AgentReady at agentready.market, operated by Villepinte Studio — the e-commerce transaction audit. Unaffiliated with other AgentReady-named products.
Every audit on this site presupposes one thing it could not show: that a store an agent can read and buy from is a store agents actually go to. The full proof needs data we do not hold yet — traffic on the merchant’s own servers, checkouts completed by agents. What we do hold, on our own surfaces, is attention: an agent asking our tools about a store by name, an assistant opening a store’s report, a person arriving on that report from an assistant’s answer. This page joins that attention to the score, store by store and week by week, and reads the result against a bar fixed before the numbers existed.
What is measured, and what is not
Four signals, all logged first-party on agentready.market, none inferred:
- Asked. An AI agent called our MCP tools or partner API and named the store. This is the primary signal: an agent chose the store on its own, independent of anything we rank or display.
- Read. An assistant, or a crawler that feeds AI answers, fetched the store’s report page. Search engines are excluded.
- Delivered. A person arrived on the store’s report from an assistant’s answer (referrer or tagged link).
- Claimed. Someone recognised the store as theirs on the report.
Alongside each week of attention: the store’s score that week (its latest scan, inherited for at most eight weeks, never mixed across scoring versions) and whether the store’s checkout door answers an agent-class user agent at all (a fact the scan records, unscored).
What this does not measure. Traffic on the merchant’s site. Carts. Orders. Nothing here says “bought”. The full chain, readiness → agent traffic → task completed → revenue, is written down in our proof schema with the fields each step must capture; this page is its first column, the one we could fill today.
The protocol, fixed on September 9
- Population. Every store in the audit corpus with at least one attention signal in the trailing 180 days.
- Qualified. A store enters the comparison only with a score, at least four weeks carrying an asked-or-read signal, and at least two distinct agent families behind them (so a single crawler cannot make the series). Nothing is published below thirty qualified stores.
- Bands. Scores 0–40, 41–79, 80–100, the cut the Index and the report’s plain-language verdict already use.
- Primary comparison. Median asked per attention-week in the 80–100 band against the 0–40 band. Bar: 1.5×. Below it, we write that a higher score does not go with more agents asking, in the same live block, in the same words.
- Secondary, confounded by construction. Read is reported but not used for the verdict: the Index ranks stores by score, so a crawler walking the Index reaches high scores first. A comparison on reads would partly measure our own page order.
- Named confounders we cannot remove. Store size (we have no traffic figure for the merchant) and vertical (the bands mix them; the verticals present are printed under the table). Read the ratio as a fact about this corpus, not a within-vertical or size-matched one.
- Correlation, never cause. A store that fixes its access pillar and is asked about more the month after would be the causal design; the series makes it possible and this page does not claim it.
The measurement, live
59 stores carry a complete series (threshold 30), out of 491 with any attention in the last 180 days. Totals over those: asked 339, read 9132, humans delivered 3, claimed 1. As of 2026-11-13.
Reading against the pre-registered bar: stores scoring 80 or more are asked about 2.0× as often per week as stores scoring 40 or less (bar: 1.5×). On this corpus, readiness goes with agent attention. A correlation on our own surfaces, not a cause, and not a purchase.
Robustness check (added after the reading): 200 of 229 asks (87%) came from a client that had read our own Index the same day, and the Index lists stores by score. Those asks may be names we gave the agent rather than stores it chose. Without them, one band has no asks left, so no ratio can be computed: the reading above is not confirmed by this check, and we do not claim it.
| Score band | Stores | Median asked / week | Median attention / week | Humans delivered | Claimed | Checkout door blocked |
|---|---|---|---|---|---|---|
| 80–100 | 24 | 0.4 | 3.1 | 0 | 0 | 0% |
| 41–79 | 25 | 0.1 | 2.7 | 0 | 0 | 0% |
| 0–40 | 10 | 0.2 | 1.8 | 3 | 0 | 0% |
Verticals in the top band: beauty 4, general 1, supplements 9, specialty_coffee 10. The bands mix verticals; read the ratio as a corpus fact, not a within-vertical one.
The rows come from proof_series, a scheduled job that joins the first-party
event log to the scan table every morning and stores one summary; the block
above renders that summary. The thresholds and the bar are printed with it, so
a later change of rule is visible on the page rather than silently absorbed.
Why publish attention at all
Because it is the only column of the proof that exists without the merchant installing anything, and because the alternative was to keep asserting the chain. The next columns need the merchant’s cooperation: an agent-traffic beacon on their side, and a buy-path run on their store that stops before payment. Both wait on merchants who want them. Until then, this table is what “agents notice agent-ready stores” rests on, with its limits stated.
FAQ
Does this prove that a better score brings more sales?
No. It relates the score to how often agents ask about, read, and send people to a store on our own surfaces. Sales are not observed here.
Why is “asked” the primary signal and not “read”?
Reads are shaped by our Index, which lists stores by score, so they cannot separate readiness from page order. Asks were meant to escape that: an agent asking about a store by name seemed to have chosen it. The data says otherwise for most of them. Many asks come from a client that read our Index the same day, so the name may have come from us. The daily job counts those asks and recomputes the ratio without them, and the live block prints that check next to the verdict; the strong reading stands only if it survives it.
Correction, 23 October 2026
An earlier version of this FAQ said an agent asking about a store “chose the store itself”, and the live block printed the verdict without the Index check beside it. Both were corrected on 23 October: the check now renders next to the verdict, in numbers computed by the code, and the claim above is withdrawn whenever the check does not hold.
What would make you withdraw the claim?
The bar is on the page: if the top band is not asked about at least 1.5× as often as the bottom band, the live block says the score does not predict attention, and that sentence stays.