The Q2 AI1000 moved 972 of 1,000 retailers. Read the count, not the ranking.

TL;DR

The Q2 2026 AI1000 reports that 972 of the thousand largest US retailers changed rank in a quarter, that the median retailer moved 35 places, and that 225 are detectably agent-ready. The churn number is close to meaningless: rank is ordinal, and 35 places out of 1,000 is three and a half percentiles. The count of 225 is solid. The interesting part is a caveat the newsletter prints and then argues against: most of those endpoints were picked up "for free". If that is true, the 22.5% is measuring which platform a retailer sells on, not what any retailer decided to build.

What the file says

The AI1000 ranks the thousand largest US retailers on how readable they are to AI agents. It is published by ReFiBuy, built on Digital Commerce 360's IR1000 list, and scored on four factors in its own words: "bot friendliness (including detectable UCP support), AI source traffic, diversity of AI sources, and 90-day momentum."

The primary carries one figure the data drop covering it left out, and it is the one worth arguing about: retailers with an endpoint have a median rank of 304, and retailers without one sit at 550. A 246-place spread attached to a single implementation decision.

The number that travelled is the weakest one

972 of 1,000 sounds like a market in upheaval. It is not.

Rank is ordinal. On a list a thousand rows deep the gaps between adjacent scores are tiny, so a scoring difference far too small to matter commercially moves a retailer dozens of places. The median move was 35 positions, which is three and a half percentiles. Stated that way it stops being a story, because it is what any densely packed ranking does when you re-run it.

It is worth noting who benefits from the framing. ReFiBuy operates in agentic commerce, and a ranking that barely moved would not justify republishing every three months or get written up when it did. That does not make the data false, and the underlying work is good enough to cite. It does tell you which numbers to lean on. Counts survive a change of formula. Ranks do not.

There is also no revenue term, no price term and no assortment term in that scoring. The AI1000 is not a retail ranking that mentions AI. It is a legibility score with retailers attached.

The contradiction in the source is the finding

Here is the passage that matters, from the data drop, in full:

Read those last two sentences together. The bulk of adopters got the endpoint for free. Therefore you should implement it as your fastest win.

Both cannot be the advice. If most of the 225 did not run a project, then the 225 is not a leaderboard of retailers who moved early. It is a map of which commerce platforms shipped the protocol on their merchants' behalf. And the retailers who did not get it for free are precisely the ones for whom it is neither fast nor easy.

The caveat is the more interesting sentence, and it is the one printed in quotation marks and then dropped.

What "for free" actually means

Shopify implements the protocol across the whole journey for merchants on it: capability negotiation and authentication, product discovery, carts and checkout, and order monitoring. Its Global Catalog lets an agent query "products across all Shopify merchants". Shopify also decides which agents may transact, through a trust tier system where "higher trust tiers unlock broader access, including direct checkout completion" and "trusted agents can complete checkouts directly".

A merchant there became agent-readable because of where it sells. It did not decide anything, and it cannot un-decide it either.

This also puts the 246-place median gap in a different light. It may be a gap between retailers on modern commerce platforms and retailers not on them, with the endpoint acting as a marker for the platform rather than a cause of the rank. Two retailers with identical intentions land 246 places apart because one of them sells somewhere that did the work.

The 4GPTs read: the agent-readable layer is being built, and mostly not by the companies being scored on it. It is being built by a small number of platforms on behalf of everyone hosted there, which from the outside, measured with a crawler, looks exactly like retailer initiative.

Which makes the advice wrong for almost everyone

Split the thousand and the recommendation falls apart in both halves.

If you are on a platform that implements the protocol, you already have the endpoint. Your score is not a to-do item, and any sprint aimed at moving it is wasted budget.

If you are not, then it is neither the easiest nor the fastest thing available to you. It is a genuine engineering commitment across discovery, cart and checkout, and it should be justified by whether you want agents to be able to transact with you at all, not by a position on a quarterly list. The honest version of the advice is that your platform is not going to do this for you.

The 22.5% does not tell you which half you are in. That is the first thing to go and find out, and the ranking cannot answer it.

What an independent read says

Forrester, who sell nothing into this ranking, describe the same market publicly:

They add that very few consumers allow agents to complete purchases without direct oversight, and that "the gap between marketing buzz and current reality is exactly where digital leaders must operate strategically, right now." A crawlability score measures how ready the infrastructure is. It does not measure whether anyone is buying through it yet, and those two things are at visibly different stages.

The counter, and the test

The strongest objection to all of this is size. Retailers who are agent-readable skew large and well resourced, and large well-resourced retailers rank highly for reasons that predate any protocol. The 246-place spread is consistent with legibility mattering, and equally consistent with size producing both the endpoint and the rank.

The dataset can settle it, because "90-day momentum" is one of the four scored factors. So the Q3 test is specific: do retailers who newly expose an endpoint climb as a cohort rather than as anecdotes, and does the median gap hold as adoption spreads past the platform-hosted group? If adoption rises and the gap collapses, the endpoint was a proxy for platform choice, not a cause of rank.

That prediction is falsifiable in twelve weeks, which is more than the headline number offers.

FAQ