What Do Shoppers Ask AI? How to Build a Useful Prompt Market for Ecommerce

Shoppers do not always ask AI for a product by name. They describe needs, constraints, trade-offs, budgets and comparisons, then expect the system to work out which products fit.

That behavior is becoming easier to see. Google says the average AI Mode search is triple the length of a traditional Search query, while OpenAI's shopping research asks follow-up questions about preferences such as brand, size, performance, comfort, style and price. Adobe's July 2026 consumer study found that 42% of surveyed US consumers had used AI assistants for online shopping, with 53% of that usage happening during inspiration and discovery.

For ecommerce teams, the implication is simple: a useful shopper-question map is not a list of thousands of generated prompts. It is a structured view of the shopping situations that matter, connected to evidence and to the products that can genuinely satisfy them.

At Gravton, the useful unit is usually the intent cluster, not the individual wording.

What should ecommerce teams know first?

A strong shopper-question map should tell you what the shopper appears to be trying to accomplish, what evidence supports the question, which market and journey stage it belongs to, and which products are genuinely eligible to compete.

One distinction matters throughout:

Intent is not demand.

Intent describes the goal inferred from a shopper's question. Demand describes how much of that need exists in the market. A question set can help map the first; it cannot establish the second simply by containing more prompts.

The objective is therefore not maximum prompt count. It is sufficient coverage of the distinct shopping situations that matter to the category.

What kinds of questions do shoppers ask AI?

Conversational shopping questions often combine several filters in one sentence. A shopper can describe the job, add a price ceiling, specify a product constraint and ask for a comparison at the same time.

These patterns overlap. "Which sandals work for swollen feet on a long flight and cost under $150?" contains a use case, physical constraint and budget in one question. Reducing that to a keyword such as "comfort sandals" throws away useful decision context.

That is one of the main differences between keyword monitoring and a useful AI-shopping question set. The natural-language question often contains the criteria that determine which products should be considered suitable.

Why can a long prompt list still be weak?

Generating more prompts does not necessarily create a better representation of the market.

A running-shoe brand can generate separate questions for flat feet, overpronation, stability shoes, walking comfort and brand comparisons, yet several of those may represent variations of the same underlying need. If every wording becomes an independent measurement unit, the market gets overweighted toward whichever ideas have the most phrasings.

A useful audit looks for problems such as these:

The same principle applies to demand. A set containing 5,000 prompts tells you that you have 5,000 prompts to test. It does not prove those prompts represent more market demand than a disciplined set of 500.

Demand sizing requires separate evidence such as first-party search behavior, marketplace signals, category research or other defensible inputs. Even then, conversational demand may need to be modeled rather than stated as an exact count.

How should shopper intent be classified and sourced?

Intent classification assigns a shopper question to an inferred goal: what the person appears to be trying to discover, evaluate, compare or purchase.

The word inferred matters. A label such as "comparison intent" is an analytical interpretation, not explicit shopper data. One primary intent can help organize the set, but the other constraints in the question should not disappear.

For ecommerce, it is useful to retain five dimensions: shopper intent, product or category, journey stage, market and shopper context. The point is not to create an elaborate taxonomy. It is to preserve the information that changes what a suitable recommendation looks like.

Provenance should be retained alongside that classification.

Observed questions come from direct evidence such as Search Console, onsite search, support conversations, reviews, retailer Q&A or sales conversations.

Derived questions combine known shopper needs with real product attributes, competitors, price bands, use cases or constraints. They were not directly observed, but the underlying evidence makes them commercially plausible.

Synthetic questions can help explore coverage gaps, but generation alone is not evidence that shoppers actually ask them.

Search Console is useful input because it reports queries that led users to a site, but Google notes that some anonymized and lower-volume queries are omitted from the report. That makes first-party search data useful evidence, not a complete category map.

What does a real shopper-question set reveal?

An anonymized Gravton category study in September 2026 shows why the structure of the question set matters.

In that dataset, 59% of the shopper questions named no brand at all. Fewer than one in ten top-of-funnel questions mentioned a brand. Earlier-stage shoppers were more likely to describe foot pain, walking requirements, support needs, price constraints or other problems and leave the AI system to decide which products belonged in the answer.

The comparison behavior was also specific. Every question that named a competitor included a reason for the comparison, such as width, arch support, durability, price or break-in time. The useful signal was not simply that two brands were being compared. It was the dimension on which the shopper expected the system to judge them.

The evidence mix also mattered. The set combined first-party inputs such as Search Console and catalog data with broader public category research. Equivalent phrasings were grouped at the intent level rather than treated as independent units.

Method note: this was an anonymized US ecommerce category study conducted in September 2026. The findings describe the structure of the monitored question space, not the frequency with which the wider market asks each question.

How should intent clusters connect to the catalog?

The shopper-question map becomes ecommerce intelligence when it connects to products the brand can genuinely sell into the need.

If a question requires a product under $150, products above that ceiling should not be counted as missed opportunities. If wide sizing is essential, a SKU that does not offer it should not be considered eligible simply because it belongs to the same category.

This keeps the denominator honest. It prevents the business from treating every absence as an AI-visibility failure when the assortment itself does not fit the shopper's requirement.

Google's Merchant Center AI reporting follows a similar move toward richer shopping context. Google separates conversational shopping activity into discovery, evaluation and ready-to-buy stages and surfaces terms, attributes and shopping intents rather than treating every shopping interaction as equivalent.

The same map can also mean different things for different business models. A retailer with thousands of SKUs may care most about category coverage and assortment gaps. A D2C brand with 80 products may care more about whether its hero SKUs enter consideration for the right needs, which competitors replace them and whether the eventual purchase route reaches the brand's own store.

A useful final record should let the team trace one need through to the catalog and current outcome:

The table itself is not the insight. The value comes from being able to explain why SKU A loses, why SKU B wins or why SKU C rarely appears despite being eligible.

Once the question universe and eligible-product set are stable, they become the measurement base for later metrics such as recommendation share and SKU coverage. Which AI-Shopping KPIs Should an Ecommerce Team Track? explains how those measures fit into the executive scorecard.

How should ecommerce teams build the shopper-question map?

A defensible public method is straightforward.

Start by defining the category and market you want to measure. Gather evidence from first-party behavior, catalog data, reviews, retailer information, competitor comparisons and broader category research. Keep provenance attached to each question so observed, derived and synthetic inputs are not treated as equivalent evidence.

Next, infer the shopper's underlying goals and retain the constraints that affect product suitability. Group equivalent needs into intent clusters so multiple phrasings do not distort the measurement set. Then map each cluster to the products that genuinely satisfy its requirements.

Finally, review the set for obvious bias. Check whether it is overly branded, whether important comparison dimensions are missing, whether early-stage needs are represented, whether the same idea appears too many times and whether every important cluster connects to eligible products.

The output should be a market definition the ecommerce team can defend, not simply a large prompt library.

What else should ecommerce teams know?

Can we see the actual prompts people type into ChatGPT?

Not through normal brand reporting. A constructed shopper-question set should therefore not be presented as a log of actual ChatGPT user prompts unless the business has direct first-party evidence supporting that claim. Observed, derived and synthetic questions should remain clearly distinguished.

Should every shopper question become a separate intent?

No. Equivalent phrasings should usually be grouped around the same underlying need. The original wording can still be retained for testing, while the intent cluster becomes the cleaner unit for analysis and reporting.

Does a larger prompt set mean there is more demand?

No. Prompt-set size describes the test design. Market demand requires separate evidence. Generating more variations of the same shopper need can make the test larger without making the underlying commercial opportunity any larger.

How does the shopper-question map connect to AI-shopping KPIs?

It creates the stable set of shopping situations and eligible products against which later performance can be interpreted. Without that stable measurement universe, changes in recommendation share or SKU coverage can reflect changes in the test itself rather than real changes in performance.

What should you read next?

Query Fan-Out Explained explains how a complex shopper question can break into several information requirements.

Intent Intelligence: Understanding Why Users Ask What They Ask goes deeper into the interpretation of intent.

How To Track ChatGPT Traffic covers the observable traffic layer once an AI-assisted journey reaches the website.

Which AI-Shopping KPIs Should an Ecommerce Team Track? shows how the measurement universe feeds the executive scorecard.

Which sources support this guide?

1. Google, “How AI Mode is changing the way people search in the U.S.”, May 19, 2026

2. OpenAI, “Using shopping research in ChatGPT”

3. Adobe, “AI Traffic Trends Report”, August 2026

4. Google Search Console, query reporting documentation

5. Google Merchant Center, AI performance insights

Want to see what your shoppers are asking AI?

Gravton's AI-shopping audit maps the questions shaping a category, the products appearing for them, the competitors intercepting those decisions and the catalog or content gaps behind the result.

See how AI shops your category