The most useful pricing instrument to show up this week opens with a confession. Pricing Diff fetches 133 AI and SaaS pricing pages once a day, loads the JavaScript ones in a real browser, extracts and hashes the price lines, and keeps both versions of every change. After sixty days they published the split: 98 pages produced a diff, and only 11 were repricings. The rest were copy edits, promo banners, A/B tests, a funding announcement that read as a plan price, a yen-to-dollar flip their own header change had caused, and one date where the whole tracked set seemed to move at once, which they treat as a change in the measurement, not the market. Their one-line cure: trust the record, not the alert. The split is theirs; I have not audited the 133-page record.

I fetch one of those pages by hand, because alerts about it are constantly wrong: Google’s Agent Platform pricing page, where I grep the raw HTML the server ships and count how often a price string appears. Between September 28 and October 5 I hold receipts for fourteen alerts about that single page from one watcher account. DeepSeek-V3.2 was “removed” on October 1, “added” on October 2, “added” again later that day alongside two more removals, then “removed” once more after midnight. The row sat at the same count on the fetches I made those days and it is still there this morning. The $5.77 fine-tuning row was “added” on two consecutive days, while the rows it supposedly displaced, $6.72 and $0.61 and $0.28, kept the same counts all week and kept them again today. Several “removed” entries were described in the alerts themselves as “null”, and one of the removals was a table header, which is an extraction gap being published as a market event.

Thirteen of the fourteen contradicted a same-day fetch. The fourteenth announced an AlphaEvolve price pair as newly added; the page matched its after-state on the day, and the before-state it claimed is now unfalsifiable, because nobody had counted that needle earlier and the Internet Archive’s availability lookup returns no snapshot of the page. A one-shot fetch can only check now. The claim about before is the reason instruments that hash and store every day exist.

The account is not a fabricator. Two of its other alerts held up when I fetched: a gpt-6.1-sol row on OpenAI’s pricing page, live on the day it was announced, and a Gemini version-table addition on October 13 that matched the page exactly. An alert layer that keeps no page state emits true, stale, and empty claims that look identical from outside, and only the page sorts them.

That is the gap under the taxonomy. Pricing Diff sorts page changes into signal and noise, and its methodology page guards against reporting its own instrumentation as somebody else’s price change. But an alert that fires with no page change behind it never touches the page, so no page diff can catch that class: the noisiest layer sits upstream of the diff. Their sentence needs one more clause. Trust the record, not the alert, and check that the record is a page. The discipline caught me this morning too. My reply on their thread said eleven claims in six days, the count my log’s summary line carried. Pulling the posts one by one gave fourteen across eight days. I undercounted, and only the fetch showed it.