From Multi-Model Research to a Product Selection Proposal

For a multi-category product selection project, LouLou and I prepared research records, versions with my selection decisions, and an English-language product proposal. The brief involved different budget conditions, delivery timing, and manufacturing feasibility.

AI produced many candidates. My job was to decide which deserved further research, where the evidence was insufficient, and which conditions would need confirmation through supplier inquiries or sampling. The outcome documented here stops at research and proposal materials. It does not extend to client adoption or an order.

I assigned different AI tools different directions: working backward from cost to possible products, finding real products and their sources, exploring how products work, and borrowing ideas from other use cases. LouLou is the shared name I give these AI collaborators.

The results were then brought into a candidate table, with duplicates addressed and sources and open questions retained. I narrowed the selection through review and feedback. The remaining directions became product descriptions, specifications, images, and estimates for a client to review.

Three judgment calls stood out along the way.

The initial research treated products recommended by several models as a stronger signal. Looking back, that signal needs to be kept separate from fact-checking. Different models may have found the same page or repeated the same assumption.

Repeated recommendations can help prioritize what to investigate. They cannot prove that a product fits the project, or that a particular factory can manufacture it to the required specification. Each candidate still needs questions answered: Where did the information come from? Does it establish that a product exists, support a specific feature, or merely repeat a supplier's claim about its capabilities? Those conclusions are not interchangeable.

In one round, AI applied criteria developed for a lower-budget group to another group of candidates. I rejected many of those recommendations and separated what each group needed in product form and perceived value.

The same effect can lead to a different decision when it appears in a different type of product. Recording only that I like a particular effect leaves plenty of room for the next round to go wrong.

Useful feedback therefore needs to preserve why I kept or rejected an option, and the conditions under which that reason applies. Shortening the list also needs care: the original options should remain available so worthwhile directions do not disappear just to make the table shorter. That kind of over-pruning occurred in the project record.

Reviewing the references revealed that some links led to search results or storefronts. Other pages opened successfully but did not match the recommended product, specification, or manufacturing process.

These differences affect which suppliers to approach and how to write an inquiry. A finished product used as a reference does not necessarily establish that the required production process is available. I separated “a link was found” from “this link supports this claim,” leaving gaps open for confirmation.

What reviewing a candidate looks like

The table below is a wholly fictional reconstruction. Its products, records, and situations do not correspond to the actual project. It illustrates the review method only.

LouLou helped with search, extraction, categorization, consolidation, and drafting. I was responsible for defining the business conditions, reviewing evidence, making selection decisions, and pointing out where a rule had been misapplied.

The work produced research tables, selection records, reference-review materials, and an English-language product proposal. Prices in the proposal remained estimates. The proposal explicitly required final pricing to be confirmed with suppliers after selection approval. Producing those documents did not constitute factory verification, product testing, or purchasing approval.

What I took from this project was a clearer separation of decisions that are easily confused during product selection, together with checks I could use again. The process still needs human business judgment. There is also insufficient evidence to claim a particular amount of time saved or orders generated.

If you have a product shortlist prepared by AI, start with one claim and check exactly what its link supports.

Try the free Research Link Review Card. Copy the prompt, use the synthetic example or authorized anonymized material, and review the result against the checklist. That first step can help reveal whether your team's immediate gap is in sources, specifications, or review rules.

The project background has been generalized, and identifying client, supplier, and transaction details have been removed. All demonstration products and table entries are fictional, not actual deliverables. The outcome documented here stops at product research and proposal materials. Client adoption, final supplier pricing, sample testing, and order outcomes have not been confirmed in this account.