Bid shading explained: how DSPs bid in first-price auctions

In a second-price auction, the winner pays the second-highest bid plus a small increment. If you bid $5.00 and the next bid is $2.50, you pay $2.51. Bidding your true value is the best strategy, because your bid only decides whether you win, not what you pay (Gligorijevic et al., CIKM 2020).

In a first-price auction, the winner pays its own bid. The same $5.00 bid now costs $5.00, even though $2.51 would have won. So a demand-side platform (DSP) adds a step after valuation: it estimates how high rival bids are likely to be and bids just above that. Bid too low and you lose impressions you wanted. Bid too high and you overpay. Shading is the search for the best point between the two.

Header bidding forced the change. When several exchanges compete in a publisher's wrapper, a second-price winner arrives at the ad server with a price that was lowered inside its own exchange, which distorts the final comparison. Exchanges moved to first price through 2017 and 2018, and Google's Ad Manager was the last large holdout (PPC Land, 23 August 2026). The auction flow itself is explained step by step in how real-time bidding works.

The April schedule slipped: the release notes show 10% of traffic moving only at the end of July. The September post explains that Google extended the test period after partner feedback (Google, 5 September 2019).

A shading model needs two inputs. The first is the value v the DSP places on the impression, the bid it would make in a second-price auction. The second is a win-rate curve W(b): the probability that a bid b beats the highest rival bid. The win rate curve is the cumulative distribution of the highest competing bid, which is why the literature treats it as a forecast of the distribution of winning prices. The DSP then picks the bid that maximises expected surplus.

Take one ad slot. The DSP values it at $4.00 CPM. Its model says rival bids follow a logistic curve centred at $3.00, the curve shape used in the win-rate method of Pan et al., so a $3.00 bid wins half the time. The table shows what happens across a million bid requests at different bids. Figures are illustrative, computed from that curve.

Bidding the full $4.00 wins the most impressions but leaves no surplus: every won impression costs exactly what it is worth. Shading to $3.00, a factor of 0.75, gives up 43% of those impressions and keeps $500 of value per million requests. The best bid is not the one that wins most often. It is the one that balances the chance of winning against the margin on each win.

The same model explains why the 2019 switch raised prices at first. In a second-price auction with this curve, a $4.00 bid wins 88.1% of the time and pays the rival's price, about $2.80 on average, for a surplus of roughly $1,056 per million requests. Under first price, even the best shaded bid earns less than half of that, unless rivals also shade and the whole curve moves left. That is what the field data show: prices jumped when publishers switched, then drifted down as bidders learned (Goke et al., arXiv, October 2021).

A minimal implementation of the grid search looks like this. Production systems replace the fixed curve with a model per segment or per request.

Published work from DSP research teams falls into six families. They differ mainly in what feedback they need from the exchange. In an open first-price auction, every bidder receives the minimum bid to win. In a closed one, losers learn nothing about the winning price (Gligorijevic et al., CIKM 2020).

Two newer lines of research address robustness. One treats bidding as an online learning problem against adversarial rivals and proves regret bounds (Han et al., arXiv, July 2020). Another builds a shading policy that stays sound when both the value estimate and the forecast of competing bids are noisy, tested on Yahoo DSP data (Qu et al., arXiv, October 2024).

The production numbers give a sense of scale. The factorization machine model improved total surplus by 20.54% online over the system it replaced (Gligorijevic et al., CIKM 2020). The deep distribution network raised advertiser ROI by 2.4% for CPM and CPC campaigns and 8.6% for CPA campaigns in an online A/B test (Zhou et al., KDD 2021). All of these results are from one DSP's own research team, so treat them as company-reported.

Shading models learn from what comes back after each auction. OpenRTB carries three signals. The win or billing notice returns ${AUCTION_PRICE}, the clearing price. The loss notice URL (lurl) can return ${AUCTION_LOSS}, a reason code, and ${AUCTION_MIN_TO_WIN}, the price that would have tied the winner (OpenRTB 2.6, section 4.4). The loss codes that matter most for shading are 100, "Bid was Below Auction Floor", 101, "Bid was Below Deal Floor", and 102, "Lost to Higher Bid" (OpenRTB 3.0, List: Loss Reason Codes).

A bid that asks for all three looks like this:

Exchanges decide what to fill. The spec lets them return an empty string when a bid was blocked, when price did not decide the loss, or when privacy rules forbid sharing it (OpenRTB 2.6, section 4.4.1). Google committed in September 2019 to give Authorized Buyers the minimum bid to win after each auction (Google, 5 September 2019). Field-level detail is in OpenRTB 2.6 explained.

Read the auction type before you shade. The at field says 1 for first price and 2 for second price plus, and a deal can override it, with 3 meaning a fixed deal price. Microsoft's Xandr exchange documentation says it sends at 1 for first-price auctions and also when soft floors make the outcome uncertain. It marks those auctions "Shade" and trusted second-price auctions "Don't Shade" (Microsoft Learn, Xandr bidder transparent auction FAQ, updated 21 October 2025). Shading a second-price bid lowers your win rate and saves nothing.

Most buyers never see the shading step. They see a bid setting and a final price. These seven steps make the behaviour visible.

1. Read the setting and its default. Ask whether shading is on by default, whether it can be switched off per line item, and whether it touches deals. PPC Land reports that Display & Video 360 applies a setting called Optimized fixed bidding to fixed CPM bids by default (PPC Land, 23 August 2026).
2. Get bid-level logs. You need, per auction, the submitted bid, the clearing price, win or loss, the loss reason and, where the exchange provides it, the minimum bid to win. The log-level data guide covers how to request and store these files.
3. Compute the gaps. For wins, compare the clearing price with your line's maximum bid: that ratio is your effective shading factor. For losses, compare your bid with the minimum to win. Group both by exchange, domain and device.
4. Run a split test. Clone one line item. Leave shading on in one copy and bid a fixed CPM with shading off in the other, with the same budget, targeting and dates. Compare win rate, average CPM and cost per outcome over at least two weeks.
5. Check the fee base. Find out whether the DSP's percentage fee is charged on the shaded media cost or on the pre-shading bid, and whether shading appears as a separate line on the invoice. Fee structures are compared in DSP pricing explained.
6. Check fixed-price deals. On a deal with at 3 the price is agreed, so there is nothing to shade. On first-price deals the deal floor is the lower bound for any shaded bid.
7. Watch the auction type per exchange. If shading is applied where at is 2, the DSP is giving up impressions for no saving.

A log row that supports steps two to four can be as small as this. The field names are illustrative, not a vendor format.

Shading shifts value from sellers to buyers, so publishers answer with floor prices. Google told publishers in May 2019 that floor tactics built to lift second-price clearing prices would no longer be relevant. Floors should reflect the true value of the inventory instead. Its unified pricing rules apply to all partners equally and cannot be set per buying platform (Google, 10 May 2019).

In OpenRTB 2.6, a seller can state floors in three places: imp.bidfloor, the video or audio object (mincpmpersec or durfloors), and the deal. Buyers read the deal floor first, then the media floor, then the impression floor (OpenRTB implementation guide, section 7.11). In header bidding, the Prebid Price Floors Module lets a publisher set static floors or fetch dynamic ones from a data service, and Prebid Server supports the same rule schema (Prebid.org, Price Floors Module).

Floors and shading models now train against each other. A floor above a DSP's shaded bid produces loss code 100, which teaches the model where the floor sits. A floor set too high cuts fill. The strategy for setting floors in first-price auctions, including dynamic floors and how to test them, is the subject of floor price strategy.

Several practices lower a price without being shading. Bid caching reuses a losing bid in a later auction.

A negotiated discount is applied by the seller to the clearing price; since release 2.6-202606, OpenRTB reports it through ${AUCTION_DISCOUNT_PCT} and ${AUCTION_DISCOUNT_CPM} (2.6-202606 release notes, 11 June 2026). Bid deflation by an intermediary is a different matter again.

Teads' August 2026 lawsuit against Google alleges that a program called Project Poirot reduced Display & Video 360 bids into non-Google exchanges (PPC Land, 23 August 2026). That is a claim in pending litigation, not a finding. Shading, by contrast, is done by or for the buyer to keep surplus on the buyer's side.

1. Shading second-price auctions. Check at on the request and on each deal. Shading where the auction already prices you at the runner-up only loses impressions.
2. Training only on wins. Won auctions show what you paid, not what you needed to pay. Without loss outcomes the model learns a biased curve (Wu et al., KDD 2015).
3. Treating an empty ${AUCTION_MIN_TO_WIN} as zero. An empty value means the exchange chose not to disclose it. Drop the row from training or treat it as censored.
4. One curve for all inventory. Rival bids differ widely by page, user and exchange. The Verizon Media team shows four segments with very different winning-price distributions (Gligorijevic et al., CIKM 2020).
5. Ignoring floors. A shaded bid below bidfloor cannot win. Clip the shaded bid at the floor, or drop the bid if the floor exceeds the value.
6. Shading fixed-price deals. If at is 3 the floor is the agreed price; bidding below it simply loses the deal.
7. Letting pacing and shading fight. Both adjust the same final number. If the budget pacer raises bids while the shader lowers them, neither works as tested. Apply them in a fixed order and log both factors.
8. Judging shading by CPM alone. Lower CPMs with a collapsing win rate can raise cost per outcome. Measure surplus or cost per action, not price.

The RTB architecture hub links the related engineering pages on bidders, latency and OpenRTB.

1. Google, "Simplifying programmatic: first price auctions for Google Ad Manager" (6 March 2019). https://blog.google/products/admanager/simplifying-programmatic-first-price-auctions-google-ad-manager/
2. Google, "An update on first price auctions for Google Ad Manager" (10 May 2019). https://blog.google/products/admanager/update-first-price-auctions-google-ad-manager/
3. Google, "Rolling out first price auctions to Google Ad Manager partners" (5 September 2019). https://www.blog.google/products/admanager/rolling-out-first-price-auctions-google-ad-manager-partners/
4. Google Ad Manager Help, "2019 Google Ad Manager releases archive" (accessed 9 November 2026). https://support.google.com/admanager/answer/9197913?hl=en
5. PPC Land, "Google moves all auctions to first price by late July" (19 April 2019). https://ppc.land/google-moves-all-auctions-to-first-price-by-late-july/
6. D. Gligorijevic, T. Zhou, B. Shetty, B. Kitts, S. Pan, J. Pan, A. Flores, "Bid Shading in The Brave New World of First-Price Auctions", CIKM 2020 (arXiv, 2 September 2020). https://arxiv.org/abs/2009.01360
7. S. Pan, B. Kitts, T. Zhou et al., "Bid Shading by Win-Rate Estimation and Surplus Maximization", AdKDD 2020 (arXiv, 19 September 2020). https://arxiv.org/abs/2009.09259
8. T. Zhou, H. He, S. Pan et al., "An Efficient Deep Distribution Network for Bid Shading in First-Price Auctions", KDD 2021 (arXiv, 12 July 2021). https://arxiv.org/abs/2107.06650
9. Y. Han, Z. Zhou, A. Flores, E. Ordentlich, T. Weissman, "Learning to Bid Optimally and Efficiently in Adversarial First-price Auctions" (arXiv, 9 July 2020). https://arxiv.org/abs/2007.04568
10. Y. Qu, R. Kant, Y. Chen et al., "Double Distributionally Robust Bid Shading for First Price Auctions" (arXiv, 18 October 2024). https://arxiv.org/abs/2410.14864