Can ChatGPT Find That Product? The Architecture Question Nobody Budgeted For

Imagine a shopper who asks an AI assistant to recommend project management software for a 10-person team, and 3 names come back within 2 seconds. None of them belong to the company that has quietly served that exact niche for a decade.

The deal disappears before a phone rings or a demo gets scheduled, because whatever holds that company’s product catalog and pricing logic was never built with an application modernization strategy in mind, and a machine cannot recommend what it cannot read. Call it an address problem: the delivery service simply does not recognize the street.

Retailers noticed the shift first, since commerce makes it visible in ways an invoice line never does. Shop-assistant features now sit inside nearly every major chat app, and behind each one runs a plan for modernizing legacy applications that few companies wrote down before 2024.

Old inventory systems answer to a handful of trusted internal tools, tuned over years to talk to people who already know where to look. They were never built to answer a question typed by an algorithm acting on someone else’s behalf. Ignore that gap long enough, and a smaller rival with a cleaner data layer starts appearing first, every single time, for reasons that have nothing to do with better products.

The pattern is not limited to shopping carts. A B2B buyer researching payroll software runs into the same filter an engineer hits while comparing embedded databases, and a hospital system vetting scheduling tools fares no better. Increasingly, the first pass at that research happens inside a chat window, not a search results page. Whatever gets summarized there becomes the shortlist; whatever doesn’t gets erased before a human ever weighs in.

Numbers back the urgency. Adobe’s Q2 2026 traffic report found AI-referred visits to US retail sites grew 393% year over year in the first quarter, with those same visitors converting 42% better than shoppers arriving through ordinary channels. A year earlier, that same traffic converted worse, not better. Something flipped, hard enough that most marketing budgets have not caught up.

The Silent Filter

Search used to be forgiving. A shopper typed a vague phrase, scrolled past 3 irrelevant listings, and eventually found what they wanted anyway, patience doing the work that good data should have done. An AI agent skips that patience entirely. It reads what it can parse and skips what it can’t, rarely explaining the omission to anyone. No feedback. No appeal process.

What gets an agent to skip a page, usually, comes down to a short list of familiar sins:

• Product data locked inside a database only a legacy front end can query
• Pricing and inventory that update on a nightly batch job instead of in real time
• No public API, so an outside system has nothing to call
• Pages built almost entirely in client-side scripts that a crawler never executes
• Descriptions written for a salesperson’s ear, not for a machine matching intent to attribute

Trust compounds, too, in a way that rewards patience. An agent that finds accurate stock levels once tends to query that same source again, the way a person returns to a store that never runs out of what the sign promised. Break that trust with stale data even once, and the agent quietly routes around the mess next time, offering the shopper somewhere else entirely.

What the Old Code Won’t Say

Ripping out a mainframe is rarely the answer, and few serious technologists still argue that it should be. McKinsey’s research on enterprise architecture in the agentic era makes a related point: legacy systems hold years of business logic worth keeping, and the practical path runs through APIs that expose what already works to newer layers built for machines to query. An insurance company’s decades-old underwriting engine, for instance, does not need a rebuild so much as a translator sitting on top of it. McKinsey calls the connective tissue an “agentic mesh,” an orchestration layer that lets new AI tools talk to old systems without anyone touching the original code.

Firms like N-iX have spent the past few years doing exactly that kind of translation work for clients who never planned to touch their core systems until an AI agent forced the question. An app modernization roadmap built around API exposure and clean, structured product data tends to look unglamorous next to a flashy front-end redesign. It also happens to be the part that decides whether a machine ever finds the product at all.

Not every fix requires new infrastructure. Sometimes it’s a matter of publishing a feed. Other times, it’s rewriting a product description so an algorithm can match “waterproof hiking boots for wide feet” to an actual SKU instead of guessing.

A Revenue Line, not a Line Item

Marketing teams have been slow to treat this as their problem. Forrester’s 2026 survey of B2B marketing leaders found that only 24% planned to work on their content’s visibility inside AI search and generative tools, even as 69% of digital strategy leaders said their companies were already piloting or fully deploying visibility efforts on AI answer engines. That gap between intent and action is where budgets quietly bleed out.

None of this shows up on a standard marketing dashboard, since most attribution models still assume a click arriving from a search engine rather than a citation buried inside a chat response. A campaign can hit every target on paid social while losing, invisibly, everywhere an AI assistant now decides what the buyer gets to see.

Framing an app modernization strategy as an IT expense misses what is actually being purchased. The real asset is a seat at a table that increasingly has no human sitting across from it, deciding who gets recommended and who gets skipped. A CFO who once measured this work in server costs and maintenance hours now needs a second column: what does it cost to be invisible to the systems doing an increasing share of the world’s shopping?

Conclusion

The question in the title is not rhetorical anymore. Somewhere today, an AI agent is reading a competitor’s clean API and skipping past a company whose product data still lives behind a login page built in 2011. Fixing that is not a marketing project, and it is not glamorous work.

It’s plumbing, mostly, done well enough that a machine can finally find the door. The businesses that get there first will not need to explain why; the ones that don’t will spend years wondering where the customers went.

Related Posts