Measuring a Spreadsheet Instead of Trusting It

This site can measure two things about a spreadsheet and refuses to estimate the rest. The two are the shape of a link string, read without sending any request, and the structure of a 195-entry pool carrying 2,806 images, read by counting fields. Link availability, parcel transit time and seller reliability are not measured here, not estimated here, and not implied by anything published here. That boundary is the design, not a gap in it, and this article exists to state it precisely enough that a reader can check it.

The sample is fixed and small enough to describe in one sentence. It contains 195 entries extracted on 2026-09-29 from the eight category views of one source marketplace, spread across 8 categories and 20 brands, carrying 2,806 images between them. Every entry has a recorded source link, a recorded price, a recorded image group, a category and a check week. Nothing in the sample was observed twice, and nothing was purchased, opened, or followed.

1. Fix the sample before counting. The pool, the date and the source views are stated first, so a figure can never be produced by widening the sample until it looks right.
2. Separate the two measurable objects. A link is a string, and the pool is a set of fields. They support different questions and they fail in different ways, so they are never combined into a single score.
3. Publish the thresholds before the count. The index threshold is an image group of five or more, the status derivation uses two fields and that threshold, and the review week is uniform across the pool.
4. State the exclusions in advance. Availability, transit and seller reliability are named as out of scope at the top, which prevents a later section from quietly answering them with a proxy.

Both passes ran against the pool of 2026-09-29, and the review week attached to every row is 2026-W40. There is no earlier pass and no later one, so every figure in the results section is a single observation of a single pool. The date matters more than the volume here, because a dated census can be reproduced and an undated one cannot.

The first pass reads link form. It detects the host and maps it to the platform the link belongs to, extracts the identifier parameter the platform uses and falls back to a trailing numeric segment when the parameter is absent, flags tracking parameters from the families that sharers add, and judges the overall form of the string. The whole pass is string analysis. It makes no network request, which is exactly why it can say something about a link without risking a false statement about a listing.

The second pass counts pool structure. It counts entries, categories, brands and distinct source platforms, orders the recorded prices to produce a minimum, a median and a maximum, counts image groups per entry and totals them, and applies the image threshold to partition the pool. Every one of those operations is a count or an ordering over fields that were captured on the extraction date, and each is reproducible from the same pool with the same rules.

One consequence of the first pass deserves stating plainly, because it is the reason no link statistics appear in the results below. The link reading is a function rather than a dataset: it takes a string and returns a classification, and it produces the same answer for the same input every time. It is exposed as an instrument at /instrument/link-health/, where a reader can paste a link and read the result. What it does not produce is a frozen count of how many links in the pool sit in each state, because no such tally was written down as a dated artifact, and inventing one now from a function would be a reconstruction rather than a measurement.

The results are three tables and one sentence. The tables report composition, price and image coverage for the pool of 2026-09-29. The sentence is that no link-availability, transit or seller result exists, because those questions were excluded before the work started rather than left unfinished.

The 37 unverified rows are the one result that could easily be misread, so it is worth being explicit: that count is produced by the threshold. The rule excludes any row whose image group holds fewer than five images, and 37 rows meet that condition. The unverified count is therefore not an independent quality finding that happens to equal 37; it is the threshold applied to one field, and a different threshold would produce a different number from the same pool without anything about the items changing.

The limits here are not incidental, and listing them takes less space than defending against them would. Each one below describes something the measurement cannot see, and each one has a specific remedy that would require work this site has not done.

1. Selection bias: the pool is one platform, eight category views and one date. It is what was visible in those views on that day, so any share computed from it describes this pool and not a market.
2. Threshold bias: the five-image rule manufactures the unverified count. Reporting 37 unverified rows without the rule attached would turn a field comparison into an apparent quality judgement.
3. Survivorship: an entry delisted before the extraction date never enters the pool and is not recorded anywhere. A pool built from what is currently visible systematically misses what has already gone.
4. No outcome data: nothing here observes arrival, condition, cost at the door, or dispute results. A spreadsheet can be measured on its fields, and its fields do not contain outcomes.
5. Determinism is not validity: the link pass returns the same answer every time, which makes it reliable and not thereby meaningful. Reliable readings of the wrong thing remain the wrong thing.
6. Summary choices: a median is one decision about how to summarise a distribution, and the Accessories category shows how much a single number can hide. The range is reported alongside every median here for that reason.

Each limit has a remedy, and none of the remedies is a small edit. Selection bias needs a second platform and a second date. Threshold bias needs the rule published beside the count, which this site does. Survivorship needs a record of rows that were present earlier and are absent now, which requires two extractions. Outcome data needs a sample of parcels followed from order to delivery with a stated sampling rule, which is the largest piece of work in the list and the one this site is least likely to attempt. Determinism and summary choices are handled by stating them, which is what this section is for.

The value of writing the boundary down is that it can be checked. A reader who wants to know what this site measures can read the rules at /method/, run a link through the checker at /instrument/link-health/, and compare the output with the description above. If any published figure turns out to come from a method outside these two passes, it is an error in this article rather than a change in policy, and it should be reported as one.