How Retail Media Auctions Work: Bid Mechanics, Floor Pricing, and Automated Auction Dynamics (2026)

Last updated: August 2026

Reviewed by Kunal Damgude, Growth and Product Marketing Manager.

Retail media auctions are sealed-bid, first-party-data-weighted competitions that clear ad inventory inside a retailer's own environment, combining bid amount, predicted relevance and shopper-level purchase signal into a single ranking score before a winner and a clearing price are set. The mechanics that decide how much revenue that auction returns to the retailer are narrow and specific: which model clears the slot, how the floor is set and what it is bound to, and how dense the bid landscape is at the moment the auction runs. For the strategic overview of how retail media automates at scale, including the full side-by-side against open programmatic display, see our pillar guide. This article goes a level deeper, into the methodology layer underneath it.

The money behind those mechanics keeps growing. Global retail media spending will exceed $300 billion by 2030, up from $184 billion in 2025 (Forrester). In the US, advertisers spent $60.32 billion on retail media in 2025 and will spend $71.09 billion in 2026, according to a December 2025 EMARKETER forecast. Auction design is what decides how much of that a given retailer captures rather than leaves at the floor.

This is a methodology-first breakdown for retailers running their own network, ad tech builders and marketplace monetization leads, not a guide to buying ads on someone else's network. We assume you know what automated bidding is; if not, our companion piece on why automated bidding transforms retail media performance covers that ground, and for the platform-by-platform view of the science of scalable auction automation see our companion explainer. What follows is auction-type mechanics, floor pricing theory, automated bidding algorithms, the MRC 2026 transparency requirements, and what happens to the auction itself when a retailer activates the advertisers it does not currently serve.

How Retail Media Auctions Differ from Open-Web Programmatic RTB

A retail media auction differs from an open-web programmatic auction in three structural ways: inventory control, signal determinism and measurement closure. The retailer owns the inventory, the auction engine and the shopper identity, so there is no supply-side platform in the middle, no probabilistic identity match, and no bid request fanning out to dozens of unrelated bidders. The signal substrate is deterministic rather than inferred: a logged-in shopper with known purchase history, loyalty status, basket composition and category-level intent. And because one system records both the impression and the transacted SKU, the feedback that trains a bidding model arrives in hours rather than days, which is why auction-level optimization in retail media converges faster than in open RTB.

Definition. The bid landscape is the distribution of all live, eligible bids on one inventory unit at the instant the auction runs. Nearly everything in this article is downstream of it.

The walls are not sealed. The IAB Tech Lab's 2025 OpenRTB updates introduced a prodfeed object for product listing advertisements, standardizing how retail media inventory can be exposed to programmatic infrastructure (PPC Land, 2025). The walled garden is not breaking down; it is being fitted with standardized doors. The complete comparison against open programmatic display, including the demand-side pressure behind that convergence, lives in our pillar guide on retail media auctions and automation. The rest of this article assumes it and moves on to the mechanics.

Auction Type Mechanics: First-Price, Second-Price, Vickrey, GSP, and Proprietary Hybrids

Retail media auctions run on one of four mathematical models. Most retailers run only one, but reasoning about all four is what lets you judge which is correct for a given inventory mix. The MRC 2026 transparency standards require a platform to disclose which model it uses, a requirement that surfaces how much variance exists across networks today.

First-Price Auction

In a first-price sealed-bid auction, every bidder submits a bid, the highest bid wins, and the winner pays exactly what they bid. Worked example with three bidders (Avenga, 2025):

• Bidder A bids $4.00
• Bidder B bids $4.50
• Bidder C bids $4.20

Bidder B wins and pays $4.50. The mechanism is simple, but it creates a strategic problem: because the winner pays their bid, every advertiser is incentivized to shade their bid below their true valuation to avoid overpaying. Bid shading reduces revenue for the platform and forces advertisers to maintain shading models, a form of optimization overhead with no downstream benefit. First-price auctions are common in display header bidding and increasingly rare in retail media, because they penalize the unsophisticated advertisers a retailer most wants to keep.

Second-Price (Vickrey) Auction

The Vickrey auction, formalized by William Vickrey in 1961, fixes the shading problem. Bidders submit sealed bids, the highest bidder wins, and the winner pays one cent above the second-highest bid (Wikipedia: Vickrey Auction). Same bids as before:

• Bidder A bids $4.00
• Bidder B bids $4.50
• Bidder C bids $4.20

Bidder B wins and pays $4.21, which is Bidder C's $4.20 plus $0.01 (Avenga, 2025). Bidder B saves $0.29 against the first-price outcome and, critically, has no incentive to shade. The dominant strategy in a Vickrey auction is to bid true valuation, which makes optimization simpler for advertisers and produces a cleaner price signal for the retailer.

The revenue equivalence theorem proves that under uniform value distributions and risk-neutral bidders, first-price and second-price auctions generate identical expected revenue (Wikipedia: Vickrey Auction). In practice the assumptions do not hold cleanly: bidders are risk-averse, value distributions are not uniform, and second-price tends to produce lower CPCs in steady state. When Walmart moved from first-price to second-price in 2022, advertiser spend held steady at plus 4% quarter over quarter while clicks rose 134% and CPCs fell 55% (Tinuiti, February 2025). That transition is still the canonical published benchmark for a second-price move in retail media, and it is Walmart's own result rather than an industry average. Target's Roundel made the same move for product ads on 4 February 2025, with no comparable published outcome data.

Generalized Second-Price (GSP) Auction

The Generalized Second-Price auction, formalized by Edelman, Ostrovsky and Schwarz in 2007, extends second-price logic to multi-slot inventory (Wikipedia: Generalized Second-Price Auction). Search results, sponsored product carousels and any unit that allocates several slots at once need a multi-slot mechanism. GSP ranks bids in descending order, allocates slots in rank order, and prices each slot at the bid of the next-ranked bidder.

Worked example from the foundational GSP literature: two slots with click-through rates of 1.0 and 0.4, and three bidders with valuations of 7, 6 and 1 who bid 7, 6 and 1.

• Bidder 1 wins slot 1 at price 6, which is Bidder 2's bid
• Bidder 2 wins slot 2 at price 1, which is Bidder 3's bid
• Bidder 1's utility: 1.0 x (7 - 6) = 1
• Bidder 2's utility: 0.4 x (6 - 1) = 2.0

The property that matters is that truth-telling is not a dominant strategy in GSP. Bidder 1 could shade to 5, drop to slot 2, pay only 1, and earn 0.4 x (7 - 1) = 2.4, better than the truthful payoff of 1. This is why GSP, despite looking like a multi-slot extension of Vickrey, behaves strategically more like first-price. Sophisticated advertisers shade, unsophisticated advertisers do not, and the platform captures the difference. For a retailer, that gap is a fairness problem as much as a revenue one.

Proprietary Quality-Weighted Hybrid

Most retail media networks do not run a pure first-price, second-price or GSP auction. They run a quality-weighted hybrid, ranking not on bid alone but on Ad Rank. The quality score is a composite of predicted click-through rate, ad relevance, advertiser history and, in retail media specifically, first-party shopper signal.

Definition. Ad Rank is the quality score multiplied by the bid. It is the number the auction actually sorts on, and it is the reason the highest bid does not reliably win.

Worked illustration, three bidders competing for one slot:

The lowest bid wins because its relevance score is highest, and the highest bid ranks last. That inversion is the entire point of quality weighting: it prices relevance rather than willingness to pay, which is what stops a busy auction from degrading the shopper experience as bid pressure rises. The clearing price in a quality-weighted hybrid usually follows second-price logic, so the winner pays the minimum bid needed to hold their rank given everyone else's quality scores. This is the model most retail media networks run today, and the model the MRC standards require a platform to disclose by name.

Bid Ranking and Clearance Price Determination

Knowing the auction type tells you the formula. The bid landscape tells you the inputs. In a steady-state retail media auction:

1. Bids enter. Every eligible advertiser's bid is collected within milliseconds of the impression opportunity firing.
2. Quality scores resolve. Each bid is multiplied by its quality score, itself a product of predicted click-through rate, relevance and first-party signal weight, to produce an Ad Rank.
3. Floor checks apply. Bids below the hard floor are discarded. Bids between the soft floor and the hard floor are evaluated under modified rules, covered in the next section.
4. Allocation determines winners. Single-slot inventory awards to the highest Ad Rank. Multi-slot inventory uses GSP-style rank-order allocation.
5. Clearing price calculates. Under second-price logic the winner pays the minimum bid that would have held their rank, which is typically the next-ranked competitor's effective bid plus one cent, or the hard floor, whichever is higher.

The distinction between winner determination and clearing price confuses advertisers new to second-price mechanics. Winning does not mean paying your bid. Winning means you out-ranked everyone else, and the price you pay is whatever it took to beat the runner-up, which is almost always less than your bid. Building that distinction into advertiser-facing reporting, showing the submitted bid, the effective bid after quality-score adjustment, and the actual clearing price, is one of the highest-trust moves a retail media network can make, and step 5 of the MRC checklist later in this article makes it close to mandatory.

Budget pacing affects clearance prices too, even though it is not a direct auction input. Pacing throttles effective bid rates through the day to stop campaigns exhausting daily budgets in the first morning hour, so the effective bid is often below the configured bid, with the gap depending on the pacing strategy. The literature classifies pacing into throttling, PID controllers, MPC controllers and online adaptive control (Chen, 2025), with simpler heuristics deployed for lighter workloads. The clearance price you observe is the auction outcome after pacing has already trimmed the bid.

Floor Pricing Mechanics: Hard Floors, Soft Floors, Dynamic Floors, and Yield Management

Floor pricing is where most retail media networks leave money on the table, not because they fail to set floors but because they set them statically, generically, and without binding them to the rest of the auction. A floor is only revenue infrastructure when it is explicitly tied to the auction type chosen, to first-party signals as inputs, and to retail-inventory yield curves as the optimization surface. Each binding is worked through below.

Hard Floors and the Fill-Rate Tradeoff

A hard floor is the absolute minimum price the platform will accept for an impression. Bids below it are discarded entirely and the slot goes unfilled if nothing clears (Avenga, 2025). The yield tradeoff is structural: a higher hard floor raises unit revenue when the slot fills and lowers fill rate, the fraction of slots that fill at all.

Definition. The yield curve for an inventory unit is total revenue equals fill rate multiplied by clearing price. The floor level that maximizes that product depends on the bid density of the inventory, which is why the same floor number is correct on one placement and wrong on the next.

Binding to auction type. A hard floor of $5.00 behaves differently under first-price and second-price mechanics. Consider three bids: $4.50, $5.50 and $5.20.

• First-price: the $4.50 bid is discarded as below floor. Bidder B wins at $5.50, paying the full bid. Clearing price is $5.50.
• Second-price: the $4.50 bid is discarded. Bidder B wins, but the clearing price is the greater of the second-highest bid plus one cent and the hard floor, so the greater of $5.21 and $5.00, which is $5.21. Bidder B saves $0.29.

Now move the floor to $5.30:

• First-price: Bidder B still wins at $5.50. Nothing changes.
• Second-price: the greater of $5.21 and $5.30 is $5.30. The floor is now binding, Bidder B pays $5.30 instead of $5.21, and the retailer captured an extra $0.09.

Under second-price the hard floor is doing more work than under first-price: it acts as a synthetic second bid when the actual second bid is weak. Under first-price the floor only matters when bids fall below it; under second-price it competes with the runner-up at every clearing event. The same number is a different lever depending on the model. Hold on to that, because it is the mechanism the auction-liquidity section later in this article turns on.

Soft Floors

A soft floor is a shadow minimum that affects the clearance price calculation without rejecting bids that fall slightly below it. As Avenga frames it, the soft floor was "implemented to 'catch' the offers that fall only slightly below the hard floor and would otherwise get rejected with no yield for the publisher" (Avenga, 2025). The Yale and Cowles Foundation paper by Bergemann and co-authors gives the formal treatment: "A soft-floor auction asks bidders to accept an opening price to participate in an ascending auction. If no bidder accepts, lower bids are considered using first-price rules" (Bergemann et al., 2025).

Bergemann's contribution is to show that soft floors improve efficiency by allowing a lower hard reserve price, cutting the frequency of no-sale outcomes while still capturing the price-shading benefit of an opening floor. For a retailer, that means the right floor architecture is rarely one number set and forgotten. It is a two-level system: a soft floor near the median of the historical bid distribution, and a hard floor at the minimum acceptable rate for inventory of that quality tier.

Dynamic Floors Bound to Shopper Signal

Dynamic floor pricing replaces static numbers with algorithmic adjustment based on observed bid density, time-of-day patterns, inventory scarcity and, uniquely in retail media, first-party signal strength. This is where retail media auctions diverge most sharply from generic supply-side floor pricing.

Binding to first-party signals. In open-web programmatic, dynamic floors take inputs like geography, device, viewability and contextual category. In retail media the dominant input is the shopper signal itself. Consider two impression opportunities for the same product placement:

• Impression 1: a logged-in shopper with a 90-day purchase history in major appliances, currently browsing the appliances category, with one item already in the cart.
• Impression 2: an anonymous browser, first session, on the homepage.

These are not the same inventory unit even though the placement is identical. The signal hierarchy, explicit purchase data above implicit browsing behavior above demographic proxies, is the dominant input to relevance scoring at auction time. A dynamic floor for impression 1 might land at $0.80 because bid density is high, with several appliance brands competing for in-market shoppers, while the floor for impression 2 lands at $0.20 because bid density is low. Same placement, different floors, because the first-party signal value is different.

Binding to retail-inventory yield curves. Retail inventory has yield dynamics that do not exist in open RTB. Sponsored product slots on a high-traffic search results page have a different curve from display banners on a category page, and in-store digital screens have a different curve again. Floor optimization is therefore not a single number but a per-placement, per-category, per-time-of-day matrix in which every cell has its own optimum, computed in continuous-update mode from historical bid density and conversion outcomes.

For mid-market retailers without in-house auction engineering, that has historically meant choosing between a static-floor baseline and building custom yield infrastructure. Osmos closes that gap inside the auction layer itself: hard, soft and dynamic floors are configurable per placement, per category and per first-party signal tier across a retailer's onsite, offsite and in-store retail media, with no custom auction-engine work required. For the deeper context on how first-party data feeds these mechanisms, see our guide on how first-party data enriches ad targeting.

Unified Pricing Rules

Unified pricing rules require a platform to apply identical floor policies to all buyers regardless of demand source or buying entity. The MRC 2026 standard requires this explicitly: "Publishers and SSPs must reveal use of reserve prices or pricing floors and apply identical floors to all buyers" (AdExchanger, February 2026).

The recent regulatory case makes the stakes explicit. Google removed unified pricing rules from Ad Manager in December 2025 following antitrust enforcement, after a 2.95 billion euro ($3.45 billion) European Commission fine for self-preferencing in ad tech (Search Engine Land, December 2025). Publishers can now set bidder-specific floors, requiring one buyer to bid $5 while others compete at $2, inverting the prior decade's direction of travel. For a retailer the lesson is twofold: floor consistency is now a regulatory expectation rather than a stylistic choice, and floor transparency is table stakes. Some estimates put publisher revenue lost to opaque auction mechanics at 15% to 30%, though that figure is a vendor-blog estimate rather than an audited one (MonetizeMore, February 2026).

Automated Bidding Theory: Algorithms, Feedback Loops, and Auction-Type Interactions

Automated bidding is the algorithmic substrate that learns the relationship between a bid and an outcome from auction win and loss data, then updates bid multipliers to hit a target KPI. We do not re-explain its rationale here; see our companion piece on why automated bidding transforms retail media performance. What follows is the algorithmic layer a retailer has to reason about when it builds or buys one.

Target KPI Types

Automated bidding systems optimize against one of four primary KPIs:

• Target CPC. Hold average cost per click at a configured ceiling. Suited to traffic-driving campaigns with consistent conversion economics.
• Target ROAS. Optimize the ratio of revenue to spend. Suited to retail campaigns with measurable basket-level outcomes, and the dominant KPI in retail media because closed-loop attribution makes the signal reliable.
• Target impression share. Hold a configured fraction of eligible impressions. Suited to brand defense and category-share campaigns.
• Target CVR or CPA. Optimize conversion volume at a target cost per conversion. Suited to funnel-bottom campaigns with deterministic conversion events.

Choosing the KPI is choosing the optimization surface. Target ROAS, for instance, needs conversion-revenue feedback inside a window short enough for the model to learn, typically 30 days in retail media, which is also the default attribution lookback in IAB Europe's Commerce Media Measurement Standards (ExchangeWire, January 2026). That standard is now at V2.1, and the compliance grace period for earlier versions ran to the end of July 2026 and has closed.

Feedback Loop Mechanics

The automated bidding loop runs continuously: submit a bid, observe the auction outcome, and on a win observe the click and the conversion, then update the model and adjust the bid multiplier before the next auction. The cycle repeats per impression opportunity with the model updating in micro-batches. A 24-author survey spanning major industry research labs and academic institutions describes this as the canonical bidding-and-auction feedback substrate (Aggarwal et al., 2024).

ML Approaches

Three methodologies dominate deployed retail media bidding systems:

1. Contextual bandits. A multi-armed bandit framework in which each arm is a bid level and the context is the impression opportunity: shopper signal, placement, time. Contextual bandits remain the dominant deployed approach because they are tractable, interpretable and need fewer training samples than deeper reinforcement-learning models.
2. Deep Q-learning networks. A reinforcement-learning approach in which the bid policy is learned as a Q-function over states, meaning impression contexts, and actions, meaning bid levels. These handle larger state spaces but need more training data and can be unstable in production.
3. Actor-critic architectures. A two-network framework in which the actor proposes bids and the critic evaluates them. Actor-critic methods generalize well across reward structures, whether CPC, ROAS or mixed, but are operationally heavier.

Transformer-based architectures for bid prediction are emerging, though contextual bandits still dominate live deployments. The practical difficulties, which is where production engineering effort concentrates, are reward shaping, delayed feedback and counterfactual estimation.

Budget Pacing Strategies

Pacing is the layer between configured bids and effective bids. The dedicated pacing literature (Chen, 2025) classifies algorithms into four families:

• Throttling. Skip a fraction of eligible auctions to slow spend.
• PID controllers. Adjust bid multipliers on proportional, integral and derivative error against target spend.
• MPC controllers. Model-predictive control optimizes pacing across a forecast horizon.
• Online adaptive optimal control. Continuous re-estimation of optimal pacing under non-stationary demand.

Simpler heuristics, including spend-as-fast-as-possible, front-loading, even pacing and modified even pacing, are approximations used where campaign volume does not justify a controller.

Bidding Under First vs Second Price

This is where the algorithmic layer couples back to the auction type. Bid shading is the primary optimization lever under first-price mechanics and meaningless under second-price, where the dominant strategy is bidding true valuation. That is why Walmart's 2022 move to second-price shows up as a steep CPC decline rather than as a spend decline: first-price-trained shading models stopped being necessary, and the runner-up bid rather than the winner's own bid became the price. Automated systems trained on first-price data overshoot under second-price; systems trained on second-price data undershoot under first-price. An auction-type change is a model retraining event, and it is one the retailer schedules, not the advertiser, which makes advance notice part of the change plan rather than a courtesy.

Data Volume Threshold

Automated bidding works best at scale. The widely-cited industry threshold for reliable machine-learned bidding is 30 or more conversions per month per campaign, and portfolio bidding typically needs 50 conversions per 30 days. Mid-market and long-tail advertisers routinely sit below both, which means generic automated bidding fails them exactly where a retailer most wants incremental demand.

This is the hinge the last third of this article turns on. The threshold is a property of the model, not of the advertiser, and the 2026 generation of campaign tooling attacks it by assembling the campaign from a catalog and a stated objective rather than from historical performance. The retailer-side consequence, a denser auction, is worked through under auction liquidity below. Osmos builds its automation to run under those thresholds rather than to wait for advertisers to cross them, which is the difference between serving the tail and shortlisting it.

A/B Testing the Switch to Automation

Switching a campaign from manual to automated bidding without data-driven validation is one of the most common failure modes in retail media. The general protocol, drawn from standard search-advertising testing practice and applicable here:

1. Split traffic 50/50. Half the campaign's eligible impressions go to the manual configuration and half to the automated one, with creative, targeting and budget pacing held constant.
2. Run for two to four weeks minimum. Shorter windows produce noise-dominated results; longer windows risk seasonal contamination.
3. Require 30 or more conversions before declaring a winner. Below that, the confidence interval on observed performance is too wide to conclude anything.
4. Instrument for KPI continuity. If the manual campaign was optimized for CPC and the automated campaign defaults to ROAS, the comparison is invalid. Set both to the same KPI first.
5. Re-test after platform changes. Auction-type switches, quality-score model updates and pacing-algorithm changes all invalidate prior tests.

This applies whether you are switching a single campaign or rolling out platform-wide automation defaults, and it is the only defensible answer when an advertiser asks why their CPCs moved.

Platform Comparison Matrix: How Auction Mechanics Differ in Practice

Methodology becomes operating guidance only when it lands against real platform configurations. The matrix below is the only place in this article where vendor names appear, and they appear as column headers in a methodology-as-rows by platform-as-columns grid rather than as section anchors. Osmos occupies the first column because the framing throughout is from the retailer's side of the auction, and Osmos is the independent mid-market option in the comparison set.

[1] Amazon does not officially confirm a hard floor. The $0.02 minimum CPC is widely cited across third-party sources, while Amazon's own public statement is that CPC is determined by an ad's ranking as well as the ranking of other related brands and products.

[2] Google removed unified pricing rules from Ad Manager in December 2025 following a DOJ antitrust ruling and a 2.95 billion euro European Commission fine.

Where mechanics converge. Every major retail media platform in the matrix uses quality-score weighting, and the trend toward second-price is universal among platforms that have publicly disclosed a transition: Walmart in 2022, Target's Roundel for product ads in February 2025 (Tinuiti, February 2025). First-party data integration is table stakes; no major platform now competes on a third-party-cookie substrate.

Where mechanics diverge. Floor pricing transparency varies sharply. Some platforms publish hard-floor numbers and others treat floor logic as proprietary. MRC participation is the most visible 2026 divergence: Google and Amazon, both dominant ad-buying platforms that also supply publisher inventory and have faced criticism over auction transparency, were absent from the standards' development (AdExchanger, February 2026). Mid-market retailers have a structural transparency advantage available here, and the format-level view of how those auctions present to a shopper is covered in our guide to sponsored and display ad formats in retail media.

First-Party Data as an Auction Signal

This section is about scoring mechanics rather than the post-cookie landscape, which the pillar guide covers. The mechanic is straightforward: a shopper's first-party identifier resolves to a profile covering purchase history, loyalty tier, browsing intent and basket state; that profile produces a feature vector at impression time; the feature vector enters the quality-score model; and the resulting quality score multiplies the bid to produce Ad Rank. Every step in that chain is now disclosable under the MRC feature-importance requirement.

The signal hierarchy is well established in retail media practice:

1. Explicit purchase data carries the highest weight. A shopper who bought running shoes 14 days ago is a different inventory unit from one who has never bought athletic apparel.
2. Implicit browsing behavior comes next: recent category-page views, search queries and basket-add events without checkout.
3. Demographic proxies carry the lowest weight, inferred from sparse signals and used only when better data is unavailable.

The consequence for auction design is that a first-party signal is not a targeting filter bolted on beside the auction. It is a multiplier inside it. Two advertisers submitting identical bids on identical placements clear at different prices and in different positions purely because one is matched against a shopper the model can predict. The market is moving accordingly: 71% of brands, agencies and publishers are currently growing or planning to grow their first-party data sets, nearly double the rate of two years earlier (AdExchanger, April 2026). Alvaro Palacios, Chief Strategy Officer at Newsweek, put the structural argument plainly: "AI decision engines optimized for outcomes (sales, retention, lift) require deterministic identity, clean feedback loops and governable data lineage."

Privacy-Preserving Auction Mechanics

The on-device bidding framing common in 2023 and 2024 industry coverage is out of date. Google retired its Protected Audience API, formerly FLEDGE and part of the Privacy Sandbox initiative, in October 2025 after low industry adoption. On-device bidding via Privacy Sandbox is no longer the privacy-preserving route for retail media.

The current path is consent-based first-party data plus clean rooms. Consent-based data flows directly from logged-in retailer relationships, which is the structural advantage retail media networks already hold, and clean rooms provide a controlled environment for combining a retailer's data with a brand's without exposing raw identifiers. Retailers that route first-party signals through that infrastructure and can already say which variables moved an auction outcome are ahead of the MRC feature-importance requirement rather than behind it.

Building a Better Auction: 2026 Best Practices and the MRC Transparency Standard

"Digital advertising runs on auctions, but buyers still can't typically see the rules." Ben Hovaness, Global Chief Media Officer, OMD (MRC, January 2026)

The Media Rating Council issued the final version of its Digital Advertising Auction Transparency Standards on 29 January 2026, ratifying a draft released in September 2025 (MediaPost, January 2026). Sponsored by the ANA, 4A's, WFA and IAB Tech Lab, and initiated by Omnicom Media, the standards cover display, text, video and audio formats across digital, search, social, retail media, streaming CTV and addressable TV. Retail media is explicitly in scope, and adoption is voluntary, applied through MRC accreditation audits rather than through regulation. Six months on, the framework's status is unchanged: published, voluntary, enforced by audit, with no named retail media network yet announcing accreditation.

What the MRC Standard Requires

The standard imposes seven core disclosure requirements, and retailers should treat them as the 2026 best-practice checklist whether or not they pursue accreditation:

1. Auction type disclosure. State publicly whether the auction is first-price, second-price or modified second-price.
2. Winner determination methodology. Disclose the exact procedure for selecting auction winners.
3. Clearing price derivation. Disclose how the price the winner pays is calculated from the bids submitted.
4. Reserve price and floor pricing disclosure, applied uniformly to all buyers (AdExchanger, February 2026).
5. Nominal to effective bid conversion. Disclose technical fees, bid multipliers and relevance scores before a bid competes (Datawrkz, February 2026).
6. Model cards for machine-learned auctions. Publish training data, intended use and performance metrics for any model used in bid evaluation.
7. Feature importance disclosure. Identify which variables, including first-party signals, influenced auction outcomes.

IAB Europe's Commerce Media Measurement Standards run in parallel on the measurement layer, setting a 30-day attribution lookback as the default alongside machine-readable line-item reporting (ExchangeWire, January 2026). Between them, MRC and IAB Europe make 2026 the year retail media auction and measurement disclosure stopped being optional in practice even where it remains optional on paper.

As MonetizeMore's chief executive Kean Graham framed the implication for advertisers: "If your current partners cannot or will not explain exactly how your money is being made, they are optimizing for themselves, not you" (MonetizeMore, February 2026). With the dominant ad-buying platforms absent from the standards' development, a mid-market retailer that can answer all seven questions has a differentiator the largest platforms have not published.

2026 Best Practices for Retailers

Translating the standards into operating practice:

• Choose the auction type by inventory characteristics, not by market trend. Second-price has strong evidence in product listing and sponsored placements, quality-weighted hybrids dominate search-driven inventory, and first-price still has a place in guaranteed display where shading models are mature.
• Set floors from yield-curve analysis, not arbitrary minimums. Compute the optimal floor per placement, category and time-of-day cell from historical bid density, and bind floor levels to first-party signal strength.
• Automate bidding where data volume supports it and provide fallbacks where it does not. Advertisers under 30 conversions a month need automation built for that regime, or they churn before they ever become a revenue line.
• Expose auction mechanics to advertisers. Auction type, floor configuration, bid multipliers, fee structure. The MRC standard expects it and advertiser trust requires it.
• Build for line-item reporting in 24 to 48 hours. The closed-loop attribution stack that pairs with this is covered in our guide on closing the attribution loop in retail media.

Agentic Bidding and OpenRTB's Limits

Looking ahead, AI agents managing spend portfolios across publishers may begin to displace bid-by-bid optimization for sophisticated advertisers. The emerging plumbing for that is the Ad Context Protocol, an open-source communication protocol that lets AI agents built by advertisers, publishers or intermediaries interact in a common language. Digiday's summary is the clearest one available: "Think of AdCP as the OpenRTB for the AI era" (Digiday, October 2025). Worth noting for anyone about to over-plan around it: the founding and supporting members announced so far are publishers, supply platforms and measurement firms, and not one is a retail media network. Agentic allocation is early-stage and directionally significant rather than imminent. The retailers that build for agentic compatibility, meaning machine-readable auction logic, model cards and feature-importance disclosure, are the ones that will integrate cleanly when it matures, and every one of those three is already on the MRC checklist above.

Auction Liquidity: What Tail-and-Torso Activation Does to Bid Density and Floor Clearance

Everything above treats the bid landscape as given. It is not. How many advertisers can reach a retailer's auction is a decision the retailer makes, and it is the largest single lever on the yield curve defined earlier. It also matters most to the smallest networks: Amazon alone is forecast to pass $75 billion in retail media revenue by 2028, more than $65 billion ahead of the next-largest network (EMARKETER), and a handful of networks continues to absorb the large majority of incremental spend. An independent retailer will not win that fight by chasing the endemic budgets everyone else is chasing. The advertisers it can win are the ones nobody is serving well, and what follows is what winning them does to the auction rather than to the software.

Why Tail Advertisers Stay Dormant

Most retail media networks run a thinner auction than their own advertiser base would support. Mid- and long-tail advertisers already generate 28% of revenue at the average network, on Forrester's numbers (Retail Media Breakfast Club, November 2025), and more than 85% at the most advanced networks. In the same research, retailers ranked the inability to manage the long tail of advertisers and marketplace sellers fourth among their scaling challenges. The constraint has never been demand. It is that an incremental advertiser costs roughly as much to onboard and support as a large one while contributing a fraction of the revenue.

Four launches across late 2025 and 2026 changed the tooling side of that equation, and all four point the same way:

• Amazon Ads Agent, introduced at unBoxed in November 2025, "works alongside advertisers to automate time-consuming tasks, like identifying targeting segments, adjusting pacing across hundreds of campaigns, and generating SQL queries for advanced analytics," and will "create your campaign structure and ad groups for you" from an uploaded media plan (Amazon Ads). Amazon's own page frames it around scale, not inexperience. The Q1 2026 beta timing, the reading that it levels the field between small sellers and large brands, and a reported average 14% cut in advertising cost of sale within 30 days all come from trade reporting rather than from Amazon (NovaData, December 2025). Treat that last figure accordingly.
• Walmart Connect's conversational assistant, in beta, answers bidding, keyword and billing questions and issues account-specific alerts. Of the traffic it has seen, 97% of user queries are unique, which suggests advertisers are asking about their own accounts rather than repeating generic questions (PPC Land, January 2026).
• Criteo GO opened to all small and mid-sized businesses and growth-stage commerce brands in the US and UK from 31 March 2026, letting an advertiser open an account, enter billing details and put a live campaign in market with no managed-service agreement (PPC Land, April 2026).
• Instacart's Ads Manager extended to retail partners in May 2026, so retailers can build basket-level offers and target high-intent segments inside the same self-serve surface their advertisers already use (Instacart).

Osmos reads those four the same way, and the reading is ours rather than theirs: none of the four companies describes its launch in these terms. What they have in common is that the campaign gets assembled from a catalog and a stated objective rather than from conversion history. If that generalizes, the binding constraint on tail activation moves from "does this advertiser have enough data for automation to work" to "can the retailer absorb the onboarding and support load at scale": a different problem, owned by a different team, with a different fix. For the tooling and workflow side of it, see our companion piece on one-click campaign creation and AI-assisted optimization tooling.

Bid Density and Floor Clearance

The floor pricing section defined the yield curve as fill rate multiplied by clearing price, and noted that the optimal floor depends on bid density. Tail activation is the lever that moves bid density, which makes it a yield-management decision before it is a demand-generation one.

Take a category-page display slot with a $0.60 hard floor and a thin auction: three eligible bidders per opportunity, of whom one typically clears the floor. Roughly 40% of opportunities go unfilled, and on the ones that do fill the clearing price sits at $0.60, because under second-price logic the floor is the only thing standing in for a runner-up. The retailer is not being paid what the impression is worth. It is being paid what it declared the impression to be worth, months ago, in a configuration screen.

Now put twelve eligible bidders into the same auction, five of whom clear the floor. Fill rate climbs toward the high nineties, and the clearing price is set by a real second bid instead of by the floor. Both terms of the yield curve move up together, which is unusual, since most floor decisions trade one against the other. The numbers are illustrative rather than measured; the mechanism is what generalizes, and it is measurable on any live network.

The counter-intuitive consequence is that a denser auction lets a retailer lower its hard floors rather than raise them. In a thin auction the floor is doing price discovery, so it has to be set defensively high, and every basis point of defensiveness costs fill rate. In a dense auction the runner-up does price discovery and the floor reverts to what it should be, a reserve below which the retailer would rather leave the slot empty. Activate the tail and leave defensive floors untouched and you capture the fill-rate gain while handing back the price gain.

Density also arrives unevenly. Tail advertisers bid into the auctions their own catalog reaches, so long-tail SKU and category queries get denser first, which is precisely the inventory with the worst fill rate to begin with. Head terms barely move. The gain concentrates where the inventory was earning nothing, which is where it is worth most, and it will not show up in a network-average CPM report.

One thing gets harder. A bigger auction puts more weight on quality scoring, because more of the bid landscape now comes from advertisers with thin listings and no performance history. If quality weighting is soft, they win rank on bid alone and the shopper experience degrades at exactly the moment revenue improves. Density is only worth having if the Ad Rank calculation described earlier is doing real work, which argues for tightening quality weighting before opening the doors rather than after.

Ad Ops Economics of a Bigger Auction

The cost side is where a retailer should be most skeptical of vendor claims, including ours. No public source gives a per-advertiser servicing cost for retail media. Across the sources reviewed for this update there is no ad operations cost-per-advertiser benchmark to cite, and any figure presented as one is almost certainly a company-wide number divided by an advertiser count. What can be described honestly is the direction and the mechanism.

The mechanism is that managed service has a roughly fixed cost per account. An ad operations specialist can carry some number of advertisers, and that number is set by how much human diagnosis each account needs: reading the report, finding the underperforming campaign, working out whether the fix is a bid, a budget, a keyword or a product, then telling the advertiser. Because that cost is roughly fixed while tail revenue per account is small, there is a revenue line below which an advertiser cannot be served profitably. That line, not advertiser appetite, is what keeps the tail dormant.

AI-assisted campaign creation and diagnosis moves that step off the specialist. It does not remove the rest: billing, policy and creative review, catalog quality, fraud, disputes and support all remain, and several get harder as advertiser count rises. The honest claim is therefore directional. The cost of serving an incremental advertiser falls, and the revenue line below which service is uneconomic falls with it. Cost does not go to zero, and a retailer shown a per-advertiser cost curve should ask which of those workloads it counted.

Retailers already believe this is where the value sits. In Forrester's survey of 160 senior retail executives across North America and EMEA at companies with more than $50 million in online revenue, nearly 75% rated fully automated campaigns and AI or ML-driven targeting valuable or highly valuable and 76% named ease of onboarding sellers for self-service the single most valuable feature, while only 53% were satisfied with what their platform's AI and ML actually delivered. As Retail Media Breakfast Club put it: "Long-tail requires frictionless self-serve onboarding, and automated campaign management." That gap is the opportunity. For the broader onsite monetization play tail activation belongs to, see our guide on monetizing onsite traffic with retail media.

Where Osmos Fits

Osmos is a retail media operating system for retailers and marketplaces, and it runs one auction across all three of a retailer's media channels: onsite retail media, offsite audience-extension retail media, and in-store retail media. Everything described in this article is configuration surface rather than roadmap. Auction type is selectable per inventory class; hard, soft and dynamic floors are set per placement, per category and per first-party signal tier; quality-score weighting, budget pacing and advertiser wallet controls sit alongside them, configurable by the retailer rather than fixed by the vendor.

Two things follow from the argument above. The MRC checklist is a configuration question before it is a compliance question, because a retailer can disclose auction type, winner determination, clearing-price derivation, floor policy, nominal-to-effective bid conversion, model cards and feature importance only if its platform exposes them. Osmos exposes auction type, floor configuration and bid multipliers to advertisers by default, which is what makes the seven-point checklist answerable rather than aspirational.

And liquidity is an operations problem wearing an auction problem's clothes. A retailer that wants a denser auction needs self-serve onboarding, campaign creation that does not require a media plan, and diagnosis that runs before a human reads a report. Osmos runs that layer next to the auction rather than as a separate managed-service motion, which is what lets the cost of serving an incremental advertiser fall instead of scaling with advertiser count. For the ad operations and workflow view of the same layer, see our companion piece on the science of scalable retail media automation.

Frequently Asked Questions

What is the difference between a first-price and second-price auction in retail media?

In a first-price auction the highest bidder wins and pays exactly their bid. In a second-price or Vickrey auction the highest bidder wins but pays one cent above the second-highest bid. First-price therefore rewards bid shading, meaning bidding below true valuation, while second-price makes bidding true valuation the dominant strategy. Most retail media networks moved to advanced second-price over the past five years, and quality-weighted hybrids that combine second-price clearing with quality-score ranking are now the dominant model.

How do retail media networks determine the clearance price after an auction?

Winner determination and pricing are two separate steps, and conflating them is the most common misreading of a second-price auction. The winner is whoever holds the highest Ad Rank once quality scores are applied and sub-floor bids are discarded; the price is then whatever it took to hold that rank, which the body section above works through in full. The practical consequence is that winning does not mean paying your bid. Reporting that shows advertisers all three numbers, submitted bid, effective bid and clearing price, removes most of the disputes this causes and is close to mandatory under the MRC checklist anyway.

What is a good floor price strategy for a mid-market retail media network?

Three components: a hard floor at the minimum acceptable rate per placement category, a soft floor set off the historical bid distribution so near-miss bids are not simply thrown away, and dynamic per-impression adjustment driven by first-party signal strength, so the floor rises for in-market shoppers and basket-add behavior and falls for anonymous browsing. The part most networks skip is re-deriving all three after any change to the advertiser base, since the correct floor is a function of bid density and density moves. Vendor estimates put the revenue left unrealized by static, opaque auction mechanics at 15% to 30%.

How does automated bidding work differently under first-price vs second-price auction mechanics?

Bid shading is the main lever under first-price and meaningless under second-price, so a system trained on one model misprices under the other. The operational point for a retailer is that an auction-type switch is a model retraining event on the advertiser's side that the retailer schedules. Give notice, expect unstable performance while their systems recalibrate, and put the change in writing. Auction-type disclosure is the first requirement on the MRC checklist, so a silent switch is now a compliance issue as well as a trust one.

What are the bidding strategy differences between enterprise and SMB retail media platforms?

Enterprise platforms, meaning the largest integrated retailers and the dominant marketplaces, typically run quality-weighted second-price auctions with proprietary machine-learned bidding that needs 30 to 50 or more conversions per campaign per month to work reliably. Platforms aimed at smaller advertisers need automation that runs below those thresholds, configurable auction types including first-price for new inventory, simpler floor pricing that does not require yield-curve infrastructure, and self-serve dashboards. Because auction logic, reporting taxonomy and API shape differ from network to network, smaller advertisers also lean harder on cross-network tools that abstract those differences away.

What is the minimum data volume needed before switching to automated bidding?

Roughly 30 conversions per campaign per month for reliable automated bidding, and about 50 conversions per 30 days for portfolio bidding. Below that, the model cannot learn a stable relationship between bid and outcome and will often underperform a manual setup, which is why the body section above recommends a 50/50 A/B test of at least two to four weeks and at least 30 conversions before declaring a winner. The threshold is a property of the model, not of the advertiser, so for the tail the answer is automation built for the sub-threshold regime rather than a waiting list. Osmos builds it that way across onsite, offsite and in-store retail media.

How will the MRC 2026 auction transparency standards affect retail media networks?

The MRC Digital Advertising Auction Transparency Standards, finalized on 29 January 2026, ask networks to disclose seven categories of auction mechanics: auction type, winner determination, clearing-price derivation, floor pricing, nominal-to-effective bid conversion, model cards for machine-learned auctions, and feature importance. Retail media is explicitly in scope. Adoption is voluntary and applied through MRC accreditation audits rather than by regulation, so nothing compels a network to comply. With the dominant ad-buying platforms absent from the standards' development, the disclosure gap is a live differentiator for independent networks rather than a box-ticking exercise.

What happens to auction liquidity when a retailer activates its long-tail advertisers?

Fill rate rises because more bids clear the floor, and the clearing price rises because a real runner-up replaces the floor as the price setter. Both effects concentrate on long-tail SKU and category inventory rather than head terms, so a network-average CPM report will understate what happened. To attribute the change rather than assume it, instrument two series before activating anything: eligible bids per auction, and the share of clearing events priced by the floor rather than by a runner-up bid. When the second series falls, the floor has stopped doing price discovery and can usually be lowered. Most retail media reporting stacks track neither by default, which is why the yield gain from activation is so often invisible afterwards.

Does AI-assisted campaign creation reduce Ad Ops cost per advertiser in retail media?

Directionally yes, and no published benchmark exists to size it. No retail media source reviewed for this update carries a per-advertiser servicing cost, so treat any specific dollar figure with suspicion and ask what it counted. Measure it internally instead: fully loaded ad operations cost, including support, billing and policy review, divided by active advertisers, tracked monthly from before activation. Watch the denominator, because a self-serve push raises advertiser count faster than it raises cost and the ratio will improve for a quarter even if nothing structural changed. The number that matters is the revenue level at which an advertiser becomes profitable to serve, and whether it is falling.

What third-party bidding management tools are used for retail media?

Brands and agencies use a mix of platform-native tools, since each large retailer offers its own buyer interface for in-network inventory, and third-party bidding management platforms that abstract across several networks at once. Those tools proliferate wherever an advertiser or agency runs campaigns on many networks with incompatible auction logic and reporting. On the retailer's side of the auction, mid-market retailers building independent networks increasingly adopt full-stack retail media operating systems carrying auction configuration, automated bidding, yield management and advertiser-facing operations in one platform. The comparison matrix above shows how those dimensions vary.

Sources

1. Media Rating Council, Digital Advertising Auction Transparency Standards, Final Release (29 January 2026)
2. Media Rating Council, Digital Advertising Auction Transparency Standards, final document
3. MediaPost, Going, Going, Gone: MRC Finalizes Ad Auction Standards (January 2026)
4. AdExchanger, The MRC Wants Ad Tech To Get Honest About How Auctions Really Work (February 2026)
5. Datawrkz, What the 2026 MRC Digital Advertising Auction Transparency Standards Mean (February 2026)
6. MonetizeMore, Why the MRC's New Auction Standards Are a Wake-Up Call (February 2026)
7. Wikipedia, Vickrey Auction
8. Wikipedia, Generalized Second-Price Auction
9. Bergemann, Breuer, Cramton, Hirsch, Ndiaye and Ockenfels, Soft-Floor Auctions: Harnessing Regret to Improve Efficiency and Revenue (Yale and Cowles Foundation Discussion Paper 2438, April 2025)
10. Aggarwal et al., Auto-bidding and Auctions in Online Advertising: A Survey (arXiv:2408.07685, August 2024)
11. Chen, A Practical Guide to Budget Pacing Algorithms in Digital Advertising (arXiv:2503.06942, March 2025, revised July 2026)
12. Avenga, First-Price vs. Second-Price Auctions in Programmatic Advertising Explained (July 2025)
13. Tinuiti, Roundel Media Studio Adopts Second-Price Auction (February 2025)
14. Search Engine Land, Google Scraps Unified Pricing Rules in Ad Manager After Antitrust Pressure (December 2025)
15. AdExchanger, AI Has Already Decided: First-Party Data Will Define Advertising's Agentic Era (April 2026)
16. PPC Land, Retail Media Networks Embrace RTB for Sponsored Products (July 2025)
17. ExchangeWire, IAB Europe Releases Commerce Media Measurement Standards and Flexi Ad Sizes Guidelines (January 2026)
18. Forrester, Global Retail Media Forecast, 2025 To 2030
19. EMARKETER, FAQ on Retail Media Networks: How Marketers Should Allocate Budgets in 2026
20. EMARKETER, Retail Media Ad Spending Forecast H1 2026 (May 2026)
21. Retail Media Breakfast Club, Long-Tail Advertisers Are a Quiet Growth Engine for Top-Performing RMNs (November 2025)
22. Amazon Ads, Boost Advertising Efficiency With Ads Agent (unBoxed, November 2025)
23. NovaData, Amazon Ads Agent: AI Campaign Management Arrives 2026 (December 2025)
24. PPC Land, Walmart Connect Bets on AI Agents to Reshape Retail Advertising (January 2026)
25. PPC Land, Criteo Opens GO to All SMBs (April 2026)
26. Instacart, Instacart Expands Ads Manager to Retailers (May 2026)
27. Digiday, WTF Is Ad Context Protocol (AdCP)? (October 2025)