Agentic Advertising
Explainable AI in bidding: what to ask of a bidding model, and how to check the answer
Published · Last updated
Next scheduled review:
Explainable AI in bidding means being able to say why a model bid what it did. A buyer or auditor can ask for five things: global feature importance, per-bid attributions such as SHAP values, decision logs, counterfactual tests and model documentation. Public DSP disclosures cover fewer of these, and EU law asks for explanations only in narrow cases.
Contents
Key facts
- SHAP made per-prediction attributions add up. Lundberg and Lee's method splits one prediction into feature contributions that sum to the model's output, a property they call local accuracy (Lundberg and Lee, NIPS 2017, arXiv).
- DSPs disclose the controls, not the weights. Display & Video 360 lets buyers set rule weights or write scripts for custom bidding, after which the algorithm "goes through a training phase and uses machine learning" (Google, DV360 Help, accessed 9 November 2026).
- The Trade Desk shows scores and modes. Kokai shows a relevance score against a seed audience, and in Q2 2026 the Koa optimisation switch became three modes, with Koa on by default (PPC Land, 21 October 2026).
- Courts want the logic, not the code. In case C-203/22 (27 February 2025), the Court of Justice said a person must be told the procedure and principles actually applied. Handing over the algorithm is neither required nor enough (Norton Rose Fulbright, Data Protection Report, 10 March 2025).
- The AI Act's high-risk rules moved. Regulation (EU) 2026/1744 was adopted on 8 July 2026 and entered into force on 27 July 2026. It moved Annex III high-risk obligations to 2 December 2027 and Annex I obligations to 2 August 2028 (EUR-Lex, OJ L 2026/1744).
Definition
Explainable AI in bidding: the methods, records and tests that let a buyer, seller or auditor reconstruct which inputs drove a bidding model's valuation of an impression and its final bid. It works at the level of a single bid and across a whole campaign.
Explainability is not the same as accuracy. A model can predict well and still be impossible to question. It is also not fee transparency. A demand-side platform (DSP) can disclose every fee and still give no reason for a bid. This page sits in the agentic advertising hub. To see explanations change as you move inputs, try the explainable AI RTB simulator.
What a bidding model decides, and what can be explained
A bid is rarely one model. Most bidders chain several steps, and each step needs a different kind of explanation. The table splits a typical chain into four layers. The layer names are ours; the structure follows the RTB bidder architecture page.
4 of 4 rows
Eligibility filters | Whether the campaign may bid at all | Rules: blocklists, targeting, consent | The rule that passed or blocked the request |
Value prediction | Probability of a click, view or conversion | Machine-learning model | Feature attributions for this prediction |
Pacing and budget | How much of the value to spend now | Controller or learned policy | The multiplier and the budget state behind it |
Price setting | How far below value to bid | Bid shading model in a first-price auction | The predicted winning price and the shade factor |
Eligibility filters
- What it decides
Whether the campaign may bid at all
- Typical method
Rules: blocklists, targeting, consent
- What an explanation looks like
The rule that passed or blocked the request
Value prediction
- What it decides
Probability of a click, view or conversion
- Typical method
Machine-learning model
- What an explanation looks like
Feature attributions for this prediction
Pacing and budget
- What it decides
How much of the value to spend now
- Typical method
Controller or learned policy
- What an explanation looks like
The multiplier and the budget state behind it
Price setting
- What it decides
How far below value to bid
- Typical method
Bid shading model in a first-price auction
- What an explanation looks like
The predicted winning price and the shade factor
Table by AdSovereign, CC BY 4.0.
An auditor who only sees the final bid cannot tell which layer moved it. That is why the first request below is for logs that keep each layer's output separate. Bid shading is explained in depth in bid shading explained, and pacing as a learning problem in reinforcement learning for bidding.
Five things a buyer or auditor can ask for
You can ask any vendor for these five items. Each answers a different question, and none is enough alone.
5 of 5 rows
| Average absolute attribution per feature over many bids | Which inputs the model relies on most | Hides interactions and rare cases |
| SHAP or LIME run on logged bids | Why this bid was high or low | Needs the model and its exact inputs; costly at scale |
| The bidder writes each layer's output per bid | What the system actually did | Only as complete as the fields logged |
| Replay logged requests with one input changed | Whether a feature changes bids, without model access | Shows effects, not internal reasons |
| A model card: purpose, data window, metrics, limits, update dates | That the vendor knows and states the model's limits | Self-reported; must be checked against logs |
- Global feature importance
- How it is produced
Average absolute attribution per feature over many bids
- What it proves
Which inputs the model relies on most
- Limits
Hides interactions and rare cases
- Per-bid attributions (SHAP-style)
- How it is produced
SHAP or LIME run on logged bids
- What it proves
Why this bid was high or low
- Limits
Needs the model and its exact inputs; costly at scale
- Decision logs
- How it is produced
The bidder writes each layer's output per bid
- What it proves
What the system actually did
- Limits
Only as complete as the fields logged
- Counterfactual tests
- How it is produced
Replay logged requests with one input changed
- What it proves
Whether a feature changes bids, without model access
- Limits
Shows effects, not internal reasons
- Model documentation
- How it is produced
A model card: purpose, data window, metrics, limits, update dates
- What it proves
That the vendor knows and states the model's limits
- Limits
Self-reported; must be checked against logs
Table by AdSovereign, CC BY 4.0.
The methods behind rows 2, 4 and 5 are published. LIME fits a simple local model around one prediction (Ribeiro, Singh and Guestrin, KDD 2016, arXiv). Counterfactual explanations state the smallest change in inputs that would change the outcome, without opening the model (Wachter, Mittelstadt and Russell, Harvard Journal of Law and Technology, 2018, arXiv). Model cards were proposed as short documents reporting a model's intended use and its performance across conditions (Mitchell et al., FAT* 2019, arXiv).
NIST's four principles of explainable AI frame the same requests from the user's side: an explanation, meaningful to its audience, accurate about the process, within the system's knowledge limits (NIST IR 8312, 29 September 2021). The last principle matters in bidding. A model should be able to say "this input is outside what I was trained on" rather than bid with confidence.
One bid, explained: a SHAP-style record
The record below shows what a per-bid explanation can look like when it is logged next to the bid. It is our illustrative format, not any DSP's output. The model predicts click probability; the bid then passes through pacing and shading.
{
"bid_id": "b-20261109-000418",
"campaign_id": "cmp-2207",
"model_version": "ctr-gbdt-2026-11-02",
"prediction_target": "click_probability",
"base_value": 0.0020,
"attributions": {
"site_category=news_sport": 0.0009,
"user_last_site_visit_days=2": 0.0011,
"hour_of_day=21": 0.0003,
"device_type=mobile_app": -0.0002,
"creative_size=300x250": -0.0001
},
"predicted_ctr": 0.0040,
"value_per_click_usd": 1.50,
"value_cpm_usd": 6.00,
"pacing_multiplier": 0.90,
"shading_factor": 0.75,
"final_bid_cpm_usd": 4.05,
"outcome": { "won": true, "clearing_price_cpm_usd": 3.70 }
}6 of 6 rows
| They must add up to |
| 0.0040 × $1.50 × 1,000 = $6.00, the value of a thousand such impressions |
| Budget state at bid time; $6.00 × 0.90 = $5.40 |
| First-price shading; $5.40 × 0.75 = $4.05 |
| Ties the explanation to a model you can request documentation for |
| Lets you compare predicted value with what was paid and what happened |
base_value plus attributions
- What to check
They must add up to
predicted_ctr: 0.0020 + 0.0020 = 0.0040. If they do not, the explanation is wrong or incomplete.
value_cpm_usd
- What to check
0.0040 × $1.50 × 1,000 = $6.00, the value of a thousand such impressions
pacing_multiplier
- What to check
Budget state at bid time; $6.00 × 0.90 = $5.40
shading_factor
- What to check
First-price shading; $5.40 × 0.75 = $4.05
model_version
- What to check
Ties the explanation to a model you can request documentation for
outcome
- What to check
Lets you compare predicted value with what was paid and what happened
Table by AdSovereign, CC BY 4.0.
The sum check in the first row is the point of SHAP's local accuracy property (Lundberg and Lee, NIPS 2017). Real models often work in log-odds rather than probability, so attributions are summed before conversion. Ask the vendor which space its numbers are in. Without that, a buyer cannot check the arithmetic. The open-source SHAP library reached version 0.54.1 on 16 October 2026 (shap/shap releases).
The record also shows what this layer cannot tell you. The user's recent visit raised the predicted click rate, but the record does not say whether using that signal was allowed. That question belongs to the consent and data checks in the audit plan below.
Counterfactual tests you can run without the model
Most buyers will never get a vendor's model. They can still test it from outside, using their own log-level data and a small budget.
- Replay tests. If the vendor offers a bid simulator or a test endpoint, send logged bid requests with one field changed, such as the domain, the hour or the device type. Compare the bids.
- Split campaigns. Run two identical line items that differ in one input, for example one with a data segment and one without. Compare bids and outcomes over the same period.
- Holdout audiences. Exclude a random share of users from a segment and compare bids on matched inventory.
- Price probes. On deals with a known floor, track how often the model bids just above it. Clustering at the floor suggests shading driven by the floor, not by value.
These tests show effects, not reasons. They are still the strongest evidence a buyer can produce alone, because they do not depend on the vendor's own account. The data side of such tests is covered in log-level data, and eight ways to combine DSP and external signals in DSP, DMP and external data strategies.
What DSPs and standards disclose publicly
This table records what we could confirm in public documentation on 9 November 2026, with a source for each row. It covers public pages only; contract-level disclosures to large clients may go further. "n/d" means not disclosed in the sources reviewed.
3 of 3 rows
Google Display & Video 360, custom bidding | Buyers assign weights to impression signals and outcomes (rules) or write scripts; the algorithm trains on past campaigns with machine learning; scripts can be uploaded and verified by API | n/d | |
The Trade Desk, Kokai with Koa | Models for bid clearing, relevance scoring, budget optimisation, KPI scoring and forecasting; relevance score against a seed; three Koa modes since Q2 2026, on by default; Zuma "applied settings" view from 27 Aug 2026 | n/d | |
IAB Tech Lab AAMP 2.3 (agent standard) | A pricing provenance field records where each price in an agent negotiation came from | Not applicable |
Google Display & Video 360, custom bidding
- What is disclosed
Buyers assign weights to impression signals and outcomes (rules) or write scripts; the algorithm trains on past campaigns with machine learning; scripts can be uploaded and verified by API
- Per-bid attributions
n/d
- Source and date
The Trade Desk, Kokai with Koa
- What is disclosed
Models for bid clearing, relevance scoring, budget optimisation, KPI scoring and forecasting; relevance score against a seed; three Koa modes since Q2 2026, on by default; Zuma "applied settings" view from 27 Aug 2026
- Per-bid attributions
n/d
- Source and date
IAB Tech Lab AAMP 2.3 (agent standard)
- What is disclosed
A pricing provenance field records where each price in an agent negotiation came from
- Per-bid attributions
Not applicable
- Source and date
Table by AdSovereign, CC BY 4.0.
The pattern is consistent. Platforms explain what the buyer controls: goals, weights, modes and settings. They do not publish what the model learned. The Trade Desk's move to an "applied settings" view is an example of explaining configuration rather than predictions. PPC Land, reporting the change, calls Kokai the place where "how much optimisation can be explained" gets tested (PPC Land, 21 October 2026).
Where EU law asks for explanations
Three EU laws touch this topic. None was written for ad bidding, and each applies only in defined situations. This section describes the rules; it is not legal advice.
GDPR. Article 22 covers decisions "based solely on automated processing" that produce legal effects or similarly significant effects. For those decisions, Article 15(1)(h) gives people a right to "meaningful information about the logic involved" (GDPR, EUR-Lex).
In C-203/22, the Court held that this means explaining the procedure and principles actually applied. Where trade secrets are involved, the information goes to the supervisory authority or court, which balances the interests (Norton Rose Fulbright, 10 March 2025). The case concerned a credit score. The commentary notes it might be enough to say how a change in the data would have changed the result, which is a counterfactual explanation. Whether a given ad decision has a "similarly significant" effect is a fact-specific question.
Digital Services Act. Article 26(1)(d) requires online platforms to show, for each ad, "meaningful information" about "the main parameters used to determine the recipient" (DSA, EUR-Lex). The duty falls on the platform showing the ad, but platforms often ask buyers and DSPs for the parameters.
AI Act. The timetable below reflects Regulation (EU) 2026/1744, published on 24 July 2026. Ordinary targeting and bidding models are not an Annex III category. The caveat is creditworthiness: Annex III point 5(b) covers systems that evaluate the creditworthiness of natural persons, so a bidder that scores people's credit for a lender's offers needs a closer look. The full analysis, including that caveat, is in the EU AI Act and advertising.
5 of 5 rows
2 February 2025 | Article 5 prohibited practices, including manipulative techniques and exploiting vulnerabilities | Applies to any AI system, including ad models, if the conduct meets the article's tests |
2 August 2026 | Article 50 transparency duties, including labelling of AI-generated content | Mainly creative generation, not bid pricing |
2 December 2026 | New bans in Article 5(1)(ba) and (bb) on generating non-consensual intimate images and child sexual abuse material; Article 111(4) marking deadline for generative systems placed on the market before 2 August 2026 | Creative generation tools, not bid pricing |
2 December 2027 | Annex III high-risk obligations, including logging, human oversight and, under Article 86, a right to explanation for certain decisions | Only if a system falls into an Annex III category, such as creditworthiness under point 5(b) |
2 August 2028 | Annex I obligations for AI in regulated products | Rarely relevant to ad bidding |
2 February 2025
- What applies
Article 5 prohibited practices, including manipulative techniques and exploiting vulnerabilities
- Relevance to bidding
Applies to any AI system, including ad models, if the conduct meets the article's tests
2 August 2026
- What applies
Article 50 transparency duties, including labelling of AI-generated content
- Relevance to bidding
Mainly creative generation, not bid pricing
2 December 2026
- What applies
New bans in Article 5(1)(ba) and (bb) on generating non-consensual intimate images and child sexual abuse material; Article 111(4) marking deadline for generative systems placed on the market before 2 August 2026
- Relevance to bidding
Creative generation tools, not bid pricing
2 December 2027
- What applies
Annex III high-risk obligations, including logging, human oversight and, under Article 86, a right to explanation for certain decisions
- Relevance to bidding
Only if a system falls into an Annex III category, such as creditworthiness under point 5(b)
2 August 2028
- What applies
Annex I obligations for AI in regulated products
- Relevance to bidding
Rarely relevant to ad bidding
Table by AdSovereign, CC BY 4.0.
Dates from Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, which also rewrote Article 4 on AI literacy.
An audit plan for a bidding model in eight steps
- Map the chain. Ask the vendor to list every layer between the bid request and the bid, with the model or rule behind each.
- Get the model documentation. Request a model card per model: purpose, training data window, evaluation metrics, known limits and the date of the last retraining.
- Agree a log schema. Each bid should carry the model version and each layer's output, as in the record above.
- Request global importance. Ask for the top features per model, updated after each retraining.
- Sample per-bid explanations. Ask for SHAP-style attributions on a random sample of logged bids, and check that they add up.
- Run counterfactual tests. Use replay or split campaigns to confirm that the features the vendor names actually move bids.
- Check data permissions. For each influential feature, confirm the data source, the consent basis and whether the feature is a sensitive category.
- Repeat on change. Re-run steps 4 to 6 after every model update the vendor announces, and write it into the contract.
FAQ
What is explainable AI in advertising?
Can I get SHAP values from my DSP?
Does the EU AI Act apply to programmatic bidding?
What should a bidding model audit include?
Is model explainability the same as fee transparency?
How this works on owned infrastructure
Every request in the audit table depends on having the model, its exact inputs and its logs in one place. When a vendor hosts the bidder, the buyer gets whatever explanation the vendor chooses to publish, usually settings and scores rather than attributions. AdSovereign is a bidding and data stack designed to run inside the buyer's own cloud account or data centre, so the value models, pacing state and bid logs sit in infrastructure the buyer controls. The data team can then compute SHAP-style attributions on its own sample and replay requests against a fixed model version without asking anyone. The architecture page shows where models and logs live, and the security page covers access controls for them.
Sources
- Scott Lundberg and Su-In Lee, "A Unified Approach to Interpreting Model Predictions" (arXiv, 22 May 2017; NIPS 2017). https://arxiv.org/abs/1705.07874
- Marco Tulio Ribeiro, Sameer Singh and Carlos Guestrin, "'Why Should I Trust You?': Explaining the Predictions of Any Classifier" (arXiv, 16 February 2016; KDD 2016). https://arxiv.org/abs/1602.04938
- Sandra Wachter, Brent Mittelstadt and Chris Russell, "Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR" (arXiv, 1 November 2017; Harvard Journal of Law and Technology, 2018). https://arxiv.org/abs/1711.00399
- Margaret Mitchell et al., "Model Cards for Model Reporting" (arXiv, 5 October 2018; FAT* 2019). https://arxiv.org/abs/1810.03993
- P. Jonathon Phillips et al., "Four Principles of Explainable Artificial Intelligence", NIST IR 8312 (29 September 2021). https://doi.org/10.6028/NIST.IR.8312
- SHAP project, GitHub releases (v0.54.1, 16 October 2026). https://github.com/shap/shap/releases
- Google, Display & Video 360 Help, "Custom bidding overview" (accessed 9 November 2026). https://support.google.com/displayvideo/answer/9723477?hl=en
- PPC Land, "Explaining Kokai: The Trade Desk's AI prices 20 million ad impressions a second, in 10 ms or less" (21 October 2026). https://ppc.land/kokai/
- PPC Land, "AAMP 2.3 blocks AI agents from inventing ad prices, IAB Tech Lab says" (30 July 2026). https://ppc.land/aamp-2-3-blocks-ai-agents-from-inventing-ad-prices-iab-tech-lab-says/
- Marcus Evans, Farah Mukaddam and Rosie Nance, "What do organisations need to disclose to individuals about AI and automated decisions?", Norton Rose Fulbright Data Protection Report (10 March 2025; on CJEU C-203/22, Dun & Bradstreet Austria). https://www.dataprotectionreport.com/2025/03/what-do-organisations-need-to-disclose-to-individuals-about-ai-and-automated-decisions/
More sources (4)
- Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 15 and 22. https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Regulation (EU) 2022/2065 (Digital Services Act), Article 26. https://eur-lex.europa.eu/eli/reg/2022/2065/oj/eng
- Regulation (EU) 2024/1689 (Artificial Intelligence Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
- Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689 (Official Journal, 24 July 2026). https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:L_202601744