Generated from the export, never typed: a proof page that cannot drift

No figure on our proof page was keyed in by hand. Each one is computed out of the Search Console exports when the page is built, which leaves no room for the text and the file to hold different numbers. That is one decision taken once, and it removes the most common failure in published evidence: a figure that was correct when somebody keyed it in and is now a little bit wrong because the file moved and the sentence did not. No amount of careful writing fixes that. A build step does.

The failure mode this removes

A typed figure has no connection to its source. The file it came from can be re-exported, corrected, extended or reclassified, and the sentence on the page will keep saying whatever it said. Three things then happen in order: the page disagrees with the file, nobody notices because both look fine on their own, and the discrepancy is discovered by a reader rather than by the author.

We have had the same class of defect in our own work, which is why this is not a theoretical preference. A classifier behind one of our published shares required a literal string to match before it would count a source correctly, so documentation from Google and from OpenAI was being counted as vendor marketing. The published figure was 91.3% and the correct one is 85.6%. Nothing threw an error, and the wrong answer looked exactly like the right one. Every source now passes through a single classifier with 24 named examples wired into it as a precondition, and the file will not load if one of them lands in the wrong bucket.

What generated actually has to mean

Reading the file at build time is necessary and not sufficient. Four properties decide whether the arrangement is worth anything:

Property	Why it matters
Every displayed figure is computed, not just the headline	A mixed page is a typed page with extra steps
The build fails rather than falls back	A fallback that renders is an error nowhere
The window and the export date come from the file too	Otherwise the dates drift away from the figures they scope
The arithmetic is asserted, not compared	Two readings sharing a blind spot cannot reveal it

The second row is the one that catches people. Our own proof page stat tiles rendered white on white in dark mode for a few hours because a style fell back to a value no token defined. It rendered, so no check failed. A fallback that produces output is the same family of bug as a missing lookup treated as nothing to check.

The fourth row, which cost us a real defect

One of our own gates was built to catch this and could not. Its method was to put the same page side by side under two identities, a named AI crawler and an ordinary browser, and raise an alarm if a figure present for one was absent for the other. Run from inside our own network the two requests came back equally wrong, so there was no gap between them to trip on. Two readings that share an assumption will agree with each other all the way to a false pass. The gate that replaced it states the figures that have to be in the document whatever the server knows about the visitor, and settles the question from a single request.

That is the general rule behind a page that cannot drift. Compare two things and you learn whether they agree. Assert what must be true and you learn whether it is.

What this is worth to a reader who does not trust you

More than the wording, which is the uncomfortable part. A paragraph about rigour is a claim. A page whose figures are computed from the file is a mechanism, and a reader can test the mechanism cheaply by coming back later and seeing whether the figures moved when the window did. Two engines asked for exactly this when we put the question to them on 15 October 2026: of 102 conversations with ChatGPT and Gemini, 99 wanted the method disclosed next to the result, and every one of them wanted a date, a last verified line and a stated refresh interval. A generated page can honour the refresh promise, because refreshing it is a build rather than an editing session.

Build it yourself or buy it

The work is a day for a simple version and the maintenance is the real cost, because the defects above are the maintenance rather than the API bill. If you are weighing that trade off, the five classes of silent wrong answer our own build produced are written up in build or buy an AI visibility stack, with the one that counted Google's documentation as marketing first. We sell the measured version of this and the pages that follow it, so take that as an interested recommendation to read the defect list before deciding.

Questions about building a proof page from its source data

Does this only work for Search Console data?

Any figure with a file behind it works the same way: an export, a log of readings, a price list. The principle is that the page reads the file rather than remembering it.

What happens when the export is re-taken?

The page changes and the dates change with it, which is the whole point. The thing to preserve is the previous version, so a reader can see what moved rather than finding a page that silently became different.

Is a spreadsheet embed the same thing?

It solves the drift and loses the scoping, because an embedded sheet carries no window, no export date and no explanation of which rows are being shown. The figures need to be computed into prose that still states its scope.

How do I stop a build quietly rendering the wrong thing?

Make the build fail on a missing input rather than filling a gap, and assert the facts that must be present rather than comparing two outputs. Both of our own failures in this area were checks that could not fail.

Is any of this visible to a reader?

One sentence of it is, and it belongs on the page: that the figures are generated from the exports rather than typed. The rest is visible over time, because a generated page and a typed page diverge the moment the data moves.

One decision, taken before the first figure went up

The choice to compute rather than type is not a quality improvement applied to a finished page. It is a decision about what the page is: a view of a file, or a document about a file. The first can be wrong only if the file is wrong. The second can be wrong in a second, independent way that nobody will find. For a page whose entire job is to be checked by somebody who has no reason to believe it, that difference is worth more than every sentence on it, including this one.