How Advertisers Are Using AI to Improve Performance, and Why Results Depend on Data and Governance

The advertising industry is implementing new forms of AI and automation while reassessing the foundations beneath the old ones. In a 2026 survey conducted by the Interactive Advertising Bureau (IAB), 60%–75% of advanced-measurement users said current approaches fall short on rigor, timeliness, trust or efficiency. About half were already scaling AI, with most of the remaining respondents expecting to follow within two years.

As the use of AI in advertising expands, teams have more room to test ideas, adapt creative, refine targeting and iterate while a campaign is still running. But as generative AI and agentic AI expand what advertisers can do, they also increase the importance of the information and governance surrounding the system.

Dennis Buchheim, Global Head of Media, Entertainment and Adtech/Martech at Snowflake, puts it this way: “The technology is moving fast, but what really determines success is whether you can control, govern and audit what’s happening.”

Advertisers using AI now must address both sides of the equation — the expanding capabilities of the technology and the data foundation and controls required to keep decisions aligned with each campaign’s purpose.

Advertising performance depends on how well brands and agencies manage a campaign’s many moving parts. AI gives them a way to evaluate more of those variables together, adapt them while the campaign is active and identify problems before they consume additional spend.

Audience targeting and predictive modeling

Audience development begins with a campaign’s intended outcome: a first purchase, a renewal, a product upgrade or reduction in churn. Propensity models rank customers or prospects according to the likelihood of that outcome, while lookalike modeling extends a known audience — often recent converters or high-value customers — to people who share relevant characteristics. Estimates of customer lifetime value, churn risk and product affinity provide ways to prioritize spend and shape the message.

The most likely buyer isn’t always the most valuable person to target. When a campaign is intended to drive incremental sales, spend shouldn’t concentrate on customers who were already planning to purchase. Uplift and persuasion models estimate whose likelihood of acting changes after exposure.

Creative generation and dynamic creative optimization

Advertising teams often need to produce the same idea in dozens of forms: different aspect ratios, products, audience segments, markets, languages and channels. Generative AI can reduce some of the production burden by drafting copy, developing visual concepts and adapting approved material for new formats.

At a larger scale, AI can also change how teams explore creative options. Advertisers can develop a wider range of initial directions and eliminate weak variants earlier. Dynamic creative optimization (DCO) carries variation into delivery. Headlines, images, offers and calls to action (CTAs) can be assembled according to the audience, placement or surrounding context, with campaign results informing which combinations continue to run.

As production expands, so does the need for control over the source material, however. Prices, product claims, disclosures and brand language should come from approved content, while regional or regulated campaigns require rules governing geography, eligibility and required wording. Generative AI can widen the creative field, but the final decision must rest with people who understand the brand, product and audience.

Advertisers should preserve records of the models, prompts, source assets and approvals used to create material campaign content. Where required — or where consumers could otherwise be misled — AI-generated or materially manipulated media should be clearly labeled. Provenance standards can supplement visible disclosures by helping platforms and reviewers verify how an asset was created or modified.

Programmatic bidding and budget optimization

Programmatic advertising automates decisions that have to be made within fractions of a second. AI can help improve the estimates behind those decisions, with the ability to analyze more signals.

Across the wider campaign, pacing and optimization models help advertisers avoid spending too quickly, leaving budget unused or continuing to fund a channel after its performance has deteriorated. These models also provide a basis for reallocating media among publishers, audiences and placements as new results become available.

The optimization target determines what those adjustments reward. Return on ad spend (ROAS) metrics connect media cost with attributed revenue, but even that result depends on complete conversion data and a defensible method for deciding which sales the campaign should receive credit for.

For a clearer view of contribution, advertisers often combine platform metrics with experiments, incrementality analysis and marketing mix modeling to distinguish campaigns that actually changed behavior.

Personalization and message sequencing

Personalization extends far beyond choosing a product recommendation. It can also determine which message best fits the customer’s current relationship with the brand, what messages the customer has already seen and whether another exposure is likely to change behavior.

A new prospect will need introductory creative, while an existing customer will need onboarding, replenishment or information tied to a product already owned. After a purchase, suppression rules can remove acquisition messages before the same customer sees them repeatedly across other channels.

Coordination is difficult when each system sees only one part of the interaction history. Email, ecommerce, mobile apps, store purchases and media exposure can each produce a different version of the customer, leading to duplicated messages, conflicting offers and poor frequency control. AI helps manage this complexity when the identity, customer state and channel permissions remain consistent across the workflow.

Brand safety and fraud detection

Campaign performance depends partly on what happens around the ad. An otherwise strong campaign can stumble if it appears beside unsuitable content or when impressions and clicks come from invalid traffic.

For brand safety, contextual models examine the subject, tone and surrounding material across webpages, images and video. This context provides a more nuanced assessment than keyword blocking alone, which can exclude appropriate journalism while missing risky content that contains no obvious prohibited term.

Fraud detection models can identify risky device behavior, implausible engagement, unusual traffic timing and discrepancies among systems that may indicate bots or other forms of invalid activity.

The information needed to plan, activate and measure advertising is usually split across organizations. Publishers and platforms hold audience, impression and placement data, while advertisers hold customer identities, transactions and other business outcomes. Connecting those events requires a reliable view of identity, a shared definition of the outcome and a way for organizations to collaborate without exposing unrestricted customer data.

As third-party identifiers have weakened, that work has shifted toward first-party relationships, identity resolution and controlled data collaboration.

Closed-loop measurement and attribution

Closed-loop measurement connects campaign exposure with an outcome recorded in the advertiser’s own systems. Purchases, qualified leads, subscriptions and store visits provide a more direct view of business performance than clicks or platform-configured conversions alone.

This metric can’t prove the campaign caused the result, however. Targeting may favor customers with stronger purchase intent to begin with, making it difficult to tell how much of the observed conversion rate came from the advertising itself. Under these conditions, traditional attribution can give a campaign credit for a sale that would have happened regardless.

Incrementality analysis focuses on the difference created by the advertising. Holdout tests, randomized experiments and causal-inference methods compare exposed and unexposed groups, helping advertisers estimate lift rather than simply assign credit.

Identity resolution across fragmented signals

The same customer can appear under a variety of identifiers, including an email address, account ID, browser identifier, connected-device ID and point-of-sale record. Unless those records are connected, reach is overstated, frequency limits break and conversion measurements are inaccurate.

Deterministic identity resolution uses stable identifiers such as authenticated accounts or hashed email addresses. Probabilistic approaches estimate relationships from less direct signals, including devices, locations and behavior. Because probabilistic matches carry uncertainty, confidence thresholds should reflect the use case.

Data clean rooms for advertising collaboration

Campaigns involve a variety of parties. Advertisers bring customer and transaction data, publishers contribute exposure and authenticated-audience data, and identity or measurement partners supply additional linking or analytical capabilities.

A data clean room gives those participants a controlled environment for comparing data without granting unrestricted access to the underlying records. Audience overlap, reach analysis, activation and campaign measurement can all occur under approved policies.

The clean room governs the collaboration, but consent, contracts and applicable privacy requirements remain the responsibility of the participating organizations.

Learn how Snowflake Data Clean Rooms support audience overlap, lookalike modeling and attribution analysis across the marketing lifecycle:

Retail media measurement

Retail media networks combine advertising inventory with detailed purchase data, creating a direct connection between campaign exposure and sales at the SKU, category or customer-segment level.

This visibility gives brands a more precise way to evaluate advertising close to the point of purchase, while the retailer can return approved analysis without handing over unrestricted customer records.

The result still depends on shared rules. Advertisers and retailers need to agree on attribution windows, eligible transactions, returns and whether the analysis reflects attributed revenue or incremental sales.

A shared data foundation gives advertisers a common view of customers, campaigns, creative, media and outcomes before that information is reused across models, applications and agents. Definitions, permissions and lineage can remain attached to the data rather than being recreated inside every tool.

Buchheim emphasizes the importance of the data foundation: “The intelligent use of data is not a nice-to-have. It’s the only way to achieve operational efficiency, better understand customer journeys, enable more personalized ads and media experiences, and drive desired business outcomes.”

Unified first-party data

First-party advertising data includes customer profiles, transactions, loyalty activity, website and app events, subscriptions, customer-service interactions and campaign history. Media data adds impressions, reach, frequency, cost, placement and creative.

A customer 360 connects the relevant parts of that history to a governed identity. Context includes consent, channel eligibility, lifecycle stage, recent exposure and previous offers.

Campaign records require the same consistency. Common IDs and definitions should connect planning, media buying, serving, measurement and finance so that spend, creative and results can be reconciled without rebuilding the campaign from several reports.

Freshness requirements should match the use case. Programmatic bidding operates within very short time windows, audience refreshes often run daily, for example, while marketing mix models usually rely on weekly or monthly data.

Shared advertising and measurement definitions

Common advertising metrics often hide incompatible definitions. Reach might refer to people, households, devices or platform identities. A conversion might mean a purchase, lead, store visit or any event configured in the platform. ROAS might rely on gross revenue, net revenue or margin, with different rules for returns and discounts.

AI systems need the approved meaning of a label, not just the name. For example, a conversational interface comparing campaigns should know which revenue measure applies and which attribution logic produced it. Semantic definitions preserve these rules alongside the underlying data. Once campaign, audience and outcome metrics carry consistent meaning, they can be reused across dashboards, models and agents without rebuilding the logic each time.

Governance across the advertising lifecycle

Advertising data includes personal information, proprietary campaign plans, publisher pricing, product-launch details and confidential performance results. Role-based access controls can help scope appropriate access to this sensitive data.

Lineage gives advertisers a traceable record of how an audience, model input or performance result was produced. It shows which source data was used, which rules and transformations were applied and which model or attribution method generated the final output.

As AI systems take on more work, governance also has to cover their actions. Generative models need approved source material and controls over the claims they produce. Predictive models require evaluation as audiences and market conditions change. Agents need explicit permissions and approval thresholds before they alter budgets, activate audiences or publish creative.

Models and agents should also be monitored after deployment for changes in input data, performance, cost and behavior. Version histories, decision logs, test results and rollback procedures give teams a way to investigate an unexpected budget shift, audience change or creative output and restore a previously approved configuration.

Responsible AI centers on three areas for advertising use cases: how customer data is collected and applied, how much authority automated systems receive and how generated or selected creative reaches the market. Weak controls in any area can expose the organization to privacy violations, wasted spend or public brand damage.

Privacy, consent and appropriate use

First-party data should be collected and used according to applicable privacy requirements, contractual restrictions and the organization’s commitments to customers. Consent needs enough specificity to determine which advertising and personalization activities are permitted.

When records are connected across systems or organizations, the resulting profile can be more sensitive than any individual source. Identity resolution, customer 360 and data clean room programs should preserve purpose limitations rather than treating technical access as permission.

Human oversight

Routine changes can operate within defined limits, while decisions with larger financial, legal or reputational consequences should require human review. A bidding system, for example, can adjust spend within an approved range while routing a major reallocation to the media team.

The same principle applies to AI agents. Reading campaign results, drafting a recommendation, changing an audience and launching a campaign represent different levels of risk. Tool permissions and approval thresholds should reflect those differences.

Creative quality and brand safety

Producing more variants doesn’t guarantee better advertising. Generative AI output can be generic, visually inconsistent or inaccurate, even when it appears polished.

Approved source assets, creative rules and evaluation criteria provide a basis for review. Before publication, teams should assess factual accuracy, brand consistency, accessibility, required disclosures and the context in which the ad will appear.

Review continues after launch. Audience response, placement quality and uneven performance can expose problems that weren’t visible during production, and those findings should inform both the campaign and later model evaluation.

Bias review and mitigation

Model evaluation should also address bias — who’s included, excluded or disproportionately exposed. Advertisers may need to test delivery, error rates, offers and outcomes across relevant groups, particularly in regulated or high-impact categories.

Advertising workflows draw on customer, campaign, creative, impression and outcome data held across several systems and organizations. Within Snowflake’s AI Data Cloud for Advertising, Media and Entertainment, those data types can be governed and analyzed together while preserving separation among advertisers, agencies, publishers, adtech providers and measurement partners.

For predictive targeting, bidding and measurement, Snowflake supports model development and inference near the data used by the application. Cortex AI supports tasks such as extracting information from briefs, classifying media content and analyzing creative assets alongside campaign and customer data. Cortex Search retrieves relevant material from approved sources, while Cortex Analyst provides natural-language access to governed metrics through semantic models.

Where the workflow crosses company boundaries, Snowflake Data Clean Rooms support audience overlap, activation and measurement through approved analysis templates. Participants collaborate on governed data while controls limit access to the underlying records.

Snowflake Horizon Catalog provides governance, discovery and lineage across analytics, machine learning and AI applications. Policies remain attached to the data as different teams and systems use it, helping organizations apply a more consistent control model across the campaign lifecycle.

AI will keep taking on more of the work involved in planning, running and measuring campaigns. The real test is whether that leads to better advertising: clearer decisions, stronger creative, better targeting and a more accurate view of what actually worked. Strong outcomes depend on how well systems stay connected to the campaign’s objective, the permissions and policies governing their actions and the signals and context available.

See how AI is being used in other industries >

Explore the AI in Industries Hub:

• AI in Cybersecurity
• AI in Financial Services
• AI in Healthcare and Life Sciences
• AI in Manufacturing
• AI in Retail