Week 19: Cloudflare Email Routing, Wikidata citations, and the benchmark that keeps citing itself

Trust foundation continues. Real numbers from Bing AI, one 10-minute email setup, and a bug I keep punting. The scoreboard Two projects continue in parallel. Convertify shipped an entity foundation over week 18: Wikidata item Q140754389 for the brand, centralized JSON-LD schema graph with @id cross-references, IndexNow integrated into the sitemap generator, and an AI-crawler-explicit robots.txt. Week 19 was about seeing what those signals actually do. Bing Copilot AI Performance: up 3.1x quarter over quarter. The top grounding query was "latest HEIC file format comparison," 43.24% share. The page Copilot cites is our own image format benchmark, published back in June with real measured numbers from a Rust + libvips pipeline. That last part is the interesting piece. The content didn't change between July and August. What changed was how readable and verifiable Convertify became as an entity, on paper. I needed a serhii@convertifyapp.net for source verification on Qwoted, which is a HARO alternative for journalist requests. Cloudflare Email Routing does exactly this: enable it, add MX + SPF + DKIM records with one click, add destination address, add routing rule, done. Forwarding to my primary inbox in under 30 seconds. If you own a domain on Cloudflare and don't have a branded email yet, this is a genuinely free 10-minute upgrade. It also matters for pitch credibility. Journalists reading pitches from founder@yoursaas.com respond at a different rate than pitches from random@gmail.com. Whether or not that's fair is a different conversation. Still not fixed. It's a 16-bit alpha channel that doesn't get cast to 8-bit after libvips bandjoin, so background-removal PNGs sometimes leave the server at 30 MB when they should be 2-3 MB. Symptom masked in the browser because the frontend canvas re-encodes to 8-bit as part of a compositing step that isn't even needed for the transparent-background case. I did close the canvas re-encode issue this week (early return when bg is transparent, skip the useless composite pass). The backend cast is next. It's a two-hour fix that keeps getting bumped by higher-leverage work. That's a reasonable priority call for one week; two weeks in a row it becomes a pattern I don't like. The clearest takeaway from Bing's AI Performance data: measured content beats theoretical content when AI grounding pipelines need a source to cite. Every benchmark I've published quotes real file sizes from actual conversions on the same production pipeline that handles user uploads. Every other image-format article on the internet is quoting codec spec figures, not measurements. Copilot seems to prefer the measurements. Entity signals amplify this. The Wikidata item and JSON-LD graph don't add new content, they add proof of who is making the claim. That's what changed between May and August, and citation counts moved with it. None of this generalizes cleanly. Bing is not Google is not ChatGPT is not Perplexity. But the principle probably does: give AI grounding pipelines something concrete and verifiable, tie your identity to it, and see what happens. Second benchmark asset covering AVIF vs WebP at real-world quality settings, more journalist pitches, and yes, actually fixing the 30 MB PNG bug.