Two shopping agents can agree on how to rank offers and still send you to different stores. The difference may be when they stop looking, or which sellers they can actually buy from.
Suppose two shopping agents recommend the same replacement filter from the same store. Both show a delivered price you’re willing to pay, both say it’ll arrive by Friday, and both ask before spending your money. One checked the store you normally use and found it couldn’t deliver in time. The other never checked it. Looking at the two approval screens, how would you tell?
The description of Stratechery’s interview with Eric Seufert puts two ideas side by side: Seufert thinks ads will be the answer in agentic commerce, and the interviewer wants retailers that cater to agents. The description doesn’t specify how those ads would work. That leaves an important question open: If a seller pays for a chance to be considered, what happens to the rest of the search once the agent finds an offer it can buy?
You can answer that question without seeing a ranking algorithm. You need to see which alternatives the agent checked, what it could confirm about them, and why it stopped. The approval button tells you who gets to authorize the order. It doesn’t tell you who got a chance to compete for it.
An ad can win by being the first offer the agent can verify
The Seufert interview description gives us the case for discussing ads, but no details about placement or ranking. So isolate the mechanism with a hypothetical. Two stores sell the exact replacement filter you want. Both meet your deadline and stay under your spending cap, but the second is cheaper. The first pays to be checked early. The agent’s ranking rule is clean: among verified offers, choose the lowest delivered total. Its stopping rule is different: once you’ve found an offer that meets every requirement, stop. There’s only one offer left for the ranking rule to evaluate. The ad has done its job without adding a point to anyone’s score.
The distinction is easier to spot if you separate your budget from your objective. Your budget says what you’re willing to spend. Your objective says what the agent should do within that limit. If you ask for the cheapest usable offer within a defined search, the first offer under budget is a candidate, not a reason to close the search. If you ask the agent to get a routine order off your plate quickly, the first verified match may be exactly what you want. The same shortcut serves you in one case and works against you in the other. The gotcha is that the final recommendation can look equally competent in both.
There’s a good objection to making every shopping agent continue after it finds a suitable offer: comparison takes time, the practical value of another check may be small, and no service can promise to inspect every seller. A paid lead might even surface an offer that wouldn’t have been checked otherwise. The reasonable demand is a stopping rule you can understand before you shop. For instance, you might ask the agent to check your usual store and an independent alternative before recommending a sponsored offer. If it can’t do that, you’ll want the recommendation labeled as a verified match, not the cheapest deal. A sponsorship label answers who paid for the introduction. A search summary answers whether that introduction ended the comparison. You need both answers to tell whether the ad merely added an option or displaced one.
A seller can be searchable without being buyable
The interviewer’s wish for retailers that cater to agents points to a different failure from early stopping. Continuing the search won’t help if the next offer can’t be verified or acted on. For your filter order, separate three questions. Discovery: did the agent find the exact item at the store? Verification: did it establish the delivered total and arrival date for your address? Checkout: can it place the order on those terms, or do you have to finish it? If the agent answers only the first question, put the store in the comparison as an unverified lead. If it answers the first two but not the third, you may already have enough to choose the better deal and buy it yourself.
Don’t assign blame from the checkout result alone. If the agent hasn’t found a listing, you don’t know whether the seller is inaccessible or the agent didn’t look in the right place. If it can quote an offer but can’t place the order, you still don’t know whether the retailer blocks that route or the shopping service doesn’t support it. That distinction calls for documentation from the side responsible for each boundary. You’ll want the agent provider’s documentation for what it can search and where it can check out, and the retailer’s own documentation for any access rule you’re attributing to the store. If neither tells you, record the observed failure without guessing at the cause.
An unrelated example makes the documentation distinction concrete. Simon Willison’s Datasette release note distinguishes formats the core product can stream from a feature that lets plugins implement other streaming formats, then describes a plugin using that feature. That’s a useful model for reading a capability claim: separate what the service does itself from what a particular connection makes possible. The release note tells us nothing about which stores any shopping assistant can reach. In commerce the distinction would matter because search coverage and checkout coverage answer different questions. If you’re choosing between a higher-priced order the agent can place and a lower-priced one you’d have to finish, the useful comparison is the price of the convenience. Suppressing the latter offer makes that trade-off for you.
A fair comparison keeps the purchase fixed and varies the search
The test to run is a side-by-side comparison of two real shopping agents on one purchase. The interview description supplies a reason to run it, not product documentation or results. The following is a proposed test, not a report of what either service does. Use the same filter model and pack size, destination, deadline, maximum delivered price, and rules on substitutions or memberships. Run the requests close together and tell both agents not to order. Use the same instruction for what counts as enough searching: you want the lowest verified offer across a stated set of stores, including any offer that requires you to finish checkout yourself.
Start with an open request to see what each agent checks by default. Ask for stores checked, verified terms, offers it couldn’t verify, any sponsorship it can identify, and why it stopped. Then give each agent the same named seller missing from the first pass. If the second request produces an eligible offer, you’ve learned that the original route missed an option the assistant could assess when directed. You haven’t established why it missed it. If the terms are verifiable but the agent can’t buy, record a handoff rather than a failed search. Before attributing an omission to ads or retailer policy, look for evidence about that particular placement or boundary. An unknown sponsorship status is an unknown, even if the agent confidently explains its recommendation.
One extra check matters here: compare the terms shown at approval with the terms that earned the recommendation. LangGraph’s release notes document a fix that gave tool calls dispatched by one route the same state as calls made through another. That’s an illustration of a documented handoff problem in agent software, not evidence of a shopping defect. The user-facing question is whether the handoff carries through the exact item, delivered total, and arrival date you compared. If the terms change, the original recommendation no longer settles the choice. Record the new terms separately and ask for fresh approval. That way the test can distinguish three otherwise similar outcomes: an offer didn’t make the shortlist, an offer needed a human to complete it, or an offer reached checkout on different terms.
The retailer is deciding how much of the sale to hand over
Take the retailer side of the Stratechery interview description seriously. A seller considering agent access would have at least two decisions to make: whether to let an assistant get a usable quote, and whether to let it finish the sale. Imagine you already know the exact part you need. The seller might welcome a straightforward order it would otherwise lose. But if you would normally buy from that seller anyway, paying an intermediary to reach you could look like an extra cost on the same sale. Whether an order is new to the store or merely arrives through a new route changes the bargain.
A seller might also want to keep a step where it explains a warranty, confirms a member price, or offers a substitution. From your side, a human handoff for any of those steps could be sensible. The problem comes when the assistant silently treats your willingness to finish checkout as irrelevant. If it only ranks offers it can buy without help, a store can be visible, competitive, and effectively absent from the recommendation. That would be a search policy favoring convenience, not necessarily a retailer refusing to work with the agent.
The Sequence’s essay on agentic swarms describes an architecture in which agents work partly independently and communicate selectively, with feedback changing what happens next. That’s an architecture analogy, not evidence about commerce deployments. Imagine separating the job so one assistant looks for better offers while another establishes whether the leading offer can be purchased. You still have to decide what happens when checkout is ready before the search is done. If readiness cancels the search, splitting the work hasn’t changed which offers get considered. It may merely hide the stopping decision behind a tidy division of labor. The commercial implication is that the service taking your request can define an eligible seller by how it organizes your task. If shoppers routinely delegate the whole order, sellers that make checkout easy could get a reason to invest further in that route. You don’t have to give up that convenience. You’ll want a way to make the assistant consider the seller you name, even when all it can give you is a verified offer and a handoff.
Give the agent permission to buy only after you’ve set its stopping rule
For a repeat order from a retailer you chose, the right stopping rule can be quite narrow. Ask the agent to reorder the exact item from that seller if it meets your delivered-price cap and deadline. If it doesn’t, the agent comes back to you. You gave it a purchasing task, not a mandate to canvass the market. For an unfamiliar purchase, keep approval and ask it to report the best verified offer, the best verified alternative you could complete yourself if there is one, and any named store it couldn’t check. You don’t need a transcript. You need enough information to tell what’s missing and whether it would be worth another check.
If an offer is sponsored, asking for the runner-up may not be enough. The runner-up could already be drawn from a search shortened by the paid lead. Ask for a fresh comparison that doesn’t use that placement as its starting point or its reason to stop. If the assistant can’t identify paid leads or rerun the search on those terms, don’t read that limitation as proof of manipulation. Treat the recommendation as one you can’t independently assess on that point. Then decide whether to do a manual check. You may happily waive one for a low-stakes order and insist on it for a purchase where the price or the delivery date matters.
Finally, distinguish today’s verified offer from tomorrow’s permission to reorder. In Daniel Stenberg’s account of work on curl, the original project kept receiving fixes while agents were producing rewrites. His example concerns software maintenance, not retail. It offers a useful warning about treating a successful snapshot as evidence that two moving things will stay in sync. For a reorder, ask the agent to recheck the agreed terms rather than relying on the last receipt. For discovery, don’t assume last week’s search boundary still suits today’s task. The Seufert interview description pairs a way to fund shopping assistance with a wish for stores that will serve it. The practical question for you is where each gets a vote. Let a retailer’s verified terms determine whether its offer qualifies. Let your search instructions determine when comparison is over. The receipt can then prove what you bought without having to stand in for what the agent considered.
/Sources
/Key Takeaways
- A paid placement can decide the winner before ranking starts if the agent stops at the first verified offer.
- Ask separately whether an agent found the seller, verified the delivered terms, and could place the order. A human checkout handoff can preserve a better offer.
- Compare two agents on the same purchase without granting payment permission. Run a default search and a named-store follow-up, then check what reaches approval.
- Set the stopping rule before giving an agent permission to spend. Your price cap defines an acceptable offer, not how long to keep looking.



