We believe in "measure continuously, honestly flag insufficient data" as a methodology โ so it should apply to ourselves first. This is a plain record of what we changed on our own site, when, and what's still missing. Not marketing copy โ an experiment log you can check against the commit history.
Why publish our own build process as an "experiment"
Most "how we work" pages show only conclusions, never the process. We're doing the opposite: recording each real batch of shipped changes as-is, including the less flattering detail that a given batch took one day. This extends the same white-hat principle we keep repeating โ the methodology says "flag insufficient data honestly"; here it's "flag a compressed timeline honestly" โ we're not pretending this was a steady, quarter-paced product cadence.
2026-07-24 ยท Wave 0: closing semantic-layer gaps
This batch answered "can an AI engine understand who we are in one fetch":
- Shipped llms.txt โ a bilingual site-description file for AI crawlers, with engine counts interpolated from the lib/engines.ts single source of truth, never hardcoded
- Completed structured data site-wide: Organization, SoftwareApplication, FAQPage JSON-LD
- Expanded the free checkup's monitored engines from 3 to 5 (aligned with the weighting rules in how we score)
- Upgraded homepage copy to cite the public University of St. Gallen and Digiday findings, replacing vague "we're professional" phrasing
2026-07-24 ยท Wave 1 batch 1: opening a door for AI agents
This batch was the "action layer" โ not just letting AI engines read us, but letting AI agents call us directly:
- Shipped an internal tool registry (G1 rule: tool parameters never accept tenant/user identity fields โ identity comes only from the API key)
- Shipped a public MCP endpoint (Model Context Protocol, standard JSON-RPC 2.0) plus a developer docs page, with a first batch of 5 read-only tools
2026-07-24 ยท Wave 1 batch 2: cold-starting the content engine
This batch stood up the resource hub: a bilingual content-hub skeleton (the site's first "URL defines language" hreflang structure) plus six pillar articles โ including the "coming soon" placeholder in this very category, filled for the first time by this article.
What we haven't done yet: cross-engine answer-diff tracking
The plan calls for asking the same question to engines on both the Chinese and Western sides and laying the differences out side by side. Honestly, as of publishing this article we have zero real cross-engine comparison data on hand โ actually calling ChatGPT, Kimi and friends costs real money and needs explicit internal authorization before we run it.
We could have written a "plausible-looking" sample answer into this article right now. That is exactly what we criticize about this industry โ "when measurement is a black box, scores are guesswork." Fabricating a finding would violate our own position worse than simply not publishing this section yet. So it stays blank for now โ once we actually collect the data, we'll publish a separate article. No promised date. Only a promise that when it ships, it's real.