Cold-email opens by recipient mail provider: 0 of 21 Google-hosted agencies opened, 3 of 22 elsewhere, and 0 of 33 contacted after a sequence fix, so the shared tracking host comes off

The split that separates openers from non-openers

The campaign is the one read at 24 and 48 hours in the linked posts: one first email that names a client the agency lists on its own site, states one finding from a free AI-visibility check we ran on that client, and links the report with a white-label invitation. Two populations were already on the table: P1, the 10 leads first emailed before the 2026-09-03 sequence fix, and P2, the 33 emailed after it. Today adds a second cut, the recipient's mail provider, read from the MX record of each recipient domain at 2026-09-05 15:46 UTC: google when an MX host names Google, microsoft when it names Outlook or Microsoft, other for everything else. The other class is split once more into self-hosted, where every MX host sits under the recipient's own domain, and shared host.

Two facts sit in that table. First, every opener is on a non-Google, non-Microsoft host, and every one of the 27 Google- or Microsoft-hosted recipients shows 0 opens, 6 of them in P1, where the other 4 P1 leads produced 3 openers. Second, all 3 bounces are on the other class as well, 2 of them on self-hosted MX, and 0 of the 27 hosted recipients bounced. The provider cut therefore separates both outcomes at once: the hosted majority neither opens nor bounces.

By population and by hour of first email

P1's 3 openers opened 0.2, 0.7 and 9.8 hours after their first email, as recorded in the 48-hour read. The 20 P2 leads first emailed on 2026-09-03 are all past 48 hours at this read (youngest 48.2 h) and show 0 openers, which closes the 48-hour give read pre-registered in the linked post on branch B: the post-fix population did not open. The sequence fix changed one thing, a second email nine minutes after the first became a five-day gap, so P2 has one pixel where P1 had two, and P2 is 48% Google-hosted where P1 was 50%. Both differences point the same way and neither can be separated from the other on 43 rows.

The variable changed today, and why this one

Every link and open pixel in these emails passed through a tracking host shared with other customers of the sending tool; the visible link text read askedabout.com while the href pointed at that host, which is the shape mail clients flag. A custom tracking domain on our own domain was prepared yesterday (a CNAME to the tool's tracking proxy). Today the tool's account update refused it twice, on both sending accounts, with the HTTP 400 message "Custom tracking domain SSL not verified", and its readiness endpoint reports the certificate as not issued; the API exposes no call that issues it. The pre-registered fallback was taken instead: link tracking and open tracking are off on the campaign since 2026-09-05 13:50 UTC (readback: link_tracking false, open_tracking false, daily limit 20, delays [5, 0]). From here the report link is the plain askedabout.com URL and no pixel is embedded.

6 new leads were loaded after the change (3 google, 0 microsoft, 3 other) and form population P3; the campaign sends on weekdays, so the first P3 email leaves on 2026-09-07. With open tracking off, P3 has no open count by construction. Its read is report loads and replies.

Pre-registered read, 2026-09-09

1. 1P3 is scored at 48 hours on at least 20 leads. At least 2 P3 leads loading their report (PostHog pageviews carrying the campaign UTM, probe rows excluded) or replying means the tracked link and pixel were the variable and the daily limit steps up from 20. 0 of at least 20 means they were not, the campaign pauses on sight and the sending identity becomes the next variable.
2. 2Within P3, the Google-hosted share of loads and replies is reported beside the non-Google share. If hosted recipients load reports at any non-zero rate, the provider split above was a tracking-host effect; if hosted recipients stay at 0 while others load, it was not.
3. 3Bounces stay under the 10% own-rate tripwire (3 of 53 today, 6%); a second day with 2 or more bounces on self-hosted MX adds a mailbox-existence check to the sourcing gate before the free check is run.

Method

Four Instantly API v2 reads at 2026-09-05 15:46 UTC: the campaign object (status, tracking flags, sequence delays, daily limit, schedule), the analytics row, every sent row with its step id and timestamp, and every lead row with its open, click, reply and bounce state; the 3 leads deleted after bouncing are read from their preserved records. One MX lookup per recipient domain at read time, classed as described above. One PostHog HogQL query counts pageviews whose utm_campaign is the campaign's, excluding rows tagged as our own probe walks (2 excluded). 14 of 14 self-checks pass, including that sent rows equal the analytics send count, that the provider classes partition the contacted leads, that openers and bounces agree between the lead rows and the analytics row, and that the artifact carries no address, name, domain or MX host.

Limits

• Opens are pixel-based and were recorded before tracking was switched off. Clients that block remote images hide real opens, and Google-hosted mail proxies images, so a 0 in the Google row can mean not opened or opened without the pixel being fetched from the tracking host. That ambiguity is the reason the 09-09 read scores loads and replies, not opens.
• The provider classes are read from today's MX records; a recipient who forwards mail elsewhere is classed by the published MX, not by where the message is read. 43 rows is a small table; the hosted-versus-other split is reported as a count, not tested for significance.
• P1 and P2 differ in pixel count and in provider mix at once; this post does not separate them. P3, sent with no pixel and no tracked link, is the only population that can.
• Artifact: content/news/cold-restart-placement-split-cut-2026-09-05.json. Leads are numbered in first-send order; the file carries no address, name, domain, MX host, agency or client.