As of November 9, 2026, an audit of 200 retailers found that 29 of them still refuse automated shopping agents at the door, and that the average readiness score across the sample sits at 65 out of 100 PPC Land: bot protection blocks AI shopping agent at 29 of 200 retailers. Read that as a distribution number rather than a web team number. If an agent cannot read what a store has on the shelf, the store is invisible at the exact moment a shopper asks what is available nearby, and the sale moves to whoever answers.
What the audit actually measured
The methodology matters more than the headline. The audit did not ask retailers whether they like agents; it asked whether an agent could complete a normal shopping task on their storefront. A block at that stage is not a philosophical position, it is a wall between a stated need and a purchasable item. Most of the retailers that failed did not fail because they had decided to exclude agents, they failed because a default rule, a JavaScript-only product page, or a missing structured field did the excluding for them.
Category scores point the same way. Baby products reached 88 while fashion reached 49, a 39-point spread inside one sample. Fashion catalogues carry variants, size runs and seasonality, which is exactly the data an agent needs and exactly the data retailers describe in prose rather than in fields. The gap is a data-modelling gap that surfaces as an availability gap.
A blocked crawl is a distribution gap
Distribution coverage used to be counted in stores: how many outlets carry the SKU, how many of those have it on shelf this week. Agent traffic changes the denominator. Coverage now also means whether the shelf state of a given store can be read by a machine at the moment of a question, because that reading is what produces a recommendation. A store that carries the product but cannot be read is, from the agent's point of view, a store that does not carry it.
Instacart's numbers show what the readable version looks like. Its assistant draws on more than 1.6 billion lifetime orders, a catalogue above 2 billion items, and more than 10 million daily inventory signals from close to 100,000 stores across more than 2,200 retail banners, and it only surfaces what is on the shelf at the chosen store at that moment Food & Beverage Department: Instacart deploys AI cart assistant across 2,200 retail banners. The commercial point is not the assistant. It is that availability data had to be normalised across 2,200 banners before any of it worked.
Shelf data nobody outside can read
There is a second-order effect that rarely appears in readiness reports. When an agent cannot confirm stock, it does not simply drop the item, it substitutes. Substitution converts an out-of-stock event into a competitor's sale, which means the cost of a blocked or unreadable shelf is not one lost basket but a permanent handover of that shopper's default choice.
Three postures toward agents
Retailers in the sample sorted into three broad postures, and each one produces a different distribution outcome. Keeping the three separate matters because the fixes are not interchangeable. A block is removed by a rule change, an unreadable page is repaired by a markup change, and a stale feed is repaired by an operational cadence that somebody has to own.
| Posture | What it looks like | What breaks | Effect on coverage |
|---|---|---|---|
| Blocked | Bot rules reject agent user agents outright | No product or stock data leaves the site | Store is invisible to agent answers |
| Open but unreadable | Pages load, but detail depends on client-side scripts | Attributes and availability missing from the fetch | Item appears, variants and stock do not |
| Structured and served | Fields, feeds and pages agree with each other | Nothing structural; freshness becomes the risk | Store can be recommended with confidence |
The middle posture is the most common and the most expensive, because it looks fine in a browser. A merchandiser opening the page sees a price, a size chart and a stock badge. An agent fetching the same page sees a shell. Teams that only test in a browser will keep concluding that everything works while the agent keeps concluding that nothing does.
White-label does not fix the catalogue
Buying an assistant is not the same as being readable by one. The enterprise versions now on the market let a regional grocer run a branded agent on its own storefront, and several chains are already live while others are in the pipeline. That solves the interface problem for the retailer's own customers. It does not solve the problem of a third-party agent arriving with no permission and no structured data.
Running the check
A distribution check for agent traffic can be run in a week without new tooling, provided the team tests what a machine receives rather than what a person sees. The output should be a list of stores and SKUs that are carried but unreadable, which is a different list from the out-of-stock report.
FAQ
Is blocking shopping agents a legitimate choice for a retailer?
It is a choice with a measurable cost. A block removes the store from agent answers, which converts availability into invisibility. Retailers that block should track the substitution rate rather than the block rate.
Why does fashion score lower than baby products?
Fashion catalogues depend on variants, size runs and seasonal phrasing that is often written as prose. Baby products tend to have stable identifiers and simpler attribute sets, so they are easier to serve in fields.
Does an on-shelf availability feed guarantee a recommendation?
No. It removes one reason to be skipped. Ranking still depends on price, assortment and how well the item matches the question asked.
How fresh does store inventory need to be for agents?
Fresh enough that a shopper routed to the store is not disappointed. For fast-moving categories that means hours, not days, and the lag should be measured rather than assumed.
