Agentic Commerce Starts With Product Understanding
Updated: Aug 19
Agentic commerce starts before checkout. AI is already helping shoppers research products, interpret their needs, narrow the options and decide what to buy. For ecommerce brands, that makes product understanding a more immediate priority than autonomous purchasing. AI needs access to clear, specific and supported product information so it can understand what a product is, who it suits, where it performs, what its limitations are and when it belongs in a buyer’s consideration set.
Most discussion around agentic commerce focuses on AI completing purchases. For ecommerce brands, there is a more immediate product problem.
Before an AI system can recommend something, it needs enough information to assess whether the product fits the buyer’s needs.
OpenAI is already building ChatGPT shopping around product research and decision-making. Shopping research helps people describe what they need, refine their requirements, assess options and work out what to buy. Google’s Universal Commerce Protocol covers the commerce process from product finding and consideration through checkout and post-purchase support.
The commercial signals are starting to appear too. Shopify says AI-referred orders grew nearly 13 times year over year in Q1 2026, while referral sessions from AI chatbots grew more than eight times. Reuters, citing Adobe Analytics, reported in August that US visitors referred by AI services generated 41% higher revenue per visit than shoppers arriving through traditional channels.
Those figures describe only part of what is changing for ecommerce teams. Product research, evaluation and recommendation can now happen before the buyer reaches the product page.
Product choice is the harder part
Consider someone shopping for a commuter e-bike. They are unlikely to care only about price, colour and motor size. Their actual ChatGPT request might be closer to this:
“I need a commuter e-bike for a hilly 18km round trip. I live in an apartment, need to carry it up one flight of stairs, and want a removable battery. Which model is the best fit?”
A well-maintained ecommerce product record may already contain the title, description, brand, product identifiers, price, availability and variants, alongside detailed specifications such as motor power, battery capacity, published range, bike weight, frame size, wheel size and braking system.
That can be a very complete product record.
It still may not resolve the buyer’s request.
The buyer wants to know whether the published range is realistic on a hilly daily commute, whether the bike is manageable to carry up a flight of stairs, how practical it is to remove and charge the battery indoors, and how the bike performs in normal city use.
Some specifications contribute directly to those questions. Weight matters if the bike needs to be carried. Battery capacity and published range matter for the commute. A removable battery matters for indoor charging.
But those specifications do not necessarily tell us whether the bike is a good fit for this particular buyer and commute.
Reviews may add another layer.
Owners might repeatedly mention strong hill climbing, noticeably lower range in cold weather, difficulty carrying the bike despite its published weight, or how easy the battery is to remove day after day.
Technical documentation may answer other parts of the question. Product or support teams may also know things that are not clearly represented in the catalog.
The information needed to answer the buyer’s question may be scattered across the product page, specifications, technical documentation, reviews and internal knowledge.
Some of it may be clear and well supported. Some may be partial or contradictory. And for some questions, the business may simply not have enough evidence to give a confident answer.
That's important when an AI system is comparing several products and trying to determine which one best fits the buyer’s constraints.
Most product records are better at describing products than assessing fit
Ecommerce catalogs were built to support websites, filters, feeds, marketplaces and internal systems. They still need to do all of those jobs well.
Titles need to be clean. Prices and availability need to be accurate. Variants, specifications and images remain fundamental.
AI-assisted shopping adds another demand. Product information increasingly has to support questions about suitability and comparison before a buyer reaches the merchant’s site.
A complete product record can still struggle here.
Take trail running shoes.
The catalog might contain:
Weight: 270g
Heel drop: 6mm
Lug depth: 4mm
Waterproof membrane: No
Now consider someone looking for a shoe that suits narrow feet, holds up over 15–20km runs on wet rocky trails and does not aggravate a tendency towards heel blisters.
The specifications help, but they do not settle much of that decision.
The more useful information concerns fit, performance in the stated conditions, longer-distance comfort and any recurring heel-rub problems. Some of that may come from specifications. Some may come from customer evidence or category expertise.
At Canonica, this is the layer we describe as product intelligence.
It includes information about suitability, buyer fit, use cases, performance conditions, limitations, evidence and review signals. The purpose is to make important product meaning more explicit instead of leaving each system or buyer to piece it together independently.
A clean catalog can still be weak for AI shopping
Catalog completeness can give teams false confidence.
A product may have every required field populated and still tell an AI system very little about when it is a good fit for a particular buyer.
Consider two trail shoes where both records say:
Grip: Good
Fit: Comfortable
Best for: Trail running
Durability: High
Nothing is obviously missing.
The weakness becomes clearer when the products need to be distinguished for a specific buyer. Perhaps one shoe tends to fit narrow feet more securely and has strong evidence around wet-rock traction, while customer feedback also shows heel rubbing on longer runs. The other may have stronger long-distance comfort signals and a more forgiving fit, with weaker evidence around slippery rock.
Those differences are much more useful in a buying decision than another broad claim about “good grip”. Simply adding more attributes does not solve this either.
A field can exist, be perfectly structured and still contain vague information.
best_for = everyday use
terrain = outdoor
fit = comfortable
performance = high
An AI system can parse those values without difficulty. They still settle very little.
The harder work is deciding which attributes matter for the category, how specific the values should be and what evidence supports them.
Evidence changes the quality of the recommendation
Reviews matter because they show what happens when product claims meet customer experience.
A jacket can be described as warm and suitable for cold weather. Reviews may show that owners find it comfortable for mild winter walks but unreliable in heavy rain.
A coffee machine might be sold as easy to use. Customer feedback may show that first-time owners struggle with cleaning and milk texture even though setup is straightforward.
That information can change whether a product is a good recommendation for a particular buyer.
Review volume alone is not enough.
Ten reviewers saying a shoe has “great traction” do not necessarily support a claim about wet-rock performance if none of them mention wet conditions. A smaller number of reviews describing actual use on slippery rock may be more relevant to that question.
The same applies to fit.
A large number of generic comfort comments does not automatically tell you whether a shoe is suitable for narrow feet or likely to cause heel rubbing after 15 kilometres.
Turning customer evidence into product intelligence requires judgement about what the evidence actually supports.
It also requires restraint. If the evidence is weak or conflicting, the product record should be able to represent that rather than force every attribute into a confident positive claim.
There is no single AI shopper
We should not assume that AI shopping systems will all evaluate products in the same way. Research is still early, and controlled studies should not be confused with live commercial systems. The differences between models are still worth noting.
A 2025 study using a mock ecommerce marketplace found that vision-language agents reacted differently across models to factors including price, ratings, reviews and platform signals. Position and ranking did not affect every model in the same way.
AgenticShop, a 2026 benchmark focused on personalised open-web product curation, found that current agentic systems still struggle with realistic product selection and personalisation.
Neither study tells us exactly how ChatGPT, Gemini or another live shopping system will behave for a particular product query. They do give ecommerce teams a reason to be careful about assuming there is one stable AI decision model to optimise around.
Clear, specific and well-supported product information is a more durable place to start.
These product gaps existed before AI
One reason I think product intelligence matters beyond the current interest in AI shopping is that the same weaknesses affect ordinary ecommerce.
If a product record does not clearly explain who the product suits, how it differs from nearby alternatives or where it performs poorly, people have to work harder too.
Product pages become generic. Buying guides repeat broad claims. Merchandising teams rely on knowledge that exists in people’s heads rather than in usable product information. Support teams answer suitability questions that could have been addressed earlier.
Reviews often contain useful detail that never makes its way back into the catalog.
AI makes these gaps easier to see because it is being asked to assess products quickly across several requirements. The underlying information problem was already there.
Better product intelligence also gives the rest of the business better material to work with. Merchandising can explain the role of similar products more clearly, while product and CX teams can use recurring suitability questions and customer evidence rather than relying on broad claims.
Agentic commerce is developing unevenly
There is good reason to avoid assuming that full autonomous purchasing is about to replace ordinary ecommerce.
EMARKETER reported in July 2026 that adoption of agentic commerce was progressing more slowly than earlier forecasts suggested. Shoppers appeared more comfortable using AI for research and product finding than allowing it to complete purchases. Its latest forecast put AI platform-driven sales at $19.74 billion for the year, around 4% below its December 2025 projection.
For ecommerce brands, the practical priority is the part of buyer behaviour that is already changing: product research, shortlisting and evaluation inside AI tools.
People may arrive at a merchant site after completing more of the consideration process elsewhere.
A shopper asking AI for a suitable trail shoe, baby monitor, vacuum cleaner or espresso machine is giving the system a set of conditions. The system then has to work from whatever product information it can access to judge which options appear to fit.
If the available information is generic, inconsistent or poorly supported, the system has more interpretation to do.
What agentic commerce readiness looks like today
A lot of agentic commerce readiness work focuses on feeds, structured data, platform access, checkout infrastructure, measurement and AI visibility. Those checks should include a more basic product test: whether the available information can answer the questions buyers actually ask before choosing.
Take a priority category and look at those questions.
For a trail shoe, they may concern terrain, fit, running distance or blister risk.
For a baby monitor, they could involve signal reliability across different house layouts, night performance, app dependence or battery life.
Then look at where the answers currently live.
Some will be in the catalog already. Others may be buried in reviews, supplier documents or buying guides. Some will exist only as knowledge held by merchandising, product or support teams.
There will also be questions where the business simply does not have enough evidence to give a confident answer.
Once those signals are identified and governed, the useful ones can be represented more clearly in the product record and passed into the storefront, feeds, marketplaces and other places where the product appears.
AI systems may still use other sources, and the final recommendation remains outside the merchant’s control. The brand can control the quality of the product information it makes available.
The practical work starts with product understanding
Agentic commerce may eventually include much more autonomous purchasing. Checkout infrastructure will continue to develop, and platforms are already building for it.
The immediate work for ecommerce teams is fairly practical.
Buyers are already asking AI to help research products, narrow their options and decide what fits. Product information therefore needs to explain more than basic identity and specifications. It needs to carry useful information about suitability, differences between nearby products, important limitations, performance conditions and the evidence behind those claims.
Those are product-understanding problems. They exist regardless of how quickly autonomous checkout develops.
FAQ
What is agentic commerce?
Agentic commerce refers to the use of AI systems to help shoppers research, compare, select and sometimes purchase products. For ecommerce teams, the impact already begins during product research and evaluation, before an AI system completes a transaction.
Why does product understanding matter for agentic commerce?
To make useful product recommendations, AI systems need enough information to assess products against a buyer’s requirements. Titles, prices and specifications provide part of that information, while more specific buying questions can require context about suitability, performance, limitations and evidence.
What is product intelligence?
Product intelligence is structured product information that helps explain who a product suits, what it is useful for, how it performs in relevant conditions, where its limits are and what evidence supports those conclusions.
Does better product intelligence guarantee an AI recommendation?
No. AI systems can use several sources and control their own ranking and recommendation logic. Ecommerce brands can improve the quality of the product information available to those systems, but they cannot control the final answer.
Sources
OpenAI — Introducing shopping research in ChatGPT
OpenAI — Powering Product Discovery in ChatGPT
Google Developers Blog — Under the Hood: Universal Commerce Protocol (UCP)
Shopify Enterprise — Agentic Commerce: An Executive Guide to What's Happening and What to Do About It
Reuters — Retailers tap AI shopping traffic but fight to keep customer data, 2026
Columbia Business School — What Is Your AI Agent Buying? Evaluation, Implications and Emerging Questions for Agentic E-Commerce
AgenticShop — Benchmarking Agentic Product Curation for Personalized Web Shopping
EMARKETER — Retailers hold AI commerce advantage over ChatGPT, July 23, 2026
Research papers
Comments