Where the counts come from
Sheet roundups in this niche compete on volume. Two thousand links, five thousand, ten thousand, fifteen thousand, twenty thousand: the advertised figures climb, and the pages carrying them almost never say what was counted or how it was counted. This ledger reports its equivalent sentence in a deliberately dull way. The 2026-W39 snapshot holds ninety-three records, made up of thirty-one wrapped addresses gathered from public sheet pages and sixty-two source listings pulled out of a public directory. Each record is one row carrying one address, and the rule that produced the total was fixed before collection opened.
Where the four-figure totals come from is not a mystery, and nobody has to be dishonest to reach them. A wrapped address and the source address it encodes describe the same item, and a sheet that stores both has stored two addresses for one product. Count every address on that sheet and the total doubles before anyone has added a row by hand. Spread the same effect across the backup tabs people keep against losing a file, and the arithmetic leaves the hundreds behind quickly. Nobody edited the file between those two headlines, and nobody counted anything twice on purpose.
Rows compound the effect in a quieter way. Some sheets carry a single address per row. Others place the wrapped address, a fallback address and the marketplace listing into three columns of one row, which is how a seven-hundred-row sheet ends up holding more than two thousand addresses without a single duplicate inside it. Counting rows and counting addresses are both defensible measurements. Reporting one total as though it were the other is the step where a comparison quietly stops being useful, and both totals are correct answers to questions that were never written down.
Duplicates finish the job. Sheets get copied between tabs, mirrored, re-shared and uploaded again, so one product ID can sit in a main tab, a backup tab and two copies passed around a group chat. Distinct product IDs and total appearances are different measurements of the same file, and they can differ by a factor rather than by a percentage. The 2026-W39 sample sidesteps that argument by fixing one rule before collection starts and writing it down on the checking method page. Which rule was used matters more here than how large the sheet happens to be.
The counting rules we found
Four counting rules cover nearly everything published in this niche. The first counts addresses: every wrapped address and every source address is one unit, and the same item can supply two of them. Applied to the current sample, that rule turns ninety-three records into one hundred and twenty-four units, because the thirty-one wrapped addresses in group A also exist as rebuilt source addresses once the rebuild is done. Nothing about the underlying items changed between those two totals. Only the definition of a unit did.
The second rule counts rows. One row is one unit no matter how many address columns sit inside it, and this is the rule behind the numbers this ledger publishes. It is the most conservative of the four and the easiest to audit, because anyone can open a sheet and count populated rows without first deciding what those rows mean. Ninety-three records stay ninety-three under it, and adding a fallback column to the same sheet does not move the total at all.
The third rule counts only what answered, and here the sample splits sharply by layer. Ask whether the wrapped address opened and group A scores zero out of thirty-one, since every wrapped address failed at the network layer when it was requested during 2026-W39. Ask whether the listing behind it is still there and the same thirty-one records score thirty-one, because all thirty-one rebuilt source addresses returned a normal response. One rule reads the period as a wipeout. The other reads it as a clean sweep.
The fourth rule counts distinct products. Fold the sample down to product IDs and the ninety-three records still describe ninety-three products, because the wrapped set and the directory set were sampled from different places and do not overlap. Duplicates are a property of the sheets being counted, not of this sample. Two sheets advertised at four thousand links each can share most of one catalogue, and under this rule the pair is worth barely more than a single sheet. No edit is required to produce that difference either.
What the same sheet scores under each rule
Run the rules over the thirty-one records in group A and the spread stops being theoretical, because those records exist in both layers at once. Address counting makes them sixty-two as soon as each rebuilt source address is written in beside its wrapper. Row counting keeps them at thirty-one. Ask whether the source listing answered and the same thirty-one return thirty-one. Ask whether the wrapper answered and they return zero. One set of lines, four totals, and no edit of any kind in between.
At sample level the same effect reads as one hundred and twenty-four against ninety-three, because address counting adds the thirty-one rebuilds to the ninety-three records and row counting does not. The version quoted most often is the largest, and it is also the version that double-counts by construction rather than by accident. A four-figure total is not automatically wrong. It answers a question about the file, while a buyer is asking a question about the items, and those two questions sit one sentence apart in the same paragraph. The rule is doing the work there, not the data.
The buyer wants the small number, meaning how many things can be ordered out of this sheet. A catalogue builder wants the big one, because an address that timed out last Tuesday may answer next Tuesday and it still belongs in the file. A verification pass wants the reachable-only figure, which is always the smallest of the three and by far the most expensive to produce. None of those positions is dishonest. A page that prints a total without naming its rule leaves the reader guessing which one is on screen.
Settle it on your own file in about five minutes. Count the populated rows and write that figure down. Then count addresses across every column and write that figure beside it. Holding both numbers next to each other removes the ambiguity permanently, because any total someone else quotes can be checked against the pair without an argument. The sheet column mapper handles the messier half of that job, since a copied sheet rarely keeps the column order it was built with. Two numbers on one line, and the argument is finished.
Outliers and why they exist
Two outliers this period deserve naming, because each one moves a published total dramatically and neither is an error. The first is the one hundred and thirteen directory addresses that never completed. The host stopped accepting connections partway through collection, so those addresses were excluded rather than written down as unknown. Count them as links and the sample reports two hundred and six. Leave them out, as this ledger does, and it reports ninety-three. Same window, same host, two very different headlines.
That decision is correct for a status count and wrong for a size count. If the question is how many addresses a sheet contains, an address that timed out still exists in the file and belongs in the total. If the question is how many answered this period, it does not count at all. Both readings of the same one hundred and thirteen addresses are legitimate, and what separates them is a definition agreed in advance rather than anything the network did.
The second outlier is marketplace identity. Every source marker observed in group A pointed at Weidian, so all ninety-three records in this period resolve to a single marketplace. That makes the sample unusually clean and unusually narrow at the same time. A sheet built around Taobao or 1688 addresses would not behave identically under these rules, and this ledger holds no observed sample from which to say how it would behave instead. The gap is stated rather than quietly filled in, because stating a limit is cheaper than defending an assumption later.
Add the mirror problem and the picture is complete. Sheets that list an item twice for redundancy, rows that carry three address columns, catalogues re-shared with a new tab added at every pass: all of them push a published total upward without changing the catalogue by one product. When a page claims thousands of links, the follow-up question is short. Rows or addresses, and does an address that no longer opens still count toward the number. The weekly snapshots page prints the sample size next to the rule that produced it.