A private marketplace does not guarantee inventory quality. No platform commits to it, and the protocol does not check it. What it guarantees is who may bid, at what price, and with what priority. Quality stays a filter the seller applies on their own side, exactly as in the open auction.

That does not make these deals useless. But the measured evidence says something more precise than the sales pitch: their quality depends almost entirely on how many domains they bundle, and the gap between the small ones and the large ones is enormous.

What the deal actually commits to

Four transaction types coexist. The open auction, available to all. The private auction, restricted to an invited list of buyers, with a floor price. The preferred deal, at a negotiated fixed price the buyer stays free not to exercise. Guaranteed buying, where both volume and price are committed and the inventory leaves the pool before the auction happens.

The first two are still auctions. The last two are not, and the specification formalizes that in a dedicated field allowing three values: first price, second price, or a third where “the value passed in bidfloor is the agreed upon deal price”. In that case there is no auction at all, and the winning buyer pays the set price even if they bid well above it.

The vocabulary also diverges between platforms, and not marginally. At one, the private auction explicitly competes with open market bids while the preferred deal does not. The standards body concedes there is no single rule:

“There are multiple common practices that depend on how the publisher prefers to sell inventory with Deal ID.”

And the deal identifier verifies nothing. The official implementation guidance is precise about what it is:

“A Deal ID should be treated as a ‘targeting token’ associated to orders, line-items or campaigns.”

A targeting token. The protocol filters who may bid, through allowed seat lists and advertiser domains. It nowhere describes what is delivered, and no field lets a buyer confirm afterwards that the impression matched the negotiated criteria. The three public supply chain files do not close that gap either: they attest to the seller’s identity and authorization, not the conformity of the content.

Which shows in what platforms promise, and what they carefully do not. None of the five documentation sets examined commits to inventory quality. They commit to access, price, priority, and for guaranteed buying alone, volume. The most explicit wording is almost defensive:

“Curated transactions are not guaranteed deals and will not override any seller account configurations, including ad quality settings or floor price settings.”

What each programmatic deal type commits to contractually and what no type commits toWhat each programmatic transaction type commits to contractually and what no type commits to. Four types coexist. The open auction is available to all buyers with no commitment on price or volume, operating under whichever auction mechanic the marketplace configures. The private auction restricts participation to an invited list of buyers and sets a floor price, but preserves the competitive mechanic: only the pool of bidders and the floor change. The preferred deal sets a negotiated price and gives the buyer an initial or preferred opportunity to bid on the inventory, without reserving that inventory and without obliging the buyer to purchase, the documentation stating that the negotiated inventory is not reserved for the buyer and that buyers are not required to buy the inventory. Guaranteed buying commits both volume and price, the inventory being removed from the available pool before the auction takes place, so that it does not participate in a priority competition but exits in advance. The technical specification formalizes this gradation in a dedicated field of the deal object allowing three values: first price, second price plus, or a third value for which the value passed in the bid floor field is the agreed upon deal price, a case in which there is no auction at all and the winning buyer pays the fixed price rather than their bid. A further field indicates whether the deal is of guaranteed type and whether the bidder must bid on it. What these mechanisms do not establish matters more: the official implementation guidance describes the deal identifier as a targeting token associated to orders, line items or campaigns, and nothing in the protocol allows a buyer to verify after the fact that a delivered impression matched the negotiated criteria of content, placement or brand safety. Allowed seat lists and advertiser domain lists filter who may bid, not what is delivered. The three public supply chain files attest to the identity and authorization of the seller, not the conformity of content to the deal. Finally, none of the platform documentation sets examined contains a commitment on inventory quality, the most explicit wording stating that curated transactions are not guaranteed deals and will not override any seller account configurations including ad quality settings and floor price settings.Four types, three possible commitmentsTypePriceVolumeQualityOpen auctionbidnonoPrivate auctionfloornonoPreferred dealfixednonoGuaranteed buyingfixedyesnoWhat the deal identifier is, officially”A Deal ID should be treated as a ‘targeting token‘“It filters who may bid. It does not describe what is delivered,and nothing lets you confirm afterwards that the impression was conforming.
Volume is committed in one case only. Inventory quality, in none. Source : IAB Tech Lab, OpenRTB 2.6, sections 3.2.11 and 3.2.12, April 2022, and implementation guidance section 7.3; official documentation of Google Ad Manager, Microsoft Xandr and Index Exchange, consulted 9 September 2026 (2022)

The measurement

Across 21 advertisers, $123 million of spend and 35.5 billion impressions, private marketplaces were compared by the number of domains they aggregate. Forty-one percent of the programmatic spend studied ran through them.

Deal typeCPMMade-for-advertisingInvalid trafficViewability
1 to 10 domains$6.030.02%0.2%67.7%
11 to 500 domains$8.491.3%0.2%75.6%
Over 500 domains$5.5219.8%1.4%72.1%
Open auction$2.7526.8%1.1%78.8%

The small ones deliver, spectacularly. A deal of ten domains or fewer carried 0.02% made-for-advertising inventory against 26.8% in the open auction. That is a ratio above a thousand to one.

The large ones do not. Above 500 domains you are back at 19.8%, and the report draws the arithmetic conclusion:

“PMPs with more than 500 domains delivered quality similar to the OMP deals, but at double the cost.”

The doubling is exact: 5.52 divided by 2.75 is 2.01. And for some participants, made-for-advertising activity “reached over 30 percent in Private Marketplace deals”.

The study also names that third profile without ambiguity. Its footnote describes these deals as sold by a supply chain intermediary that bundles sites from many publishers, rarely carrying auction priority, where “the value proposition is inventory curation”.

Quality and price of private marketplaces by number of domains aggregatedQuality and price of private marketplaces by the number of domains they aggregate, measured across twenty one advertisers, one hundred and twenty three million dollars of spend and thirty five point five billion impressions between September two thousand twenty two and January two thousand twenty three. Forty one percent of the programmatic spend studied ran through private marketplaces and fifty nine percent through the open auction. Private marketplaces containing one to ten domains, described as publisher specific, show a cost per thousand of six dollars and three cents, two hundredths of one percent of impressions from made for advertising sites, two tenths of one percent invalid traffic and sixty seven point seven percent viewability. Those containing eleven to five hundred domains, described as seller specific auction packages, show a cost per thousand of eight dollars and forty nine cents, one point three percent made for advertising, two tenths of one percent invalid traffic and seventy five point six percent viewability. Those containing more than five hundred domains, described by the study as sold by a supply chain intermediary that bundles together sites and apps from many publishers, rarely carrying auction priority, and whose value proposition is inventory curation, show a cost per thousand of five dollars and fifty two cents, nineteen point eight percent made for advertising, one point four percent invalid traffic and seventy two point one percent viewability. The open auction, taken as reference, shows a cost per thousand of two dollars and seventy five cents, twenty six point eight percent made for advertising, one point one percent invalid traffic and seventy eight point eight percent viewability. The study concludes that private marketplaces with more than five hundred domains delivered quality similar to open marketplace deals but at double the cost, a ratio that is arithmetically exact since five dollars and fifty two cents divided by two dollars and seventy five cents equals two point zero one. It further reports that made for advertising activity reached over thirty percent within private marketplace deals for some study participants.The bigger the deal, the less it protectsDeal typeCPMMade-for-advertisingViewable1 to 10 domains$6.030.02%67.7%11 to 500 domains$8.491.3%75.6%Over 500 domains$5.5219.8%72.1%Open auction$2.7526.8%78.8%What the study concludes”PMPs with more than 500 domains delivered quality similar to the OMP deals,but at double the cost.”And for some participants, junk inventory exceeded 30% inside private marketplace deals.
Under ten domains, 0.02% junk. Over five hundred, 19.8%, at double the open auction price. Source : ANA, Programmatic Media Supply Chain Transparency Study, December 2023, pages 67 to 70 (2023)

Why the good inventory does not stay in the auction

This is where the article stops being a complaint and becomes an explanation. The pattern is not a failure of diligence. It is what a rational seller does.

Auctions beat negotiation, and the result is old. A foundational 1996 paper establishes that a seller negotiating optimally with a given set of buyers earns less than one running a simple auction with a single additional bidder. One more bidder is worth more than all the bargaining power in the world. Moving inventory out of a wide auction into a bilateral fixed-price deal costs the seller that advantage, unless something compensates.

A 2023 paper models what happens when both channels coexist. It identifies a devaluation effect: the existence of a private channel informationally disadvantages advertisers excluded from it, which lowers their willingness to pay in the open auction, which lets the connected advertisers win with lower bids. The authors conclude that under identifiable conditions the publisher is “better off not introducing a private exchange, even if it is costless for the publisher to do so.”

And in 2025 the adverse-selection prediction was tested. A paper in Management Science models two channels and finds that publishers running both “can leverage their private information on impression quality to sell lower-quality impressions at higher prices in RTB, leading to adverse selection and exposing their RTB-only counterparts to losses.” Its empirical section confirms the prediction: real-time bidding impressions from dual-channel publishers “are of significantly lower quality compared with those from single-channel publishers.”

A publisher knows the quality of its inventory better than any buyer does. The theory predicted decades ago that the best of it would be reserved for the negotiated channel. The measurement caught up in 2025.

One caveat on that paper: I could not obtain the full text, so the sample size and the quality measure are unverified. The conclusion is confirmed by two independent institutional listings of the publication.

Theoretical predictions and empirical confirmation on inventory allocation between negotiated channel and open auctionTheoretical predictions and empirical confirmation on the allocation of advertising inventory between a negotiated channel and the open auction. A foundational paper published in nineteen ninety six in a leading economics journal establishes that a seller negotiating optimally with a given number of buyers earns less revenue than one running a simple English auction with one additional bidder, a result robust under both independent private values and common values, whose consequence is that an extra bidder is worth more to the seller than all of its bargaining power. A theoretical paper published in twenty twenty three in a marketing science journal, devoted specifically to open and private exchanges in display advertising, identifies two effects. The first is competition thinning, as advertisers are spread across multiple auctions. The second, called the devaluation effect, arises because the existence of a private exchange informationally disadvantages the unconnected advertiser, lowering its willingness to pay for impressions in the open exchange, which in turn allows the connected advertiser to win impressions with lower bids; the authors conclude that if the baseline fraud intensity is mild and advertisers’ average willingness to pay for a legitimate impression is high, the devaluation effect outweighs the gains from mitigating fraud, so that the publisher is better off not introducing a private exchange even if doing so were costless. A paper published in a management science journal in twenty twenty five models two channels and concludes that publishers using both real time bidding and private marketplaces can leverage their private information on impression quality to sell lower quality impressions at higher prices in real time bidding, producing adverse selection that exposes publishers operating only that channel to losses; its empirical analysis finds that real time bidding impressions from dual channel publishers are of significantly lower quality compared with those from single channel publishers. A literature review published in twenty twenty observes separately that many publishers still prefer selling premium high quality inventory through guaranteed contracts and limit dynamic allocation to lower quality inventory. A measurement gap remains: all of this literature reasons from the publisher’s profit, and no institutional or academic source compares the return obtained by the advertiser across the two channels on equal terms.The good inventory does not stay in the auction1996, general auction theoryOne additional bidder is worth more to a seller than all of its bargaining power.2023, theory applied to private exchangesA publisher may be better off not opening one, even if it were free to do so.2025, the prediction is tested on real dataAuction impressions from publishers running both channels areof significantly lower quality than those from single-channel publishers.And nobody, anywhere, compares the return the advertiser gets across the two channels.
The prediction is decades old. The empirical confirmation is one year old. Source : Bulow and Klemperer, American Economic Review 86(1), 1996; Choi and Sayedi, Marketing Science 42(3), 2023; Balocco, Lu, Li and Gupta, Management Science, 2025; Choi, Mela, Balseiro and Leary, Information Systems Research 31(2), 2020 (2025)

The bundle, and why it prices badly

The study offers a financial comparison to explain the mechanism, and it is too apt to paraphrase:

“CDOs typically bundle good or ‘prime’ mortgages with a low risk of default alongside subprime mortgages with a higher risk of default. If bond rating agencies assign high ratings like AA to these CDOs, which masks the underlying risk closer to a B- rating, then buyers will likely misprice the investment. […] Similarly, if the average PMP mixes high-quality premium inventory with low-quality inventory (including MFA sites) to maintain an overall low CPM, it creates a perception of high quality that does not always align with reality.”

A large private marketplace is a bundle. The word “private” functions as a rating: it sets a quality expectation that need not match the contents. And the price forms on the rating.

The recommendation that follows is blunter than you might expect from an advertiser association:

“the days of simply assuming all PMP inventory is worth the premium are behind us. […] With proper checks, controls, and optimizations in place, OMP inventory can drive comparable quality at a lower cost.”

Curation, and the field that omits the number

A layer has settled on top of all this, and it is what composes the large deals: curation. An intermediary, often the supply-side platform itself, bundles inventory from many publishers, enriches it with data, and sells it as a single deal.

Three things are worth knowing.

The reference definition of curation was written by a curation vendor. The most-cited text, hosted by the standards body, is filed by that same site as a member perspective, authored by an executive of a curation company. It is not a specification voted by a working group. The conflict of interest starts at the definition.

Standardized transparency covers the type of fee, never the amount. A dedicated deals interface was finalized in February 2026. It defines a curator role and a curation fee field, and the specification states what that field holds:

“curationfee provides information about the type of fee being applied in the bidstream, but not what that fee is.”

The type, not the number. And “undisclosed” is a permitted value in the schema, not an anomaly. That interface also sits outside the real-time protocol, which the specification says explicitly, so nothing in the bid request itself currently carries curation information.

On actual amounts, the one quantified analysis covers 1.5 billion bid requests and comes from an independent consultancy, whose business admittedly benefits from advertiser distrust of intermediaries. It finds half of curated deals carry no additional curation fee, a third apply a margin-based fee with a 14% median, and 7% use a static pricing model that can exceed half the transaction value. Median effective margin across all curated deals: 30%.

And the institutional gap is documented. The most recent joint framework on programmatic auction transparency, published in January 2026 by the measurement council together with both major advertiser associations and the agency association, contains no occurrence of the words “curator” or “curation” across its nineteen pages. It recommends voluntary annual independent audits for auction operators without ever placing curators in that category.

Scope of standardized curation fee transparency and the actual level of those feesScope of standardized curation fee transparency and the actual level of those fees. A dedicated deals application programming interface, finalized in February two thousand twenty six by the standards body, defines three roles attached to a deal identifier, namely the origin system that receives bid responses, the seller that sold the deal, and the curator, defined as the business entity that did the packaging of inventory, technology or data. It provides a field for the curation fee, and the specification states that this field provides information about the type of fee being applied in the bidstream but not what that fee is, the example given being that a curator charging a five dollar cost per thousand will have the field set to the value corresponding to the cost per thousand type without the amount appearing anywhere. The value corresponding to undisclosed is a value provided for by the specification rather than an anomaly. That interface is moreover separate from the real time bidding protocol, the specification stating explicitly that it is not included in the protocol request or response, so nothing in the bid request itself currently carries curation information. On amounts actually charged, the one quantified analysis available covers one point five billion bid requests and comes from an independent consultancy whose business benefits from advertiser distrust of intermediaries, a bias worth noting. It finds that fifty percent of curated deals carry no additional curation fee, that thirty four percent apply a margin based fee with a median of fourteen percent, that ten percent apply a fixed cost per thousand with a median of forty eight cents, and that seven percent use a static pricing model that can exceed fifty percent of the transaction value, with the median effective margin across all curated deals standing at thirty percent. Finally, the most recent joint framework on programmatic auction transparency, published in January two thousand twenty six by the measurement council together with both major advertiser associations and the agency association, contains no occurrence of the terms curator or curation across its nineteen pages, and recommends voluntary annual independent audits for auction operators without ever placing curators in that category.The type of fee, never the numberWhat the specification provides”provides information about the type of fee being applied in the bidstream,but not what that fee is”What the fees actually are, across 1.5 billion bid requestsNo additional curation fee50%Margin-based, 14% median34%Fixed CPM, $0.48 median10%Static pricing, can exceed half the transaction7%The January 2026 joint auction transparency framework mentions curation exactly zero times.
The field reports the type of fee, never the amount. And undisclosed is a permitted value. Source : IAB Tech Lab, Deals API v1.0, finalized 6 February 2026; Jounce Media analysis, April 2025, across 1.5 billion bid requests; MRC, Digital Advertising Auction Transparency Standards, January 2026 (2026)

What holds in their favour

There is real evidence on the other side, and it concerns something other than inventory quality.

Value distribution is better. A 2020 British study measured that 54% of advertiser spend reaches the publisher in private marketplaces against 49% in the open auction. Its second edition found unattributable spend falling below 1% against 3% overall, with an impression match rate above 70% because deal identifiers make it easier to reconcile data between buy side and sell side.

A regulator confirms it independently. The British competition authority, having asked supply-side platforms directly, reports that the revenue share retained “tended to vary depending on whether the inventory was sold via a private marketplace or an open auction (with a higher revenue share retained for open auctions).”

But the recommendation carries an admission in parentheses. The British study’s advice to invest more in well-curated private marketplaces reads in full:

“Advertisers, agencies, adtech vendors and publishers should consider investing more in well-curated PMPs, given their higher impression match rates and publisher revenues (and, although outside this study, lower risks in fraud, viewability, brand safety and data leakage)”

Two measured benefits, four asserted ones, with a note in the same sentence saying the four were not studied. That is honest of them, and almost never quoted.

Evidence favouring private marketplaces and limitations declared by the studies recommending themEvidence favouring private marketplaces and the limitations declared by the studies that recommend them. On value distribution, a British study from May two thousand twenty measures that fifty four percent of advertiser spend reaches the publisher for private marketplace transactions against forty nine percent for open auction transactions, and that video returns sixty five percent against fifty four percent for display. Its second edition, published in January two thousand twenty three and covering one point three billion impressions, measures that unattributable spend falls below one percent for private marketplaces against three percent overall, and that those transactions, comprising approximately one fifth of matchable impressions, achieve a match rate above seventy percent against fifty eight percent overall, partly because deal identifiers facilitate impression matching between the demand side platform and the supply side platform. A British competition regulator, having questioned supply side platforms directly, reports that the scale of the revenue share retained tended to vary depending on whether the inventory was sold via a private marketplace or an open auction, with a higher revenue share retained for open auctions. However, the British recommendation in favour of greater investment in well curated private marketplaces rests on two benefits actually measured, namely higher impression match rates and higher publisher revenues, and mentions four further benefits, namely lower risks in fraud, viewability, brand safety and data leakage, while stating explicitly in parentheses that these four are outside the scope of that study. A separate measurement gap remains, since all available literature reasons from the publisher’s profit and no institutional or academic source compares the return obtained by the advertiser across the two channels holding brand, period, audience and creative equal.What argues in their favourWhat is measuredShare of spend reaching the publisher54% against 49%Unattributable spendunder 1% against 3%Impressions matched buy side to sell sideover 70% against 58%What a regulator confirms separatelySupply-side platforms retain a higher revenue share on open auctions.What the recommendation admits it did not measureFraud, viewability, brand safety, data leakage: “although outside this study”.
Two measured benefits, four asserted. The parenthesis belongs to the study itself. Source : ISBA and PwC, Programmatic Supply Chain Transparency Study, May 2020, page 9, and Study II, 18 January 2023, pages 7 and 11; Competition and Markets Authority, Online Platforms and Digital Advertising, Appendix R (2023)

And a gap remains that nobody fills. All of this literature reasons from the publisher’s profit. No institutional or academic source compares the return the advertiser gets across the two channels on equal terms: same brand, same period, same audience, same creative. Not the advertiser associations, not the regulators, not the research. What exists on the advertiser’s side of that question is sales material.

What to do

Ask how many domains the deal contains. That single question separates 0.02% junk inventory from 19.8%. It fits in one line of an email, and it discriminates better than every use of the word “premium” in a deck.

Do not pay the premium for a bundle. A deal aggregating hundreds of domains delivers, on average, open-auction quality. Paying double for it means paying for the label.

Separate the argument that holds from the one that does not. These deals return more to the publisher and reconcile better: that is measured, real, and a good reason to use them if your objective is a publisher relationship. That they are cleaner or more effective is not measured, and the study recommending them says so itself in parentheses.

Ask for the amount of the curation fee, not its type. The standard gives you only the type. The amount has to be requested, and the one available analysis shows a minority of deals carrying margins above half the transaction.

And keep in mind why inventory sorts this way. A publisher knows its inventory better than you do. Theory has predicted for decades that the best of it goes to the negotiated channel, and the empirical test arrived in 2025. That does not make deals useless. It means what you are buying depends entirely on who assembled the bundle, and that knowing its contents is the only protection available. That constraint is why our B2B paid acquisition work keeps a B2B budget on search and on the social platforms, where the inventory can be enumerated, rather than on anything sold by label.