Read this first
This page is the short, plain form of the paper. Each part links to the part of the paper that backs it. The research stops on 3 November 2026. The main market is the US.
The abstract has the full story.
What this is about
When a brand buys an ad online, a machine bids for a slot on a web page or a TV app. Someone on the seller's side may have put that slot in a pack first. The pack has a code, called a deal ID. The seller's side picks the slots, adds data about the people who may see the ad, and sells the pack.
People call this curation. The firm that builds the pack is the curator. It may be the ad exchange, a data firm, an ad agency, or a TV or app owner.
This paper asks a few plain questions. How much of the trade goes this way? What is in a pack? What does it cost? Does it work as sold? Who owns the curators? Do the courts change it? Do AI agents now make the picks?
The paper says what it counts and how.
How big it is
One firm counts the chances to bid that buyers are offered. By its count for one month, March 2025, nearly three in four of them worldwide came through these packs. That is a count of chances to bid, not of money spent.
A trade group tracks what some big US brands spend. For the typical brand, more than nine in ten dollars went through deals in late 2025. But that count lumps all deals in one box, so it cannot tell us how much went through curated packs. The two counts must never be mixed. No count we found measures the US share of curated packs. So we cannot say if it went up.
What is in a pack
We looked at 21 products. None of them shows all three things a buyer needs to know: which sites are in the pack, the rule that picked them, and where the data came from.
Most say what kind of rule they use, but not the details. Few show who sells in the pack. One names the publishers in small packs. No one shows proof that the ads ran on the sites they named.
The files that tell buyers who sells what have no place for a curator. A new spec can name the curator, but we found no sign that anyone uses it yet.
The paper shows what each product tells a buyer.
Who gets to bid first
Does a pack get to jump the line, or bid at a lower price than others? For most products, their own help pages do not say.
One says a pack gets no head start. Two let the person who builds a pack pick a kind of deal that goes first. For the rest, we could not tell.
The paper shows each product's rules.
Who owns them
Many data firms that build packs were bought. A credit firm bought one. A big ad agency group bought two. A TV data firm bought one. Some ad tech was shut down.
We could not tell who controls many of the firms in 2023, so we cannot prove there are fewer free ones now. But every firm we could track that changed hands joined a larger group or left the market.
The paper lists each deal and its date.
What it costs
None of the products we scored publishes its fee rates. Some explain how the fee is worked out. Many say "no extra fee", but each of those claims covers just one firm's cut. Most of the products we scored do not even say what kind of fee they take.
In one big study of US brands, $710 of each $1,000 that went into the buying tools reached the site that ran the ad. Curation had no line of its own. No one has shown, on the same ads, that a pack costs more or less than the open market.
The paper shows each fee and who takes it.
Does it work
Packs are sold as cleaner, safer and better seen. Every number we found to back that comes from the firms that sell the packs, or their friends. We found no check of a named pack by someone without a stake.
The paper lists each claim and who made it.
The courts
A US court found that Google broke competition law in parts of its ad tech. It chose fixes in September 2026. The fixes are not in force yet. None of the court papers we read talks about curated packs. Europe fined Google, but had not chosen a fix by March 2026.
AI agents
AI agents now help sell some ads. One firm says it ran 30 live campaigns with agents by April 2026. Only one story we read says what the agents chose: in Spain in 2026, agents picked spots on streaming TV for a phone company. Other stories, for a drinks brand and for car dealers, say agents set up the buy but not what they picked. Every story comes from the firms that ran it. No one else has checked the results.
The author helps lead a group that writes one of the rules for these agents. His own firm runs two agents listed there. None of that counts as proof in this paper.
The paper shows what each agent has done.
The scores
We scored what each product shows in public, not how well it works. Products that show little get low scores, even if they work well. 4 exchange products sit in the top band any product reached. Two groups had too few well-documented products to rank at all. Most scores rest on the firms' own ads and news posts, so the bands could move.
The paper explains the scores.
What to ask for
Before you pay for a pack, get these in writing:
- the list of sites in it;
- every fee, with who pays and who gets it;
- whether it gets to bid first or at a lower price;
- a check of its claims by someone with no stake;
- who can change it after you buy, and how you will know.
The paper turns this into a checklist.
About this paper
AI tools did the research. A second AI tool checked each file and tried to prove it wrong. In all, 389 sources were read. The scores were set in one pass, with no second scorer. Apart from one AI review of the plan and one of the finished paper, no outside party checked the work. We fixed what held up from the second review on 4 November 2026. No vendor paid for it or saw it.
The author works at Samba, which owns a data firm in this paper. He once worked at Verve and Lotame, which are in it too. These links are marked where they show up.