A second hostname served the same documentation in 325 readings, and most source tables would count it twice

Count hosts and you get a bigger number than you should. In the run we took on 16 October 2026, developers.google.cn was a source in 325 of 1,260 readings while developers.google.com was a source in 446. Those are two hostnames carrying one body of developer documentation. Any table that lists them as separate rows is telling its reader that an engine reached for two publishers when it reached for one, and the overstatement is 325 readings deep.

The fix is not complicated and it is almost never applied: before you count distinct publishers, decide which hostnames are the same publication, write the rule down, and publish the rule beside the total.

Why a mirror host is a counting problem and not a curiosity

A source table has two useful columns. One is the frequency of each row. The other is the total number of distinct domains, which is what tells a publisher whether the field is crowded or open. The second column is the one a duplicate host breaks.

Our own 16 October run recorded 20,143 citations across 4,947 distinct domains. The 4,947 is the figure a reader uses to decide whether there is room. Every duplicate hostname inside it inflates the apparent breadth of the field while narrowing the real one, because a reading that cited the mirror was a reading spent on documentation rather than on a page anybody could have written.

Three kinds of duplicate, and only one of them is honest to merge

What you are looking at	How to tell	What to do with it
A regional hostname run by the publisher	Same owner, same path structure, content intended for a region	Keep the row, and say in a note that it serves the same documentation as the main host
A proxy or a mirror run by somebody else	Open the cited address. It resolves to a page published by a different company	Resolve it back to the page it serves before counting, and never list it as a publisher
A subdomain doing a different job	Same owner, different purpose, such as a help centre against a developer site	Keep both rows. They are different writing for different readers

All three rules have been applied to our own data, and the middle one cost us a published correction. 712 of the addresses cited in the 16 October run belonged to our data provider's fetching setup rather than to any publisher: an iframe proxy and a reverse proxy, both handing over genuine Google documentation under a name of their own. Keep rows like that and a reader finds an unfamiliar proxy host sitting in a list of publishers. Point each one back at the page it hands over and the vendor own site share lands at 82.0% instead of 82.4%. The earlier run moved by the same distance, 85.6% to 85.3%.

The country hostname is the first row, not the second, which is why it stays in our table with the 325 attached. It is Google's own documentation under Google's own name. The honest treatment is to print it, and to tell the reader what it is.

The test to run on any source table somebody shows you

This takes a few minutes and it is the single fastest way to work out whether a published citation table was assembled carefully.

• Open three cited addresses from the top of the table. If any of them resolves to a page published by a company other than the one in the hostname, the table has not been cleaned.
• Look for a pair of hostnames with the same path shape. Two rows ending in different country suffixes with identical paths are one publication.
• Ask what the distinct domain total was counted on. Hosts, registered domains or publishers are three different totals, and a table that does not say which one it used cannot be compared to another table.
• Ask for the counting unit as well. Ours is one domain in one reading, deduped. The identical data counted by raw address occurrence gives 36,576 rather than 20,143, and the two figures are not comparable in any sentence.

That second check has a close cousin in ordinary publishing, where the question is whether two pages are the same page rather than whether two hosts are the same site. The decision rule is similar and is worked through in the two jobs a content duplication checker does.

What this changes about what you publish

For a publisher deciding where to spend, the merged view is the one that matters. On the merged view the documentation layer in our 16 October data is larger than the row order suggests and the number of distinct publishers behind the field is smaller. Both corrections point the same way: the questions worth writing for are the ones no operator documents, because that is where the field stays genuinely wide.

We publish these counts and sell the work of getting a business into them, so the counting rules get printed next to the figures rather than kept in a spreadsheet. The Search Console side of the same disclosure habit, two domains with their windows and export date on the page, is our proof page.

Questions about duplicate hosts in a citation count

How many readings did the second Google documentation host appear in?

325 of 1,260, on 16 October 2026, across ChatGPT and Gemini. The main documentation host appeared in 446 of the same 1,260.

Is a country hostname the same thing as a spam mirror?

No, and that difference decides what you do with the row. A country hostname belongs to the publisher and carries its documentation. A third party proxy serves somebody else's page under its own name, and 712 addresses of that kind came out of our own run before the shares were calculated.

Does merging duplicate hosts change the headline share?

Pointing the proxies back at the pages they hand over took the vendor own site share down four tenths of a point, to 82.0% from 82.4%. The direction of the correction matters more than the size: the clean figure is the one that gets published, and the difference gets recorded rather than quietly absorbed.

What should a distinct domain total actually count?

State the unit. Ours counts hostnames as the engines returned them, with provider infrastructure resolved away first, and the total is 4,947 for 20,143 citations. A table that gives a distinct total without naming its unit has given a number without a unit.

Which engines and which date is this?

ChatGPT and Gemini only, read on 16 October 2026 across 630 questions in three markets. Two apps produced every row above. Our product reads three further surfaces and none of them was asked these questions. The method behind the earlier run in the same programme is published as 597 readings across ChatGPT and Gemini in three markets.

Count the publishers, then count the hosts

The row that reads 325 is the reason to do it in that order. A table of hostnames answers a question about addresses. A publisher planning what to write next needs the answer to a different question, which is how many organisations were actually in the field, and the two answers differ by every mirror and every proxy you did not resolve. Write down which one your number is, and the next person to read it can use it.