.png){kind=logo}
As advertising budgets continue to move closer to the moment of purchase, the role of ads is quietly changing. They no longer stop at influence. Increasingly, they participate directly in conversion. Retail media alone is projected to reach nearly $70 billion in annual U.S. ad spend within the next two years, and with that growth comes a new class of expectations. Ads are no longer evaluated only on reach or engagement. They are judged on whether they behave correctly at the point where money changes hands.
This transition has exposed a fault line in how most advertising systems were designed. Many were built to optimize persuasion, not transaction integrity. Promo codes, discounts, and offers were treated as secondary enhancements, often applied downstream at checkout. Today, those same mechanics are appearing directly inside the ad experience itself.
That shift turns a marketing surface into a form of checkout infrastructure.
Vishal Jain, a seasoned Software Engineering Leader and a judge in the IEEE Senior Member Application Review Panel, with over two decades of experience, has built his career at the boundary where advertising systems stop being abstract optimization engines and start carrying real economic consequences. His work centers on how large-scale platforms behave once automated decisions directly shape revenue outcomes, user trust, and the operational credibility of the system itself. Rather than treating scale and correctness as competing priorities, he approaches them as inseparable constraints, designing platforms that are expected to act predictably even when signals are incomplete and the stakes are high.
We spoke with him about what it takes to engineer conversion-grade advertising systems and why promotional logic has become a first-class systems problem.
Because intent is forgiving. Transactions are not.
Traditional advertising systems were designed around probabilistic outcomes. Clicks, impressions, and even conversion rates tolerate ambiguity. A user might click twice. Attribution might be slightly off. Those imperfections were acceptable because the system itself was not making binding decisions. Responsibility still lived with the user at checkout.
Once an ad automatically applies a promotion or discount, that tolerance disappears. The platform is now committing on behalf of both the advertiser and the user. If the offer has expired, misapplied, or incorrect, the failure is immediate and visible. There is no downstream buffer to absorb it.
At that point, advertising systems inherit responsibilities that look much closer to pricing or payments infrastructure. Correctness matters more than optimization. Explainability matters more than clever ranking and perhaps most importantly, the system needs to know when not to act. That distinction is what separates influence-grade systems from conversion-grade systems. One optimizes likelihood. The other must guarantee behavior.
Because promotional data is inherently unstable.
Promotions originate from many sources; they have eligibility rules that evolve, expiration windows that change, and constraints that are often inconsistently defined. At a small scale, humans compensate for this messiness. At platform scale, automation amplifies it.
Most systems were never designed to reason deeply about promotional integrity. Promo logic was frequently treated as a marketing concern or an edge case, bolted on late and handled downstream. That approach works as long as promotions remain peripheral. It breaks the moment promotions move upstream into the ad experience itself.
When that shift happens, every assumption changes. Discovery latency matters because stale offers erode trust, and expiration accuracy matters because errors multiply across millions of impressions. What used to be a minor inconsistency becomes a platform-level failure mode. I see the same pattern more broadly in my role as an invited paper reviewer for the ECIS 2026 Conference: the challenge is no longer generating offers. It is determining which ones are safe to apply, under what conditions, and with what level of confidence.
The first realization was that this could not be treated like a typical product launch.
Once promotions become part of the ad surface, the system needs end-to-end ownership of the offer lifecycle. That includes sourcing, validation, eligibility evaluation, application, and retirement. Each stage introduces its own failure modes, and those failure modes compound if they are not designed explicitly.
From an engineering perspective, the work focused on building a pipeline that could ingest promotional signals from multiple inputs, normalize them into a consistent representation, and enforce strict correctness checks before anything reached users. Coverage and speed were important, but they were never allowed to outrank confidence. The system had to prefer restraint over aggressiveness.
One of the hardest aspects was coordination. Advertising delivery, commerce signals, data ingestion, and governance teams historically operated with different priorities and abstractions. Building this system required aligning them around a shared definition of what constituted a valid promotion and what conditions justified automatic application.
That discipline paid off quickly. The platform was able to support rapid advertiser adoption and scale revenue contribution far faster than historical ad format rollouts, while maintaining stability under load. The speed was visible. More importantly, the system held up as usage accelerated.
Trust relocates.
When users manually enter a promo code, errors feel personal. When the system applies a promotion automatically, errors feel institutional. Users assume the platform understands the rules better than they do. That assumption raises the cost of being wrong.
From a systems standpoint, this means the platform must be able to explain its decisions internally. It must know why a promotion was applied or withheld. It must handle ambiguity conservatively rather than optimistically.
In practice, that meant designing eligibility and expiration checks that were stricter than what many advertisers initially expected. It also meant building fallback behavior that preserved the ad experience without forcing a questionable promotion through. The objective was not maximum discount application. It was predictable behavior. Once advertisers and users trust that the system behaves consistently, adoption follows naturally.
It reinforced the need for first-party confidence. As measurement defaults became less stable and privacy expectations tightened, relying on downstream correction became riskier. Recent industry surveys show that more than half of digital marketing leaders report declining confidence in downstream attribution models, making it harder to rely on post-hoc correction when systems get things wrong.
If a promotion is misapplied, you cannot assume attribution models or post-hoc analysis will smooth it out.
That environment favors systems that can make deterministic decisions at the moment they matter. It also increases the importance of auditability. When something goes wrong, you need to trace why the system behaved the way it did.
In that sense, promotional infrastructure is not just a growth lever but a stability mechanism. It reduces dependence on probabilistic correction and shifts accountability earlier in the flow.
By designing for partial certainty.
No system has perfect information. The mistake many teams make is assuming automation should always act. In high-stakes environments, the more important capability is knowing when not to.
That principle influenced how scaling was handled. Monitoring focused on confidence signals, not just throughput. Rollback paths were treated as core functionality, not emergency measures. Ambiguous inputs were deprioritized rather than forced through.
This approach is consistent with how I think about large-scale decision systems, including in my invited ICESIC 2026 Conference keynote and session chair role. In that keynote, I examine how systems processing high-volume signals in real time preserve consistency and reliability by separating ingestion, evaluation, and action into distinct layers. The takeaway is not that systems should be faster, but that they should be more deliberate about when and how they act.
This approach helped the platform absorb rapid growth without introducing brittle behavior. Stability during expansion is rarely about raw performance. It is about maintaining decision quality as complexity increases.
Ads are becoming transactional surfaces. That trend will continue.
As systems take on more responsibility for applying prices, discounts, and recommendations automatically, the line between advertising infrastructure and financial infrastructure will keep blurring. At the same time, roughly one-quarter of large organizations are already operating AI-driven decision systems in production, expanding the blast radius when automated decisions behave unpredictably.
The platforms that succeed will not be the ones that optimize the hardest. They will be the ones who behave predictably under uncertainty.
Conversion-grade advertising is not about making ads smarter. It is about making them accountable.
