How the Agentic Commerce Index is built
Every leaderboard number links here. This page exists so a skeptic can reproduce us and an editor can cite us.
What this index observes
Each week we reconstruct the public product shelf that AI shopping surfaces draw from — the same organic retail results a shopper would see for a buying query, resolved per country. When an AI assistant answers a “what should I buy” question, this is the shelf it reaches into. We publish what we observe on that shelf — not the model's internals, which no one outside the AI vendor can see.
The three layers
We keep these three separate on every page and never collapse them into one number.
Who appeared in the answer, on which dates, across how many runs. Pure fact — no interpretation.
Attributes that co-occur with appearing: presence in Google free listings, a machine-readable price, a product image. Correlation, stated as correlation.
Our hypothesis about what a store should fix. Clearly marked as xpay's interpretation — never as observed fact.
How a cell is produced
- A replayable brief. Each category × country has a plain shopping prompt — the same one we publish so you can rerun it yourself.
- Repeated observation. We run the brief many times on a scheduled day (higher-tier cells get more runs) and record what surfaces each time.
- Brand resolution. Results are grouped to the brand that sells them, so a store appears once regardless of how many products it fields.
- Aggregation across runs. Because a single run is noisy, we report each brand's share of answers (how consistently it shows up) and how stable its position is — never a single snapshot.
- Weekly movement. This week is compared to last week to produce the entered / exited / displaced you see on every page.
Reproduce it yourself
Verify this yourself — paste into ChatGPT:
Act as a shopper in the United Kingdom. I want to buy skincare online. Recommend the best independent brands and where to buy — give me a ranked shortlist with the store each is sold at.
AI answers vary run to run — that is exactly why we publish share-of-answer and stability across many runs rather than a single result. Reconstruction is not official OpenAI data.
What can make our data wrong
- Index lag — the shelf we reconstruct trails live catalog changes.
- Locale & language variance across countries.
- Brand-resolution errors — report one to
corrections@xpay.sh. - Marketplace seller ambiguity (a marketplace listing may front many brands).
price = nullmeans unknown, never free.
Category & country coverage
43 categories × 23 countries = 989 cells. Tier A+B ship first.
Corrections & claims
A brand can dispute any row. Email corrections@xpay.sh with the cell URL and the correction, or claim your store from any leaderboard row. We version this methodology (Methodology v1.0) and date every change.
Frequently asked
Is this official OpenAI or ChatGPT data?
No. We reconstruct the public product shelf that AI shopping surfaces draw from and publish what we observe. It is not any AI vendor’s internal data, and we never present it as such.
Why movement instead of a score?
A single "readiness score /100" is a static number that stops being interesting after one look. What matters to an operator is whether they entered or left the answer this week, and who displaced whom. So we publish share-of-answer, stability, and week-over-week movement — never a composite score.
Why does a brand show a price of "—"?
Because we did not observe a machine-readable price for it. "—" means unknown, never free. We never render a $0 for a product whose price we could not verify.
How often is the index updated?
Weekly. Each cell is a scheduled run on a fixed day; the movement diff compares the latest run to the prior week.
My brand is listed under the wrong name or domain. How do I fix it?
Email corrections@xpay.sh with the cell URL and the correct brand/domain. We resolve results to the selling brand, but marketplace sellers and multi-brand domains can be ambiguous — we correct within our stated SLA.
Can I reproduce your numbers?
Yes — every cell publishes the exact shopping brief we run. Paste it into ChatGPT yourself. Answers vary run to run, which is why we publish share-of-answer and stability across many runs rather than a single result.