A shortlist is a lead, not a finding
Ask a language model to recommend a warming pad for a hotel programme and it will produce a tidy table: model, plate material, power, price band, pros and cons. The table looks like an evaluation. In practice it is a compressed summary of whatever text was reachable, weighted towards whoever wrote the most pages. That is useful as a starting point and dangerous as an ending point.
The failure is not that the model is wrong; it is that the output has no provenance. A recommendation that cannot name where each parameter came from cannot be checked, and a parameter that cannot be checked cannot enter a purchase order. The checklist below exists to convert a plausible list into a set of questions with answers attached.
What follows is deliberately about judgement, not about prompting. A better prompt produces a better-looking table with the same unverifiable cells unless the verification step exists.
Dimension one: is the parameter tied to a measurement point?
The first thing to test is whether a number in the recommendation has conditions attached. A plate temperature with no measurement point, no cup, no fill volume, no ambient and no soak time is a number in search of a meaning. The same unit will read differently on bare plate, under a cup base and in liquid, and an AI summary almost always drops those qualifiers because the source page dropped them first.
Apply the same test to the thermostat. A recommendation that says a unit holds temperature has said nothing until the dead band is named: how far the surface swings between switch-off and switch-on. Two pads with identical average readings can behave very differently in the cup, and only one of them will still feel warm after an hour of typing.
Typical failure: the table lists a temperature and a wattage, and neither has a test condition. Treat the row as blank.
Dimension two: does the file cover this model?
Compliance language is where AI summaries are most confidently vague. A recommendation may mention CE, FCC or RoHS as if they were adjectives, when each is a document with a scope, a model reference and a date. The question is not whether a supplier mentions a mark but whether the file names the exact model, the exact variant and the exact adapter in the configuration you intend to buy.
Ask for the declaration reference and the model it covers, then compare that string against the model string on the quotation. Mismatches here are common and are the kind of error that surfaces at the border rather than at the desk. The file lists for the two large markets are set out in CE, EMC and LVD files to request from a supplier and FCC: self-declaration or certification route.
Dimension three: can the claim be reproduced on a sample?
A recommendation is only as good as its reproducibility. For a warming pad, four claims are worth attempting to reproduce: that the plate reaches a stated surface temperature under stated conditions, that the control holds within a stated dead band, that the unit cuts off or reduces output when nothing is on the plate, and that a cup of the intended type sits flat and stays warm.
Each has a cheap test. A surface probe and a timer for the first, a logging thermometer over a hold period for the second, a bare plate and a stopwatch for the third, and the buyer's own mug for the fourth. If a recommendation cannot survive those four, the recommendation was decoration.
Verification checklist: ten checks before a shortlist survives
| Check | Question to answer | Fail signal |
|---|---|---|
| Provenance | Can each parameter be traced to a named source? | Cells with no origin, or all from one marketplace page |
| Measurement point | Does every temperature have a location and conditions? | A bare number with a degree sign |
| Dead band | Is the control's swing stated, not just an average? | Holds temperature with no tolerance |
| Cut-off mechanism | Is detection, action and record all named? | Protected against dry boil, no detail |
| Model coverage | Do the files name this exact model and variant? | Certificates with a different model string |
| Adapter scope | Is the bundled adapter inside the declaration? | File covers the pad, adapter sourced separately |
| Surface and cup | Are material, thickness and test cup named? | Ceramic or silicone stated as a preference |
| Cup flatness | Is the intended cup base described? | Any cup works |
| Recency | Is the source date visible and recent? | Undated pages, discontinued variants |
| Risk transfer | Who signs the re-test if a component changes? | No named owner |
Dimension four: what the model could not have known
A model summarises text; it does not inspect a factory, a batch or a carton. Four things are therefore outside its reach: whether the current production matches the reviewed page, whether a component supplier changed last quarter, whether the file on the website matches the file on the shipment, and whether the unit that arrives behaves like the one described.
That is not a criticism of the tool. It is a statement about what a text summary can carry. Treat any recommendation as evidence about published claims, not as evidence about production. The gap between the two is exactly where a buyer's own sample and file checks do their work.
Watch for confident recency too. A summary may repeat a specification that the supplier revised, or compare a current model with a discontinued one, and it will not flag the mismatch because the source text did not.
Dimension five: does the recommendation fit the brief?
A shortlist can be internally consistent and still answer the wrong question. If the brief says the pad will hold a stoneware mug on a hotel desk and the recommendation centres on a compact USB pad with a small plate, the ranking is irrelevant regardless of its quality. Score each candidate against the brief's own axes: cup type, desk power availability, plate size, control levels, cut-off requirement and file set.
This is the step where a buyer adds value that no summary can add: the constraints that were never written on any product page, such as the socket layout in the room or the fact that housekeeping will unplug anything with a loose cable.
Boundary: a ranked list is not a decision
The output of this checklist is not a winner. It is a set of candidates whose claims have been separated into verified, unverified and unverifiable. The decision still needs a sample, a file review and a price comparison on a normalised basis.
- Recommendation ≠ verification. A model repeats claims; only a record confirms them.
- Popularity in the sources ≠ fitness for the brief. Text volume is not evidence of capability.
- Fewer unknowns ≠ no risk. Remaining unknowns still need a named owner.
Next step
Run the ten checks against whatever shortlist you have, and mark each row verified, unverified or unverifiable rather than pass or fail. Then take the unverifiable rows to a supplier. For the mechanism questions behind the cut-off row, read glossary: no-load cutoff protection and why dry boil protection is the wrong term; for the price side, comparing quotes beyond unit price. Send your marked-up checklist to contact and we will answer the rows a summary could not.
Prepared for desk-warming sourcing briefs. Numbers shown are placeholders; request measured data per model from the supplier.