Written for the person who has to explain a figure that shifted. Anything that changes how a number is computed appears here with the reason — because “the score dropped” and “we changed how the score is computed” are different conversations, and your client is entitled to know which one they are in.
Four kinds of entry. Shipped is new capability. Changed is behaviour that already existed behaving differently — the ones to read. Decided is a standing choice being locked, usually one we had been explicit about leaving open. Measured is a finding from our own runs that changed what we build.
Six engines are bundled in every tier — that is a pricing commitment and it does not move. Which adapters are sourcing answers today is a separate, more perishable fact, and the site was presenting only the first. The engines page now carries a status table generated from the same constant the monitoring loop reads, so it cannot stay flattering after the product changes.
The platform is deployed on its own domain with the public site, the documentation set and the first methodology page. Sign-in, tenant isolation and the worker tier are verified against the deployed application rather than locally.
Structured logs now reach a durable destination with a 30-day retention window. Until this landed, a failure that happened overnight was gone by morning — which made every reliability claim unverifiable after the fact.
A brand can now declare every domain it owns rather than one. Re-analysing live South African data moved one brand's owned citations from 6 of 618 to 31 of 618 — a 5.2× understatement that had been silent. If you configured a brand before this date, check its domain set.
Result rows are permanent; answer payloads are erased on a twelve-month window or on request, leaving a tombstone. This resolves the standing conflict between provenance (which wants everything kept) and erasure obligations (which do not).
Cost per run fell from $0.159 to $0.0337 — a 79% reduction — across 210 live runs, with the mention rate unchanged at 30/30 against 30/30. This is what makes an all-engines-included bundle survivable at agency pricing.
Every figure in the product now renders as a range with the number of observations behind it, and per-prompt output became a stability class rather than a percentage the sample could not support. Historical figures were not restated; periods before this date show the same underlying data in the new form.
Grounded model APIs perform a web search on only a fraction of runs. Recording which runs actually searched separates “the engine looked and you were not there” from “the engine never looked”, which were previously indistinguishable.
Every source in the citation table now opens the stored answer that cited it. This is the feature independent reviewers use to separate credible tools from dashboards, and it was worth shipping before anything decorative.
The agency's logo, accent and domain on a client-facing report that exports by printing the page it is already on — one component, one set of figures, one methodology footer, and therefore no way for the export to disagree with the screen.
Language explains roughly a third of the variance in AI answers, and four South African languages have been live in Google's AI surfaces since March 2026. Afrikaans ships as a first-class prompt language rather than a translation layer. isiZulu follows.
Checks are ordered by demonstrated effect: per-bot crawler access first, then server-rendered content, then chunk structure, then schema. llms.txt is reported as informational and not scored.
Changelog entries that affect how a figure is computed go out to every agency on the platform before they land, not after.