first product titles matter. The queries that convert for us contain exact dimensions, materials, thicknesses and grades. If the spec tokens are not in your product data, you are not in the candidate set.
• Correct product taxonomy matters. Miscategorized products are silently filtered out of category-level shopping queries.
• Things that produced no measurable lift for us: llms.txt files (90 days of server logs, near-zero bot interest — consistent with Google’s public statements that it does not use them) and adding more structured-data markup in pursuit of ranking (we keep schema for data fidelity, not rank).

Context worth reading alongside this data: OpenAI’s agentic commerce product feed specification and Shopify’s public engineering material on how its catalog search weighs titles over descriptions.

Methodology and limitations

• Attribution is GA4 session-source (source-level, not channel-group level — GA4 moved ChatGPT between channel groups mid-July), cross-checked against our Shopify order ledger.
• AI traffic is undercounted by nature. Agents strip referrers, buy through in-assistant surfaces, or send the human back later as “direct.” Treat every number here as a floor.
• n is small. This is a young store publishing early data — dozens of orders, not thousands. We are sharing it because first-party numbers on agentic commerce are scarce, not because the sample is large.
• One store, one category. B2B-leaning commodity packaging is probably an ideal early category for agentic buying: spec-heavy, comparison-friendly, repeat-purchase. Your category may differ.

For researchers and journalists

We will share anonymized monthly data (sessions, orders and revenue shares by assistant) with researchers and journalists on request — reach us via our contact page. We update this page monthly as the data grows.

Browse the store this data comes from: corrugated boxes, poly bags, or the packaging buying guides.