Executive Summary
An AI chatbot does not know your products. Its ability to answer a question depends on whatever records you give it—give it bad data, and it will give you bad answers.
Conversational commerce is product discovery, advice, and purchasing through natural-language dialogue. It can happen on your own site, in a messaging channel, or through an external AI assistant.
Conversational commerce agents answer questions by first interpreting the question, then retrieving structured product facts, price, stock, and delivery information from live commerce systems. If those sources are incomplete or unreliable, the answer will be too.
- Your catalog is the source of errors. An assistant cannot construct a reliable answer from missing or stale product facts, no matter how good the model is.
- Incomplete and incorrect answers both have costs. A hedged answer (vague or likely wrong) loses the sale. A wrong answer creates a return and damages the shopper’s willingness to believe you next time.
- Fixing the data pays dividends in traditional search and AI-enabled commerce. Fixing data improves search and merchandising now, whether or not you implement conversational commerce or if users leverage their own AI search tools to find you.
Automotive parts make an unforgiving example: a shopper asks whether a set of front brake pads will fit her 2016 Focus 1.5 TDCi, and whether they will arrive before Friday.
If the product record is complete, the assistant answers in full: yes, the FP-2216 front pad set is a confirmed fit, £42.99, 14 in stock, next-day if she orders before four, and the matching front-disc set is there too.
If the product record is incomplete, as it is for most catalogs, the same assistant running the same model can see only the price. It offers the same pad set but cannot confirm fitment for her vehicle or see live stock, so it asks her to check a compatibility chart or call. In this article, product data is shorthand for two things: the structured catalog, and the live commerce services that supply price, stock, and delivery.
Illustrative scenario. Same assistant, same model, two catalogs.
The natural reaction to the second answer is to blame the model, so teams switch providers and spend months tuning prompts. The disappointment usually survives because nobody changed the data underneath.
A language model cannot retrieve a field that does not exist.
The solution is “grounding,” the process of attaching your AI model to verified and complete data inputs.
Why Grounding Is Capped By Your Product Data
Grounding is your catalog, real products, and Live data.
Grounding is not a feature you switch on. Its usefulness is limited by the completeness and freshness of the systems underneath it. For product questions, grounding means the assistant answers from live catalog and commerce data rather than relying on its training memory. That is the correct design. It also clearly indicates that the quality of your data determines the quality of the answer.
In the first case above, the one with grounded data, the AI assistant helps make a sale. In the second instance, with incomplete (“ungrounded”) data, it sends your customer away to phone a call center. We all know that customers are unlikely to place that call, and AI agents cannot. In fact, an AI agent faced with that result will simply ignore you and move to a competitor whose product record it can verify.
The hedged answer (the one requiring you to confirm if the part works) leaks the sale. On the flip side, if the agent had come back with confidence but provided a wrong answer (e.g., the wrong part was offered), this confidently wrong answer costs the sale, the return, and something harder to win back: the shopper’s willingness to believe the next answer from you. A customer who has been burned once double-checks you, and a customer who double-checks you is already comparing you. And increasingly, the one doing the comparing is not a person at all—fail an AI enough, it will learn not to trust you (or recommend you) at all.
You Are Now Selling To People And AI Agents
You now answer to two buyers: the shopper reading your storefront, and the AI agent reading your catalog on that shopper’s behalf.
What we know from working with many retailers on conversational commerce is that the “conversational layer” (the chatbot) is the least critical part of conversational commerce. Everything that decides whether it works sits underneath it, in systems nobody even talks about.
The presence of AI agents in the commerce journey matters more every quarter, because even if you are not rolling out AI tools on your own site, they are being used already. AI tools increasingly mediate product discovery, comparison, and shortlisting. In an agent-assisted journey, software may read your catalog, resolve compatibility, and check stock before the shopper ever reaches your storefront.
Two buyers, one factual foundation.
The reassuring part is that both people and AI agents assisting people depend on the same factual foundation: complete attributes, clean taxonomy, accurate stock, and machine-readable structure. Build that once and build it properly, and you can serve both with a single piece of work. But knowing what to fix and what to expose is nuanced.
What To Publish, Expose Selectively, And Keep Private
If machine-readable data is this valuable, why expose yours at all?
For an assistant you own, running on your own site and app, the answer is governance. Your catalog does not become a public asset because a bot can read it; it becomes one when nothing stops a script walking it question by question: does this fit a 2015 Honda Civic EX, does this fit a 2016 Honda Civic EX, does this fit a 2016 Honda Civic LX, until the fitment table has been rebuilt from the outside. Rate limiting, authentication, and query budgets prevent that. All three are ordinary engineering.
The harder judgment is what kinds of data you publish deliberately, making it open to external agents, feeds, and protocol endpoints you do not control.
Publish, expose selectively, keep inside.
Let’s break down the discussion by data asset type:
- Public product facts exist to be seen. Hold them back from agents, and you do not stay protected; you simply stay out of the answer.
- Operational facts are the ones that need judgment. Prices can be contract-specific, stock can be channel-specific, and some attributes are licensed from suppliers on terms you do not own.
- The third asset, your decisioning intelligence, is what lets an assistant built on your own systems do what a general-purpose agent never can: rank, personalize, and recommend from data only you hold.
Exactly where operational data ends and decisioning intelligence begins is still unsettled, and that boundary deserves its own article. For now, the point is simpler: you are managing three assets rather than one.
Where A Shopping Assistant Sits
Think of conversational commerce as a ladder you climb from the bottom, not a product you drop in at the top. If you buy it too soon, you are missing critical steps on your ladder.
The assistant is rung three; most programs start there.
What The Assistant Must Get Right
Even if the data is right, the conversational commerce piece itself is not trivial.
Once the foundation exists, the assistant still has real work to do, and the builds that succeed are ones that avoid hedged answers altogether.
To avoid ambiguity, the assistant should ask for clarity when needed. For example, when a required field is ambiguous, such as which engine variant the shopper drives, the AI agent should ask one clarifying question instead of guessing. To improve answer clarity, the agent should have access to live systems (vs an index) for volatile fields such as price, stock, and delivery promises. Further, the AI agent should state what it does not know, rather than filling the gap with plausible language. And when a question exceeds what the data can support, it should hand off cleanly, routing a warranty question to a person rather than generating a plausible paragraph.
Yet even a disciplined assistant is only as strong as the data beneath it, and we have seen what solid data delivers on a live storefront.
Conversational Commerce Delivers 2.4x Revenue For GSF
At GSF Car Parts, a UK automotive parts retailer whose platform we rebuilt on Adobe Commerce, the product data, search, and merchandising work moved real numbers.
Product data, search, merchandising, and the storefront all changed together, so the results cannot be pinned on any one ‘fix.’ Still, all relied on the same strong foundation: a structured catalog, fitment relationships, and live stock that a shopping assistant can read. Product data pays off now, before autonomous buying does.
The speculative claim is that autonomous agents will be buying on your customers’ behalf at scale, very soon, and that you must prepare for that precise future today. The evidence says slow down. In a Gartner survey of 322 US consumers in January 2026, willingness to let AI make purchase decisions topped out at 11%, and that was in the lowest-stakes categories such as personal care and household supplies.
In a separate Gartner survey of 846 US consumers run through November and December 2025, 54% of shoppers who used generative AI while buying said they had to double-check the accuracy of everything it told them, and 62% said the information wasted their time. Kate Muhl, the Gartner analyst behind the research, puts it plainly: accuracy is now a brand issue. An assistant that has to be fact-checked is not a convenience; it directly acts against you (whether you own that agent or not).
Structured, complete, machine-readable product data compounds in value whether or not the agents-as-autonomous-buyers wave arrives. It makes search, merchandising, and a competent assistant better today, and if agents do arrive in force, the very same records are what let you be chosen instead of skipped. You were going to do this work anyway.
That is the unglamorous, load-bearing case for starting with your product data before you buy the fun chatbot.
Frequently Asked Questions
Conversational commerce is shopping through natural dialogue instead of menus and filters. A shopper describes what they want in plain language, and an assistant grounded in the retailer’s live catalog asks clarifying questions, recommends products, and supports the purchase. The same conversational layer can also serve AI agents shopping on a customer’s behalf.
Usually because of the product data, not the model. A grounded assistant can only retrieve fields that exist in the catalog. If fitment, stock, or delivery data is missing, stale, or stored as free text, the assistant either refuses to answer or guesses. Fixing the records removes a large and expensive class of failures.
Grounding means the assistant answers from your live product data at the moment of the question, rather than from the model’s training memory. That is the correct architecture, and it makes your data quality the ceiling on answer quality.
Not necessarily. Core attributes, compatibility claims, descriptions, and customer-facing availability need to be machine-readable. Contract pricing, channel-specific stock, and delivery logic should be exposed selectively. Behavioral data, margin logic, and recommendation intelligence remain governed inside your environment.
Complete structured attributes, a clean taxonomy, accurate real-time stock and pricing, and explicit product relationships such as fitment and compatibility. All of it should be machine-readable and kept current by pipelines. These are the fields the assistant retrieves when a shopper asks a question, and they are the same fields external AI agents read.
Not necessarily. A conversational assistant can often sit on top of your existing commerce platform, provided the platform can expose reliable product, pricing, inventory, and order capabilities. The prerequisite is usually not a new platform but a catalog in good enough shape to answer from.
Prepare the data, not the hype. Gartner’s January 2026 survey found willingness to let AI make purchase decisions topping out at 11% of US consumers, so fully autonomous buying is early. But the structured product data agents will need is the same as what improves search, merchandising, and on-site assistants today, so the work pays off either way.
Where to Go Next
This article stops where your own storefront ends. The buyer that never visits it at all, the external agent reading your catalog through feeds and protocols, is covered in AI agents becoming customers, and the automotive version of the data problem is in the product data quality piece.
Find the Data Gaps that Would Make a Shopping Assistant Fail
The AI Readiness Assessment scores your business on four pillars: AI Visibility, Data Foundation, Customer Experience, and Operating Model. Each is rated from L1 Invisible to L5 Agent-Native, and the result identifies the highest-priority gaps to address first.
