RESEARCH PROPOSALRESEARCH PLAN

Why does AI recommend our brand?

Do not stop at whether the name appears. Read the reason for the recommendation.

Document version
Draft 01
Updated
Design status
Design proposal / pre-experiment
Schedule
To be determined

What we want to learn

Research question and hypothesis

When the same brand is recommended, how accurately do the reasons reflect the user situation and the strength the brand wants to communicate?

Hypothesis

Providing product facts together with use situations may produce more question-specific reasons instead of generic reasons such as price.

This is an expected direction, not an established finding.

Why ask this question?

A brand mention alone does not show that the intended meaning was conveyed. Recording reasons separately distinguishes incorrect description from meaning transfer.

What we will compare

Keep the campaign
change the explanation

Compares whether recommendation reasons fit the question and intended brand meaning, not merely whether the brand is mentioned.

Held constantSame product facts · same question · same AI condition
A

Basic product description

Provide only general facts such as size, material, and price.

B

Situation-linked description

Connect the same facts to the customer situations they fit.

Measure each condition separately
Check in the answerDoes the reason fit the question?

Record how the situation, facts, and poor-fit conditions appear in the answer.

Check against the brandIs the intended meaning conveyed?

Separately score whether reasons overlap with the strength the brand intended to communicate.

Human response and AI output are measured separately. One does not stand in for the other.

Controls

Keep product facts, question, service, model, search use, and repeats constant, changing only how the explanation connects facts to situations.

Example: travel luggage

Ask for the same fictional suitcase, then evaluate separately an answer that only says it is inexpensive and one that explains why it fits a short weekend trip.

How we will proceed

Proposed sequence

  1. Lock evaluation criteria

    Define intended brand strengths, product facts, and fit conditions for each question before viewing answers.

  2. Repeat under the same conditions

    Provide both descriptions under the same service and question conditions, mixing order across repeats.

  3. Review reasons blind

    Record brand mention, factual accuracy, question fit, and brand-meaning match separately.

This sequence is a draft. Sample size, repetitions, and decision rules will be fixed before execution.

What we will record

Measures and limits

MeasureMethodLimit
Factual accuracyCompare the answer with the product fact listAn accurate answer may still give a reason unrelated to the question
Question fitCheck whether recommendation reasons directly connect to the use situationDo not impose evaluator expectations after seeing the answer
Brand-meaning matchCompare answer reasons with pre-defined brand strengthsSimple repetition of ad copy is not success

How we will judge results

Rules fixed in advance

Question fit and meaning match both rise

Check whether the direction repeats for other products and questions

Only brand mentions rise

Keep mention and meaning transfer as separate outcomes

Factual errors rise or results are mixed

Do not confirm the format effect; investigate error causes first

No description is judged superior until answer count, evaluators, and repeat criteria are set.

Decisions before launch

Items not yet fixed

Target
Brand and product not set
Questions
Situation-specific question set not set
Measurement
Service, repeats, and evaluators not set
Schedule
Separate candidate from monthly experiments A–C

Scope of this plan

  • Recommendation reasons in AI answers do not substitute for actual purchase impact.
  • Directly supplied materials and autonomous web discovery must be separated.
  • This is a research proposal, not evidence of execution or effect.

Background and sources

Reviewed guidance
and research rationale

OpenAI guide to shopping results in ChatGPT

Official explanation that shopping results use query intent, context, product information, and third-party content. It is not evidence that this design works.

Checked 2026-09-26

Read the full backgroundFull explanation and hypothetical example

The same product can be recommended for different reasons

Our travel bag appeared in an AI recommendation list. That is welcome, but the rationale deserves attention because the reason we want to communicate may differ from the one AI gives.

Imagine a campaign showing friends leaving for an unplanned weekend. Its goal is to make people feel, “I want to go too,” not to make them memorize the bag’s weight.

An AI could recommend the bag because it is inexpensive or because it holds what a short trip requires. It selected the same bag, but those answers tell different stories.

If the customer asked about price, affordability is a natural response. But for someone looking for weekend luggage, we should also examine why the bag is recommended. The presence of a brand name alone does not show that the campaign meaning carried through.

Look beyond price and performance to taste and situation

OpenAI says ChatGPT shopping results consider the intent and context of a question, using sources such as product information, prices, and third-party content, while acknowledging that intent can be misread. Official explanation ↗

Its shopping-research guidance also says the system may ask whether a user values performance, comfort, style, or price, then provide reasons, strengths, and weaknesses suited to those constraints. Official guidance ↗

This supports looking at taste and use situation as well as price and performance. It does not establish how well AI understands campaign mood or shared cultural taste, nor whether changing an ad changes recommendations.

AI does not need to repeat the campaign line verbatim

An AI need not repeat “an unplanned trip” word for word. If it accurately explains why the bag suits someone who takes short weekend trips, the intended meaning may have carried through in different language.

Conversely, repeating the campaign line while adding a nonexistent feature is not a good answer. Examine both factual accuracy and whether the intended appeal is explained.

What to inspectQuestion to ask
Who is it recommended to?Does the person and use situation match the intended customer?
Why is it recommended?Does it explain a reason this person would choose the brand?
What supports the claim?Can it be verified in actual product information or use experience?
When is it unsuitable?Does it disclose drawbacks or situations where it should not be used?

These are review questions, not validated scores that guarantee more recommendations.

Diagnose the failure before revising the campaign

The response depends on where the problem occurs.

  • Customers do not understand the campaign. Examine its message and expression first. More product copy for AI will not automatically make the ad easier to understand.
  • Customers understand the campaign, but AI describes the product incorrectly. Check the sources used, stale information, product confusion, and unsupported additions.
  • The description is accurate, but there is no reason to choose the brand. First ask whether the customer question calls for that reason. If the question concerns weekend travel, examine whether use stories or the product page explain why the bag suits short trips.

One ad does not need to contain everything. A campaign can create the desire to leave, a trip review can show real use, and a product page can explain capacity and cautions. Whether this division helps must be tested.

Choose one campaign and read the rationale

Write one sentence: “Who should choose our brand, in what situation, and for what reason?”

One active campaign is enough. Place the intended story, the customer situation that needs it, and verifiable product or review evidence side by side. Find customer language in inquiries, interviews, or search data. Then read AI answers to see whether the recommendation rationale fits.

Do not insert the campaign line into the question merely to obtain the desired answer. That may only test whether the system repeats a supplied phrase.

The proposal is simple: do not look only for the brand name in an AI answer. Read who it is recommended to and why.

Sources reviewed and remaining unknowns

  • We reviewed the official guidance on September 26, 2026. It describes features; it does not show adoption or brand impact.
  • This article does not compare actual brand answers with customer response.
  • It contains no same-condition comparison between the previous two weeks and the most recent two weeks. We cannot say whether a tendency persisted, appeared, weakened, or reversed.
  • A formal biweekly report needs comparable period data and counterexamples before describing direction. This article is an earlier-stage research proposal.