hipobuy finds spreadsheet: how a finds list gets built and pruned

The table records eight publicly visible finds lists against six fields. Everything in it was established by hand: entries counted as rendered, review dates taken as the most recent date found anywhere on the list, link shapes classified by resolving each address, and duplication measured on the first fifty entries by seller identifier.

We chose these six fields because each one changes a purchasing decision. Fields that would have been interesting but unverifiable, such as time spent curating or number of buyers served, were excluded rather than estimated. That exclusion is why the table has eight rows and not thirty.

Two conventions bound the whole exercise. Anything we could establish by hand went in; anything that required trusting a stated figure did not, which is why the table carries a duplication rate but no popularity column. And every value describes the list as rendered on the day we looked rather than as described by its maintainer, which makes a row a photograph of a moment rather than a description of a practice.

The genre also has a lifecycle, and a row reads differently depending on which stage its list is in. Collection comes first: entries arrive from browsing, from other lists and from reader suggestions, with almost no filtering. Consolidation follows, where duplicates are merged and dead pointers dropped, and the entry count usually falls. Organisation comes next, where the maintainer adds facets and moves items between them, and the count barely moves while the facet column changes. Pruning is the last stage, where items nobody has asked about in a while are removed. A list can sit in any one of the four for months, and the stage explains the shape of its row better than any single column does.

Entry counts in the sample ranged from roughly two hundred to over four thousand. Review dates clustered into three groups separated by about two months each. Link shape distribution leaned toward marketplace addresses in five lists and toward short addresses in three. Four lists referenced inspection material, four did not. Facet counts ranged from one to four. Duplication on the first fifty entries ranged from under five percent to over a third.

Reading the columns together rather than separately is the point. A high entry count with an old review date and a high duplication rate describes a list that accumulated rather than one that is maintained, and three lists in the sample had exactly that profile.

The columns are also meant to be read in pairs. Entry count against duplication rate tells you whether a list is large or merely repetitive. Review date against link shape tells you whether the pointers were checked, since a fresh date on a list of short addresses has not been verified in the sense a buyer needs. Facet count against entry count tells you where the maintainer spent their effort. What the table deliberately does not do is sum anything into a score, because a weighted total would hide the trade-offs the columns exist to expose.

One pattern recurs outside our sample as well, and it is that the six columns move at different speeds. Review dates move in steps, because they follow rebuilds rather than daily attention. Entry counts move slowly and often downward, because consolidation removes more than collection adds. Duplication rates move fastest of all, since a single afternoon spent copying another list can lift the figure substantially. Facet counts barely move at all, because facets are expensive to add and awkward to remove once readers depend on them. A row whose facet count has held steady across three reviews is therefore ordinary rather than neglected.

Entry count is a rendered count, not a stated count. Where a list stated a figure, we recorded ours and marked theirs as unverified; they differed in every case we checked.

Review date is the most recent date present. A list with no date anywhere is recorded as undated rather than as old, because the absence of a date is a property of the maintainer practice rather than of the content.

Link shape classification is per address, and a list is assigned the shape that dominates its sample. Duplication is measured by seller identifier rather than by item title, because the same seller frequently lists one product under several titles.

A QC reference counts only when it points at material about a specific item rather than at a general explanation of what inspection is. A list that sends readers to a page describing the inspection process has not referenced anything for our purposes; a list that attaches photographs or notes to individual entries has. We record the presence of per-item references and we do not score their quality, because the quality of inspection material is not something a rendered page can show us.

Facet count is the number of distinct regrouping dimensions a list supports, not the number of groups inside one. A list split into a dozen categories offers one facet. A list that can be regrouped by category, by destination region and by provider offers three. The column earns its place because facets answer questions the maintainer did not anticipate, while values inside a single dimension only answer the question they were built for.

Two measurement conventions apply across all six columns. Everything is counted as rendered on a desktop viewport, because some lists present a different structure to a narrow screen and a collapsed version cannot be counted reliably. And where a field cannot be established it is recorded as Not verified rather than left empty or estimated, with the row still standing. An incomplete row is more useful than a filled-in guess, and the gaps are as much a finding as the values.

The first outlier is the duplication rate above a third. That list stated the largest curated figure in the sample and rendered fewer distinct items than two smaller lists. Duplication, not size, was the reason.

The second outlier is the list with four facets and a small entry count. Facets cost effort to build and a small count suggests the maintainer spent their time on structure rather than on volume. For a buyer who filters rather than scrolls, that is the more useful artefact, and it is the reason we do not rank by size.

The duplication outlier deserves one more sentence than the table can carry. A high duplication rate is not automatically a fault; a list organised around sellers will naturally repeat one seller across several of their items, and that is a legitimate way to present a catalogue. What makes the outlier in our sample notable is that the repeats were the same item under different titles rather than different items from one seller. That distinction is invisible in the rate itself and only visible when you read the entries, which is why we describe the finding rather than publishing the number on its own.

The four-facet outlier points at the opposite failure mode. Structure costs effort, and a maintainer who spends that effort on organisation instead of volume produces a list that is worse for browsing and better for filtering. Neither is a defect. A buyer who filters should read the facet column first and a buyer who scrolls should read the entry count first, and the fact that two readers want different rows out of one table is the argument against collapsing it into a ranking.

Match the list to what you are doing rather than to what looks biggest. Buying one considered item: pick a list with a recent date and inspection references. Buying volume in one category: pick a list with a high facet count. Shipping a mixed haul: pick a list that shows link shapes, because that is where address problems surface.

If you only have time for one check, take ten addresses at random and resolve them. It is the highest-information ninety seconds available, and it works on any list regardless of how it presents itself.

Three steps, in order, and the order is what makes the ninety seconds work. Step one takes twenty of them: find the newest date anywhere on the list and read the entry count beside it. That pair alone eliminates most lists in front of you, because a large count with an old date is a list that accumulated rather than one that is kept up. Step two takes thirty: count the facets and decide whether the way you shop matches them. If you arrived with a specific item in mind, facets are irrelevant and you can move straight to the third step.

The third step takes the remaining forty seconds and it is the sampling step. Take ten addresses spread across the list rather than drawn from the top, resolve each one all the way through to its destination, and count how many arrive at the item the list promised. Do not stop at the first failure, because a single dead pointer is noise while a pattern is a finding. You finish with three facts about a list, all of them checkable by anyone else, and none of them dependent on what the list says about itself.

What to do with the three facts is the part the table cannot do for you. A list that fails the first step is not automatically useless, since a buyer hunting one particular item may still find it there, and a stale list occasionally holds the only surviving pointer to something nobody else kept. A list that passes the first step and fails the third is the more dangerous case, because it looks maintained and behaves like a museum, and a reader who trusts its date will spend an hour discovering what ten resolutions would have shown in forty seconds. Treat the three steps as a filter for where to spend attention rather than as a verdict on whether a list deserves to exist.