The Shelf You Can Restock
Friday's poll asked what becomes the new shelf space once agents decide what gets considered, and the smallest turnout of the run, 16 ballots, came back tied: Retrieval results and Paid recommendations at 31% each, Training data and Default integrations at 18% apiece. A tie that thin settles nothing, and the essay says so before it says anything else. What survives the noise is the grouping, and the Restock Test explains it: sort every route into a consideration set by how long a change takes to land and whose approval it needs, and you get minutes for a bid, weeks for retrieval, quarters for an integration deal, and a model release for training data. The room picked the two it could move inside a planning cycle. The vote here went to paid recommendations for one reason — it is the only route of the four that carries a price, and a price brings an auction, an inventory owner, and somebody who answers when the placement disappears.
In English, please
A recurring reader poll asked how a company gets noticed once AI assistants, rather than people, are the ones drawing up the shortlist. Ask an assistant to recommend a supplier, a tool, or a product and it names a handful of options; landing in that handful is the new version of landing on a store shelf. The poll offered four routes in: being written into the AI model itself while it was trained (Training data), being findable by the AI's live lookup step (Retrieval results), being the tool the assistant is wired to reach for automatically (Default integrations), and paying for the spot (Paid recommendations).
Only 16 people voted, which the essay flags at the top as too few to prove anything. The top two answers tied at 31% each and the bottom two tied at 18% each. In whole ballots that is five, five, three, and three, so two people changing their minds would have produced a four-way tie. It was the smallest turnout of any poll in the series, where the biggest drew 91.
The essay's tool for reading the result is called the Restock Test, borrowed from the shelf idea: if you want to change your position, how long does that take, and whose approval does it need? Paying for a spot changes in minutes and needs only your own decision. Becoming findable takes days or weeks of publishing and waiting to be indexed. Becoming a built-in default takes a business deal, so months. Being baked into the model waits for the next model, and you cannot take it back. The two answers people picked were the two they could act on inside a normal budget cycle.
The author voted for paid recommendations, for a commercial reason rather than a technical one: it is the only one of the four with a price on it. A price means there is an owner, an auction, an invoice, and somebody on the hook if what you bought fails to show up. The other three are influence with no counterparty attached. If your findability drops because the AI's ranking rules changed overnight, there is nobody to complain to. The essay is explicit that this is not advice to spend there yet, since that market is still small and barely measured; it is a claim about where the money eventually collects.
On this page
The cold open
Here’s the question as it ran: when agents decide what gets considered, what becomes the new shelf space? Options: Training data. Retrieval results. Default integrations. Paid recommendations. Core thesis at launch: distribution shifts from reaching people to gaining eligibility inside the agent’s consideration set.
Shelf space is a useful metaphor with one thing hidden inside it. A supermarket shelf is a fixed asset. Somebody built it, somebody owns it, and the number of facings is known before a single shopper walks through the door. An agent’s consideration set has none of those properties. It gets assembled at the moment the question is asked, out of whatever the agent can reach right then, and between questions it doesn’t exist at all. So the four options in this poll aren’t four kinds of shelf. They’re four mechanisms by which a company becomes reachable when the shortlist gets built, and they run on very different clocks.
The industry has already watched a smaller version of this play out. Curation turned media buying into a shelf of pre-cut packages in under two years, and the open question there was what happens to that shelf once the buyer arrives with an agent that can assemble supply per brief. This poll asks the same question one layer up, about the assistant rather than the exchange.
The vote
The poll closed with 16 votes. (Shares are LinkedIn’s rounding, which is why they sum to 98.)
| Answer | Share |
|---|---|
| Retrieval results | 31% |
| Paid recommendations | 31% |
| Training data | 18% |
| Default integrations | 18% |
Start with the least flattering number on the page. Sixteen is the smallest turnout of the eleven Friday polls run across both seasons so far; the largest was 91. At that size the shares resolve to five ballots, five ballots, three, and three. One person changing their mind produces a different headline. Two people produce a four-way tie. Nothing in this result is a verdict, and that belongs at the top of the essay rather than in a footnote at the bottom.
What survives the noise is weaker than a ranking and still worth a paragraph. The two answers at the top are the two routes into a consideration set that an operator can change without anyone else’s approval, on a timescale shorter than a fiscal year. The two at the bottom are slower and belong to somebody else. Sixteen people can’t establish that split as a fact about the market. They can only be consistent or inconsistent with it, and this room came out consistent, down to a tie between the two options a media buyer would recognize as things you can actually go and buy.
One more caveat before the mechanism. Operators answer poll questions from inside their own budget authority, and both winners sit squarely inside it. Treat the result as a description of how the job currently feels, and hold the forecasting claims to a higher bar than 16 ballots can clear.
The Restock Test
The Restock Test: for any route into an agent’s consideration set, ask two things. How long between deciding to change your position and the change showing up in what the agent considers? And whose approval does that change require?
A route with short latency and the switch in your own hands behaves like media. You buy it, you measure it, you turn it off when it stops paying. A route with long latency and the switch in somebody else’s hands behaves like infrastructure. You invest in it, you defend it, and you can’t correct it mid-quarter. Both are worth having. They don’t get the same budget line or the same review cadence, and confusing the two is how a company ends up filing a brand-equity program under performance marketing.
Run the four.
Paid recommendations. Latency: minutes. Approval: yours, assuming somebody is selling the slot. This is a bid. You raise it, cut it, or switch it off before lunch, and the effect on what the agent surfaces is immediate and reversible. It’s also the only one of the four that arrives with a price attached, which drags an auction, an invoice, and a counterparty who has to answer for delivery along behind it.
Retrieval results. Latency: days to weeks. You publish, a crawler finds it, an index or an embedding store absorbs it, and your material becomes reachable for the questions it answers. Approval: mostly yours, with one large asterisk — the ranking function belongs to whoever runs the agent and can change without notice. Operators recognized this option fast because they have lived it. It is search optimization with a different retrieval layer underneath, and it inherits the economics: cheap to attempt, open to everyone, and competed down toward parity as soon as everyone attempts it. The work itself is real and already reasonably well specified. Structuring what you publish so a reasoning system can evaluate it is most of what context agents actually consume.
Training data. Latency: a model release. Approval: nobody’s, in any usable sense. Whether your company, your product, or your category shows up in a model’s weights was settled by what existed on the open web before the cut-off. You can move it over years by publishing consistently. You cannot restock it for Q4, and you cannot take back what is already in there. Slow to acquire and slow to lose — that is the profile of brand equity, which is a real asset and a poor answer to a question about shelf space.
Default integrations. Latency: a business-development cycle, so quarters at best. Approval: theirs. Being the tool an assistant reaches for by default is a distribution deal in new clothes, and the industry has decades of experience with the old version of it. This is the most valuable position of the four once you hold it and the least actionable of the four while you don’t, because the only lever available is persuading someone else’s roadmap.
Line the four up by latency and the vote sorts itself: minutes, weeks, quarters, model releases. The room’s top two are the first two on that list. The gap between the pairs is the gap between what an operator can act on this quarter and what an operator can only lobby for.
Those latencies aren’t arbitrary. Each route acts at a different point inside the agent’s own pipeline, and one of the four never has to enter that pipeline at all.
What would change the ranking
The Restock Test sorts by current latencies, and each of the four has a mechanism that could shorten its own.
Training data shortens if frequent post-training becomes routine rather than annual, and it partly dissolves in any case. When an agent grounds an answer in a live index, the weights stop being the only path into what it can say about you. That is the direction most production systems have already taken, which is part of why retrieval polled where it did.
Default integrations shorten the moment a platform swaps business development for enrollment: a published registry, a manifest format, stated criteria for what qualifies as a connector. That is the difference between negotiating carriage and submitting to an app store, and it is a decision the platforms make on their own schedule.
Retrieval results get harder as more companies do the work, in the ordinary way that any open, non-exclusive channel decays toward parity. The advantage there is a timing advantage, and timing advantages expire.
Paid recommendations lengthen if disclosure rules or platform policy reduce how much a purchased slot actually moves a shortlist. Nobody knows yet how much a disclosed recommendation is worth compared with an organic one, because there isn’t enough of either to measure.
None of these is far-fetched, and any single one of them reorders the list. Re-run the test when the latencies move.
My vote
I voted Paid recommendations, and the reason is the price, not the volume.
A price is the only thing on this list that creates a market. When a slot has a price it also has an auction, an owner of the inventory, a clearing mechanism, and a contract that states what was sold. The other three routes are influence. You can work at them, and the work pays, but nobody can sell you a guaranteed position in a consideration set, which means nobody can build a business on selling you one either. Markets concentrate around whatever has an owner and a price, and this season’s question is where money and power move.
The second reason is that the inventory already exists and somebody is already selling it. An agent platform doesn’t own the supply, the demand, or the product it recommends. What it owns is which options get named, and that is the whole of its inventory. OpenAI put a rate card on the answer at Cannes, the minimums have already collapsed from $200K to zero, and every trust metric attached to the format is self-reported with no referee. That is the ordinary opening sequence of a new ad format, and it is running now rather than arriving later.
The uncomfortable part of this vote is that it isn’t an endorsement. Paid recommendation being the biggest shelf doesn’t make it the healthiest route for advertisers or the most useful one for the person asking the agent a question. It makes it the route with a counterparty attached, and that matters more than it sounds like it should. If your retrieval position drops because a ranking function changed overnight, there is no invoice, no service level, and nobody who owes you an explanation. A purchased slot at least has somebody who answers for it, which is the difference between a transaction and an outcome you were hoping for.
The other 31% has the better near-term case, and I’d say so to anyone planning next year’s budget. Most companies will get more out of retrieval work over the next eighteen months than out of buying placement, because the placement market is small and barely instrumented, while retrieval work compounds and costs a fraction. Those two claims sit together fine. Retrieval is where the effort belongs right now. Paid recommendation is where the shelf ends up.
What the next weeks test
Eligibility was the question of getting into a consideration set. Everything after it is about what happens once you’re in one. Week 3 has since closed and answered: Outcomes took 57% of 73 ballots — the room voted for the one unit nobody can sell yet, against this week’s 16. Week 4 asks what buyer and seller agents negotiate hardest once they transact directly. Week 7 asks who owns the customer relationship when the agent picks the brand, which is this week’s question asked after the transaction instead of before it.
The reading in this essay is checkable in an ordinary way. If paid placement inside assistants stays small through 2027 while retrieval optimization turns into the line item every marketing organization funds, then the room’s other 31% was right and I was wrong. Both numbers become public on a normal schedule: platform ad-revenue disclosure on one side, agency and in-house line items on the other.