When two of your own numbers disagree, ask whether they describe the same rows

When two figures you produced from one dataset disagree, the first question is not which one is wrong. It is whether the two describe the same rows, counted by the same unit, over the same scope. Four such pairs are on record in our own measurement programme, and only one of the four was a defect. The other three were legitimate figures that had been given the same name. Knowing which kind you are holding takes about a minute and decides whether you are about to correct a number or correct a sentence.

The reason this matters commercially is that a reader who finds two of your figures disagreeing will assume the worse of the two explanations. Printing the reason beside the pair removes that assumption. We are paid to take these readings and to write the pages that carry them, which is why our own four pairs are listed by name rather than described in general terms.

The four pairs, and what differed in each

Every row below comes from work done on 15 and 16 October 2026 on ChatGPT and Gemini. Two engines, not five.

The pair	The two figures	What actually differed	A defect
The access run	176 attempts and 220 dated fetches	Scope. The smaller figure covers a single pass: eleven agent names put to sixteen public addresses. The second is that pass plus a later confirming reading	No
The citations recorded	20,143 and 36,576	The unit each one counted. One domain in one reading deduped against raw url occurrences, on identical rows	No
The own site share, first run	85.6% and 85.3%	A cleaning rule applied before counting, resolving provider fetch infrastructure back to the pages it served	No
The own site share before publication	91.3% and 85.6%	A classifier that could not match a whole class of hostname, so platform documentation sorted as vendor marketing	Yes

Three of the four needed a sentence. One needed a fix. That ratio is why the order below starts with the cheap checks.

Check the scope first, because it costs nothing

The access run is the clearest case. The same work is on record as 176 attempts, of which 23 were refused outright, and as 220 dated fetches. Both are correct. The smaller figure is the first pass alone. The larger one is that pass plus the reading taken hours later to see whether the refusals repeated. Anybody quoting one of those two without saying which has handed a reader a number that will not reconcile against the other.

Scope disagreements have a signature: the larger figure is usually the smaller one plus something that happened later, or plus a second population. Before anything else, write down what each figure counted over and compare the two descriptions rather than the two numbers.

Then check the unit, which is where the biggest gaps live

The 16 October 2026 run holds 20,143 citations across 4,947 distinct domains when a citation means one domain appearing in one reading, deduped. Count every occurrence of an address in those same stored rows instead, and the total is 36,576. Nothing was recounted and nothing was corrected. A domain quoted four times inside one answer is one citation under the first rule and four under the second.

Those two figures are not comparable in any sentence, and the gap between them is larger than most people expect from a definition. Every figure we publish uses the deduped rule, including the earlier run of 15 October 2026, because changing the rule between rounds would make a definition look like growth.

Then check for a rule applied before counting

The third pair is the one people miss, because both figures come out of the same script on the same rows. In the 16 October run, 712 of the cited urls belonged to the data provider's own fetch infrastructure rather than to anybody who publishes anything. With those addresses attributed to their real publishers, the own site share settles at 82.0% rather than 82.4%. Applied to the 15 October rows, the same rule reads 85.3% where 85.6% was published.

Two independent datasets, the same four tenths of a point. A correction that lands the same distance twice is itself a small piece of evidence that the cleaning rule is doing what it claims, rather than absorbing something it should have printed.

Only now is it a bug

The fourth pair is the real one. The same share stood at 91.3% before the correct figure, 85.6%, was established. A sorting rule insisted on a particular text fragment being present before it would accept a page as official documentation, which pushed platform documentation into the bucket labelled vendor marketing. There was no error output and no log line to read, which is why the existence of a second figure was the only thing that could have revealed it.

The repair was a precondition rather than a better pattern. The sorter now halts on any known example that comes out in the wrong bucket, so the next version of this defect stops the pipeline rather than publishing a share. A rule nobody has tried on a known case will misfile a whole category for as long as it runs.

The order, as four questions

• Do the two figures cover the same rows, or is one of them a later pass added on?
• Do they count the same event as one, or does one of them count every occurrence?
• Was a cleaning rule applied to one and not the other, before either was divided?
• If all three answers are yes, you have a defect, and the thing to look for is a rule that cannot match a class of input.

Three of those four questions are answered by reading your own notes. Only the fourth costs engineering time, and running the first three first is what stops a definition being fixed as though it were a fault.

What this page does not establish

Four pairs is four pairs. We hold no measured rate for how often a disagreement between two of your own figures turns out to be a definition rather than a defect, and we are not going to put one on this page by dividing one by four. What the four do establish is that the cheap checks are worth running first, because three of ours were settled by a sentence.

The pairs above also come from two products and nothing else. Three more surfaces sit inside what we sell and inside neither of these datasets. Which of our own datasets may be quoted for which question, and which pairs of them may never be added, is kept as a register of the four readings behind this site.

Questions about two figures from one dataset

Which of two disagreeing figures should be published?

Both, with the reason beside them. We publish 85.6% as measured and record the 85.3% that results from resolving the provider hostnames, because printing one and hiding the other removes the only evidence a reader has that the cleaning rule exists.

What is the fastest way to separate a definition from a defect?

Take ten rows by hand and count them under both figures' descriptions. If both descriptions reproduce their own figure on those ten rows, the disagreement is a definition. If one of them cannot reproduce itself, you have found the defect in ten rows instead of in a day.

Is a four tenths of a point difference worth printing?

Yes, and the reason is the second occurrence. One such shift is a rounding story. The same shift appearing in two independently designed datasets, 85.6% to 85.3% and 82.4% to 82.0%, is a check on the rule that produced both.

What counted as one citation in the 20,143 figure?

A domain scores once per reading, however many times it turns up inside that answer. Under that rule the 1,260 readings taken on 16 October 2026 produced the 20,143 figure. The same rows counted as raw url occurrences give 36,576, and the two totals are never set beside each other as though one had grown out of the other.

Will two versions of one figure cost us the reader?

An unexplained one does. A pair printed with its cause is the clearest evidence a report has that somebody is reading it. A reader who already suspects the publisher goes looking for exactly this, and finding the explanation already written is the outcome worth engineering for.

Two figures, one name, and the sentence that separates them

Every pair above was settled by naming what each figure counted, and three of the four needed nothing beyond that. The habit transfers to any number you publish about your own work: print the rows, the unit and the scope above the figure, and the next person who finds a second version of it will know which one they are holding. Our own invoices depend on readings like these being believed, which is exactly why the superseded value stays in print next to the current one. A reader who wants the search side of the same record will find two domains at aiknowsus.com/proof, each with its window above it.