A founder asks which market has the strongest demand for an adjustable home-storage product. The analyst uploads support and enquiry messages, asks AI to identify themes, and receives a confident answer: fit questions dominate in the United States, so the next inventory allocation should favour that market.
The answer may describe the file while failing to answer the business question. A customer seeking help after purchase and a person considering a first order are not interchangeable observations of demand.
A convincing ranking built from unlike inputs
Fictional teaching example: The messages, people, counts and candidate AI conclusion below are invented. No real customer data, market research or Aright campaign is described.
A retailer has 120 retained messages from one period. Its authorized internal matching records link 80 US post-purchase support messages to 50 distinct customers, and 40 Singapore pre-purchase enquiry messages to 20 distinct people. The two groups do not overlap. “Fit” is a reviewer-defined theme, not a purchase or return event.
| Retained source | Raw messages | Messages mentioning fit | Distinct people in retained source | People with at least one fit message |
|---|---|---|---|---|
| US post-purchase support | 80 | 48 | 50 | 28 |
| Singapore pre-purchase enquiries | 40 | 12 | 20 | 9 |
| Total retained file | 120 | 60 | 70 | 37 |
The grouping of people is an explicit assumption of the example, not permission to identify or join real customers across systems. Actual analysis needs an appropriate purpose, lawful access and a method that preserves uncertainty about unmatched records.
A deliberately bad summary says:
Fit demand is twice as strong in the US: 60% of US messages concern fit, against 30% in Singapore. Increase US stock by 50%.
The percentages are arithmetically correct: 48/80 = 60% and 12/40 = 30%. The conclusion changes what they measure. These are proportions of selected messages using a theme, not proportions of people who want the product, and certainly not an estimate of future orders.
Fixing duplicates still does not fix the sample
After linking repeat messages, the within-file person figures are 28/50 = 56% and 9/20 = 45%. Neither is a market demand estimate. The US group contains people who already bought and then contacted support. The Singapore group contains people who asked a question before purchase. People who understood the page, never saw it, bought elsewhere or abandoned it silently are absent.
The business also lacks the number and composition of eligible product viewers in each market. It does not know whether support access, advertising exposure, language options or response practices made one group more likely to appear. Dividing by distinct people reduces one form of repetition; it does not make the two collection processes comparable.
Even a perfect theme classifier cannot recover those missing populations from the text. More sophisticated clustering can improve the description of the retained file while leaving the inference unchanged.
This matters when adding platform data. Google's Search Console query documentation says the API returns top rows rather than guaranteeing all rows. Those records describe a property's observed search traffic under the chosen query, not total market demand. An absent row should not quietly become proof that nobody in a market has a question.
Preserve the useful finding without inventing a ranking
A defensible analyst note would read:
Observed: In the retained US post-purchase source, 28 of 50 linked customers raised at least one fit question. In the retained Singapore pre-purchase source, 9 of 20 linked people did so. These are different contact populations.
Useful question: Which measurement or compatibility distinction is unclear, on which product/page version, and at what point in the buying journey?
Not established: Relative market demand, the incidence among all buyers, a cause of returns, or the value of changing inventory.
Action: Examine the fit-question evidence separately in each source. Keep the inventory proposal open until comparable commercial evidence supports it.
That note is more useful than deleting the analysis because the sample is imperfect. It keeps a real kind of observation while giving it an appropriate job.
For the US branch, inspect the order variant, product instructions and page seen by the customer. A fit question after delivery might reflect unclear measurements, a wrong item, a use case outside the stated conditions, or simply a request for confirmation. Each explanation points to a different action. For Singapore, examine what information the enquirer lacked before ordering. Do not force both groups into one “market preference” story.
Use AI where the uncertainty is inspectable
AI can propose themes and retrieve candidate passages for human review. Give each proposed theme an operational meaning. “Fit” might otherwise combine dimension questions, material preference, incorrect fulfilment and installation problems. Preserve a passage that shows why the label was proposed, and let reviewers disagree or leave it unresolved.
A small double-reviewed sample can expose label confusion; it does not establish a universal accuracy rate. NIST's Generative AI Profile cautions against extrapolating from narrow assessments. Here, the classifier's quality and the dataset's ability to answer a market question are two separate evaluations.
For a future market comparison, first name the outcome: suitable enquiries, completed orders, retained contribution or something else. Then design a reasonably comparable collection process around the same product, offer, period and observable exposure, recording differences that cannot be aligned. Even that may describe only the business's reached audience rather than the whole market. An appropriately designed new test can change the investment decision; rewording the old prompt cannot supply missing exposure or sales evidence.
The immediate choice is therefore to investigate the fit obstacle and improve the evidence for a later allocation—not to increase US inventory from a theme count. A useful AI analysis sometimes ends with a narrower recommendation because it has made the original question more precise.