IA in an AI World: Retail Finally Has the Tools. Now It Needs the Architecture

I spent thirteen years working in nearly every facet of omnichannel retail, from a small department store in a local mall to a fourteen-floor flagship. Across every one of those environments, regardless of size, location, role, or annual sales figures, I kept running into the same problem.

Retail has an information problem, and it’s disguised as a merchandising one.

Where does this product live? What do we call it? How does a customer find it without already knowing it exists? These questions predate the internet. They predate e-commerce. Walk into any department store and you’re navigating a taxonomy that someone built, labeled, and argued over, a set of decisions that determine whether you leave with what you came for, or empty-handed.  There’s a name for this area of expertise: information architecture (IA), the discipline of organizing, labeling, and structuring information so people can find what they need. Retail has been practicing it for a century without calling it that and often without the customer in mind.

The hierarchies still found in many department stores’ reporting – department, class, subclass – started life as charts of accounts. They existed so buyers could plan assortments and accountants could track margin; the shopper was never the audience.  Because decades of sales history hung off every category code, the structure couldn’t change without breaking the reporting. If you’ve ever sat in a meeting and heard something to the effect of “we can’t restructure the categories, it’ll break our reporting,” you’ve watched a hundred-year-old constraint dictate your modern-day experience.

Moving online didn’t magically fix this issue. A website’s navigation is its taxonomy exposed with no floor layout, no signage, no associate to step in when the structure fails a customer. The first reckoning came in the early 2000s with faceted filters, the “Color: Blue (47)” pattern that Endeca pioneered and that now sits on virtually every retail site. Facets made data quality impossible to hide.  A Sleeve Length filter is worthless if only a third of the shirts carry the attribute. Nearly every search project of that era became a data-cleanup project. Twenty years and several generations of tooling later, that’s still true. The tools kept improving. The data underneath them never did.

AI is just the newest technology to stress-test that data, and the harshest yet. But it’s also the first one that offers a way out. Fixing the data was always the unaffordable part; enriching tens of thousands of SKUs (sometimes hundreds of thousands) by hand is why retailers lived with broken search for two decades. AI collapses that cost, not by replacing the discipline of information architecture, but by giving IA practitioners tools that make their work faster, more evidence-based, and more impactful than ever before.

Here’s where that impact is being felt.

Building Taxonomies That Reflect How Customers Actually Think

Here’s the fundamental tension at the heart of retail IA, shoppers don’t think in categories, they think in situations.

A shopper rarely walks into a department store, physical or digital, thinking of a category. The category is where they end up, the last step in a translation they don’t even notice making. What they start with is a situation: they think “I’m in need of comfy work pants,” “I have a wedding in three weeks and nothing to wear,” “It’s getting cold and I need something for the commute.” The occasion, the problem, the feeling, that’s the organizing principle in the customer’s mind. The product category is an abstraction they must translate before the store can assist them with their needs.

Traditional retail is built the other way around. It’s organized by what the product is, by the logic of the buying team, the supplier relationship, the warehouse structure. Joggers live under Ready to Wear/Activewear/Bottoms/Joggers because that’s where the buying team files them. It makes sense internally. It makes no sense to the customer who types “comfortable pants for working from home” and doesn’t find what they are looking for, not because joggers aren’t readily available, but because the entire category logic wasn’t built for that intent. They weren’t browsing Activewear. They were never going to browse Activewear, and they didn’t even know they were called joggers.  The problem isn’t the placement. It’s the architecture.

This is where AI gives IA practitioners something genuinely powerful. By analyzing search queries, browse paths, purchase behavior, and return patterns, AI can surface the mental models customers are actually using, the situations, occasions, and intents that drive their decisions, and make them visible to the people responsible for building the structure. The taxonomy stops reflecting how the business organizes its inventory and starts becoming a reflection of how customers think about and search for what they need.

The practitioner’s role isn’t diminished by this, it’s sharpened. The data surfaces where the architecture is failing. IA judgment decides how to fix it. What gets its own category versus what becomes a filter? When does a taxonomy change need to ripple through the navigation? How do you restructure a category without breaking the mental models of existing customers? These are strategic decisions. AI conducts the analysis and surfaces patterns no human team could find at that scale. IA decides what the architecture does about them.

Closing the Vocabulary Gap at Scale

Ask a retailer what their biggest search problem is and the answer is almost always the same: customers don’t search the way we label things. You call it Outerwear. They type “coat.” You call it Home Fragrance. They type “candles.” You file it under Footwear/Boots/Chukka. They type “ankle boot.”

Bridging this gap is one of the core tasks of information architecture, specifically, the discipline of controlled vocabulary: building and maintaining the mappings between internal labels and customer language. Traditionally, it required human curation.  Someone compiled synonym lists, reviewed search logs manually, and refreshed the mappings periodically. The work was never finished, always out of date, and perpetually under-resourced.

Stores had a fix for this gap: the associate. A customer could ask for “something for my daughter’s graduation” and the associate would walk them to the appropriate section of the store.  E-commerce replaced the associate with a lexical search box that only matched exact words. Shoppers figured that out fast. They learned to take their real questions to Google and bring back keywords the site could handle and when Google pointed them to a competitor instead, they went. The search box didn’t just fail to answer questions; it walked customers right out the door.

This is the part of the story where AI stops being an abstraction. Semantic search powered by natural language processing, the branch of AI that interprets human language as it’s actually spoken rather than as exact keywords, has transformed what’s possible here, and IA is the discipline best positioned to govern it.  Modern search handles “couch” and “sofa” without anyone building a synonym table, but that’s the easy part. It also understands that “going out top” is a style intent rather than a product category, that “something for a hot, summer wedding” should return occasion wear in the right formality register and for the right region rather than a literal keyword match, and that “dupe” means “affordable alternative to a luxury product.” (Sit with that last one for a moment. “Dupe” meant nothing in retail six years ago. No merchandiser added it to a synonym table; no taxonomy committee approved it.  The system learned it because customers kept saying it.) A century of retail asked shoppers to decipher its vernacular. This is the first technology that quickly allows the retailer to decipher the shoppers.

The controlled vocabulary problem, at the level of synonym mapping, is now largely automated, but automation without governance creates its own problems. When the vocabulary layer is invisible and algorithmic, who ensures the mappings are accurate? Who catches the cases where semantic search is confidently returning the wrong results? Who ensures that the language the system is learning from doesn’t encode the biases of a skewed historical dataset?

The answer is IA practitioners, working at a higher level of abstraction than they used to, auditing outputs rather than building synonym tables by hand but exercising exactly the same core judgment.  Does this label serve the customer? Does this mapping reflect how people actually think?

Using Search Behavior as a Continuous IA Audit

One of the most powerful things AI puts in the hands of IA practitioners is the ability to hear what customers are telling them: continuously, at scale, without a qualitative research study that is often-times very expensive.

Every failed search is a signal. Every zero-results page, every reformulated query, every session that ends in a bounce from a category landing page: these are customers encountering a structural failure in real time and telling you exactly what it is. The challenge has never been a lack of data. Retail sites generate search log data at enormous volume. The challenge has been analysis: the logs are too large to process manually, the patterns too buried to surface quickly enough to be actionable.

AI makes that analysis tractable. AI can process search logs continuously, cluster failed queries by type, and surface structural problems with specificity.  Customers searching “jacket for rainy commute” get nothing usable, because the catalog knows the jacket is a “soft-shell” and knows it’s “water-resistant,” but nothing in the data connects either fact to a rainy commute.  The situation exists in the customer’s head and nowhere in the taxonomy.  Customers who search “ethical cashmere” are arriving at zero results, which is both a navigation failure and a potential signal about product demand. A spike in searches for a term that didn’t exist six months ago is an early warning that a new category is forming and the taxonomy hasn’t caught up.

For IA practitioners, this transforms the evidence base available for structural decisions. Problems that would once have required a commissioned research study to surface are now visible in the data, if you know what to look for and how to translate the signal into structural action. That translation is the IA practitioner’s task. AI surfaces the problem; the practitioner redesigns the architecture.

The Decisions That Still Belong to People

It would be a mistake to read any of the above as a case for removing human judgment from retail IA. The opposite is true: AI absorbs the computational work and elevates the practitioner to the decisions that require strategic and cultural thinking, which are also the decisions were getting it wrong costs the most.

Should “Plus Size” be a top-level department or a filter applied across categories? AI can tell you which approach produces more conversions. It cannot tell you whether a department structure that routes plus-size shoppers away from the main site is inclusive or othering, whether it reflects the brand’s values or contradicts them, whether the conversion-optimizing choice is the right one for the long-term customer relationship. That is a values question, a brand question, and a structural one and it belongs to a practitioner.

Should a luxury retailer organize its navigation by price point or by occasion? The data will contain the necessary insights to generate an optimal answer. But whether leading with price is consistent with the positioning of a brand that has spent decades signaling that its customers don’t check price tags, that requires understanding the brand, the customer, and the cultural register simultaneously. No pattern-matching model is equipped to hold all three.

The same applies across the structural decisions that define a retail site’s architecture: what earns its own top-level section versus what lives as a filter, what gets a dedicated landing page versus what lives in search, when a category is large enough and distinct enough to be split, what the hierarchy says about the brand’s priorities not just its products. These decisions have data inputs. They are not exclusively data-driven decisions.

The most valuable IA practitioners in retail right now are the ones who understand AI well enough to work with it, to read what the models surface, to govern what AI optimized, and to make the structural calls that fall outside the model’s competence. That combination is rare, and it is where the discipline is heading.

Why IA Practitioners Are More Important, Not Less

The case for information architecture in retail has never been stronger precisely because AI has raised the stakes of structural decisions, while also making their consequences more visible.  It has also changed the math. For twenty years, fixing retail’s data was too expensive to justify, so nearly everyone lived with it, which meant a broken catalog was a level playing field.  Those days are over.  AI has made doing it right affordable enough that it’s now a choice, and the moment it becomes a choice it becomes a differentiator.  The retailer who governs its architecture creates a search that understands customers, navigation that adapts, and an experience that feels coherent.  The retailer who doesn’t is losing customers to the competitor who did, one abandoned search at a time. 

When navigation adapts dynamically, someone has to define the rules it adapts within. When semantic search automates vocabulary mapping, someone has to audit the outputs and govern the system. When AI surfaces structural problems at scale, someone has to diagnose the root cause and redesign the architecture. When personalization engines optimize for conversion, someone has to ask what they’re doing to the coherence of the experience over time.

That someone is an IA practitioner.

So where does a retailer start?  Every era’s search technology has outrun the data underneath it.  Endeca’s filters were only as good as the attributes behind them, and today’s semantic search is only as good as the catalog it reads. The difference now is that the technology excuse is gone.  Search that understands customers is available to any retailer off the shelf, yet independent benchmarks keep finding major sites failing basic queries. When a fix is affordable, available, and provably profitable but still hasn’t happened, the problem isn’t technical, it’s organizational.  Merchandising owns the products, IT owns the platform, marketing owns the copy, and no one seems to own the taxonomy.

Three moves change that.  Give the taxonomy a named owner with cross-functional authority; a taxonomy owned by everyone is owned by no one. Treat search performance as a data-quality metric, not a platform feature, and route failed searches directly into taxonomy governance as a standing input; the search log is a requirements document written by customers in their own words, and at most retailers it goes unread.  Lastly, govern the machine tagging.  LLM (Large Language Model) auto-enrichment makes it economically possible to describe products in customer language at scale, but ungoverned, it simply creates the old mistakes faster. Governance matters more in the AI era, not less.

Thirteen years in department stores taught me that customers almost never ask for what the sign says. Retail’s information problem was always a translation problem, and for the first time, the translating can run in the customer’s direction. But the technology only closes the gap if the architecture allows. Someone still has to own the structure, govern the language, and make the calls that data can’t. That someone has never mattered more.