by Greg Madison
The logic behind loyalty programs has always been precise and straightforward: collect more data… to better understand the shopper… to target them more narrowly and effectively… to generate more incrementals and extract more margin through increasingly granular personalization… to collect more data… and so on.
No surprise, then, that the biggest and smartest retailers invested heavily – an estimated $15 billion in 2025 – in customer data platforms, advanced analytics and, more recently, AI-driven targeting engines. Major CDP and loyalty vendors, like SalesForce, Adobe, and Oracle, just to name a few, have been reporting double-digit growth in retail-loyalty and personalization spending since 2023. All of it designed to turn customer loyalty into a very defensible competitive moat, a market fortress.
But any “fortress” is only as good as the stuff it’s made of. And in far too many cases, these defensive positions are built on extremely iffy, dirty data.
Too often this data is really just probability and inference. Retailers infer income from ZIP codes and household size from basket patterns. They guess at a life stage from purchase proxies. Making matters more complicated, third-party enrichment layers fill in gaps with assumptions that look tidy in dashboards but fray badly in the real world.
How badly? Well…
Speaking personally, I’ve used my wife’s loyalty card at the grocery store. I’ve used my in-laws’, my own parents’, my loyalty card, my boss’s. In one case – and this is true – I even used my poker buddy’s card. I can’t imagine a marketing outfit making much sense out of those patterns. And yet it’s packaged and sold along with all the rest of the data.
Hard research from Alyson Lloyd and James Cheshire, and Tim Raines and Paul Longley, and others show I’m far from alone in my “scattershot” approach to using loyalty cards. Their papers show grocery loyalty data is uniquely noisy, with shared households, pooled cards, and proxy usage far more common than in most other retail categories.
What’s more, researchers have repeatedly shown that inferred demographic attributes routinely miss the mark at the individual level, particularly when built from ZIP codes, basket proxies, and third-party enrichment.
It’s starting to make headlines – literally.
Consumer Reports reported recently that a male, college-educated six-figure earner requested his customer profile, only to find out the marketers had him pegged as a female high school graduate who earned about half the income. And who was likely to want to go on a cruise for some reason.
It’s Hanlon’s razor: Never attribute to malice what can be adequately explained by stupidity. Or inertia, I guess. To be clear, I don’t mean to suggest there’s malice or malfeasance going on here – no willful wrongdoing. The problem is structural; dirty data is inherent in how these programs work. It comes with the territory.
Certainly, data quality is only one of the risks facing loyalty programs. Regulators, particularly at the state level in places like New York, California, and Colorado, are asking pointed questions about usage and fairness, customers are uneasy about how much of their information is “out there,” and whether they’re really benefiting. Across the industry, retailers are increasingly favoring loyalty mechanics that are easier to explain, easier to scale, and easier to defend — even when more granular personalization is technically possible. Kroger is playing up fuel rewards, spend thresholds, and basket-level incentives over risky data-dependent incentives – and this is despite having the data to go further. Walmart+ is focusing on value – delivery, speed, convenience – and sidestepping potential price discrimination.
These are smart strategic moves but, as I noted, they’re leaving data on the table, untapped, because of the real concern that at least part of it might be questionable. So why risk it?
The better question is becoming: How to use it?
Let’s take a look at dirty data, the big risks it poses, and, importantly, what kind of cleaning solutions might be at hand.
Why Dirty Data Is So Dangerous
Regulators don’t care whether a model meant well. They want it explained, audited, and justified. In this way, dirty data turns precision targeting into a compliance liability.
And then there’s trust risk. Shoppers don’t experience dirty data as some kind of modeling error. They experience it as “creepiness” or, even worse, unfairness.
When an offer feels mismatched – or oddly specific in the wrong way – it naturally triggers suspicion. “Why makes them think I live there? What gave them the idea I earn that? Why didn’t I get that deal?” Once shoppers lose confidence in the fairness of loyalty systems, the entire construct weakens. Loyalty stops feeling like a reward and starts feeling like an ominous black box.
Third, there’s operational risk. Dirty data doesn’t stay in the marketing department. Very often it bleeds into execution in the store. Store teams field complaints about missing or inconsistent offers; they can’t answer the awkward questions and customer service absorbs confusion. Promo performance becomes harder to forecast because targeting logic is built on shaky foundations. In a labor-constrained environment, there’s little tolerance for explaining algorithms that don’t make sense on the floor.
And finally, and perhaps most dangerous of all in this environment, there’s margin risk. Precision targeting only works if the precision is, well, actually precise. When it isn’t, retailers over-incentivize shoppers who would have purchased anyway. They miss those who actually need and would respond to an offer. Margin – critically low at the best of times – leaks out across untold numbers of transactions.
In the end… ouch: the systems designed to make pricing and promotions more efficient often introduce more variability instead.
This is one of the core reasons retailers are de-risking loyalty and personalization strategies. It’s clear at this point that confidence in data has been overestimated.
Hyper-personalization, in the real world, requires clean inputs, and quality data. Grocery doesn’t have it.
Yet…
Where the Industry Is Looking for Dirty Data Help
Here’s where the story turns forward-looking – and, it has to be said, more than a little ironic. AI finds itself at the center of a firestorm in grocery right now, raging around Instacart’s algorithmic “price test” experiments and Wegmans’ biometrics pilot programs in New York.
And in half-baked deployments, AI intensified the wave of “personalization” that increased the gathering of questionable, sometimes dirty, data.
But as of early 2026, large language models (LLMs) and next-generation AI systems are beginning to show promise in areas traditional analytics struggled with: anomaly detection, context resolution and mapping, data reconciliation, and outright explainability.
Enterprise platforms from vendors such as Palantir, Databricks, Snowflake, Amperity, and IBM are increasingly being used to surface data inconsistencies, flag low-confidence inferences, and introduce explainability layers that traditional analytics lacked.
Instead of blindly trusting inferred attributes, LLM-powered systems can flag inconsistencies, question improbable assumptions, and surface uncertainty rather than hiding it. Needless to say, it might take large teams of people months, if not years, to crack the dirty data problem. These newer systems are better able to help retailers distinguish between high-confidence behavioral signals and low-confidence demographic guesses.
In other words, they can separate valuable marketing wheat from probabilistic chaff.
This doesn’t mean retailers will suddenly achieve perfect data; that’s probably unrealistic. But it does mean systems may become more honest – about what they know, what they don’t, and, most importantly, what shouldn’t under any circumstances be used to drive decisions.
The cleaner data that AI helps produce will let human managers know exactly what’s reliable enough to act on.
The data quality question – the problem, really – is a strategic one. It helps explain why loyalty is being simplified and why personalization is being restrained. Why pricing algorithms are under scrutiny. And why trust, which was once assumed, now has to be actively protected.
