Candidate-set inclusion is not the same as ranking.

When a shopper asks an AI system for product recommendations, the visible order is only the end of a longer decision path.

Four questions make the result useful

1. 01RetrievedCould the system find usable product information?
2. 02IncludedWas the product treated as a plausible option?
3. 03ComparedWere the right facts used?
4. 04OrderedWhere did it appear in the final answer?

Those questions are related, but they do not mean the same thing.

01 · RETRIEVABILITY

Could the system find usable product information?

Retrievability is the first gate. The system needs enough public information to connect a product to the shopper's need.

A product can fail here even when its page is technically accessible. The facts may be scattered, the category may be unclear, or outside sources may describe it differently.

If the product never appears, the test has not yet shown why. It has only established an omission under the recorded conditions.

02 · CANDIDATE-SET INCLUSION

Was the product treated as a plausible option?

The candidate set is the group of products the system appears to consider relevant enough to present or discuss.

Entering that set is different from being placed first. A product can be a credible match and still lose to another option on size, price, availability, buyer fit or a more specific use-case detail.

That is not automatically a visibility failure.

03 · COMPARISON OUTCOME

Were the right facts used?

Once the product is included, the next question is whether the comparison was fair.

Look for three things:

• Were the important product facts accurate?
• Were the buyer-fit advantages connected to the question?
• Was the reason for choosing another product defensible?

This is often more commercially useful than the visible order. An inaccurate comparison can reveal a source-clarity problem. An accurate comparison may show that the product simply was not the best fit for that particular shopper.

04 · FINAL ORDERING

What did the answer put first?

Final ordering is the easiest result to notice and the easiest to overinterpret.

The order can change with the wording of the question, available sources, inventory, model behavior, session conditions and time. One captured answer is evidence of what happened in that test—not a stable market ranking.

The practical question is not “Did we rank number one?” It is “Where did the product leave the decision, and was that outcome supported by accurate information?”

PRACTICAL CHECK

A five-minute merchant check

To examine one product without forcing it into the answer:

1. 1

   Start a fresh AI conversation and record the product, date, visible model or mode, and market you are testing.

2. 2

   Ask a natural buyer question without the brand name, product URL or copied proprietary specifications.

3. 3

   Save the complete question, answer and cited sources before interpreting the result.

4. 4

   Mark retrieval, candidate-set inclusion, comparison accuracy and final ordering separately.

5. 5

   Repeat only when you have a clear reason, such as a different buyer need, market or implemented source change.

This check does not explain every cause, and it does not create a stable ranking. It does help distinguish a real diagnostic question from a disappointing but reasonable comparison.

NEXT DECISION

When deeper diagnosis is useful

A deeper product-level diagnosis is warranted when the product is:

• absent from a relevant candidate set;
• associated with the wrong category or buyer need;
• included but rejected using inaccurate or incomplete facts;
• contradicted by public sources; or
• recommended in principle but unsupported by the landing product page.

The next step should be the smallest controllable change that addresses the observed failure, followed by a controlled retest. The goal is not to promise a ranking. It is to produce clearer evidence about why one real product did or did not survive the shopping decision.