Home › Deliverability Hub › Top-Level Domains
3,415,449 sends over 30 days, every one we made, split by the TLD they went out on and scored against Google’s permanent domain-reputation verdict. It landed on 0.177% of sends overall — and a .link domain drew it 16 times more often than a .com sending the same copy to the same list.
0.177%
Of 3,415,449 sends refused on reputation
16×
.link against .com, the same copy and list
43%
Worse reply rate on a domain once it is refused
51 of 3,493
Domains past the retirement line. 96.4% are clean
This is not a fleet-wide collapse. 3,366 of 3,493 sending domains drew no refusal at all. 127 drew at least one, and 51 sit at or above the 5% line where we retire a domain. The problem is 1.5% of the estate.
It is a sourcing problem, not a sending problem. The single best predictor of whether a domain draws this verdict is the top-level domain it was registered on — not the platform it sends from, not the copy, not the list. That survives holding the client constant, which is the test that rules the other two out.
And unlike a blocklist listing, it costs results. Our SURBL study found a listing moved reply rate from 1.14% to 1.16%, which is nothing. This verdict moves it by 43%. The two sound alike and should not be treated alike.
Several campaigns hit volume ceilings in the same fortnight, with domains going quiet and bounce rates climbing. The question put to us was whether the infrastructure was at fault.
The verdict behind it is one of the few in email with no appeal. Google returns 550 5.7.1 … very low reputation of the sending domain, it is permanent, and it attaches to the domain. Warming does not clear it. Moving the mailbox to another tenant does not clear it. There is no form to fill in and no counterparty to argue with, so prevention is the only lever there is, and the only questions worth asking are what it costs and what predicts it.
An earlier look at this had eliminated seven candidate causes and left copy and audience standing. It was five weeks old and predated a round of retirements aimed at this exact verdict, so it needed redoing rather than citing.
| Population | Every send we made in the window: 3,415,449 across 34 client workspaces with 5,000 sends or more. Not a sample. |
| Signal | The bounce text itself, matched the same way our daily retirement scan matches it, so the study and the automation cannot disagree about what counts. |
| Window | 30 days to 8 October 2026 |
| Rate floor | A domain needs 50 sends before it carries a rate. The verdict can only fire when the recipient is on Gmail or Workspace, so a domain with three sends and one refusal has measured who it wrote to, not its own reputation. |
| Denominator | Sends, not bounces. Against bounces, a quiet domain whose few bounces happen to be Google-issued reads as catastrophic and a loud one reads as fine. The bounce analyzer splits a real mailbox the same way. |
| Not measured | Placement. 98% of sends produce neither a bounce nor a reply, and inbox, spam folder and silent discard are indistinguishable inside that. This study measures refusals, which are loud, and says nothing about the quiet majority. |
Every bounce in the window, by cause. The reputation verdict is an eighth of them.
| Outcome | Count | % of sends |
|---|---|---|
| Ordinary bounces — bad address, full mailbox | 27,786 | 0.814% |
| Recipient-side gateway spam blocks | 14,097 | 0.413% |
| Google reputation refusals | 6,061 | 0.177% |
| All bounces | 47,944 | 1.404% |
The concentration is the load-bearing fact. Of 3,493 domains carrying enough volume to be rated, 3,366 drew no refusal at all, 127 drew at least one, and 51 sit past the retirement line. An estate that is otherwise being accepted has 51 domains doing the damage.
Our own estates against bought-in stock. Read the two zeroes with the note underneath.
| Estate | Source | Sends | Refusals | Rate |
|---|---|---|---|---|
| Azure, estate A | Built by us | 1,640,757 | 4,693 | 0.2860% |
| Azure, aged estate | Built by us | 579,788 | 1,332 | 0.2297% |
| Google, estate A | Built by us | 43,364 | 21 | 0.0484% |
| Google, estate B | Built by us | 172,074 | 16 | 0.0093% |
| Google, bought-in reseller stock | Bought in | 721,109 | 0 | no visibility |
| Third-party SMTP | Bought in | 258,467 | 0 | no visibility |
Our own Google stock refuses at 0.0093%, about 31 times cleaner than our own Azure stock at 0.2860% — while carrying roughly a tenth of the volume.
The two zeroes are not clean records. Those estates send through infrastructure that never returns this verdict to us in a form we can match, so zero there means no visibility. That is 979,576 sends a month, about a quarter of everything we send, unmeasured on this question. The same gap was open at the top of the SURBL study in July and it is still open.
Across the whole fleet first, then held inside single clients.
| TLD | Sends | Refusals | Rate |
|---|---|---|---|
| .link | 197,451 | 2,619 | 1.3264% |
| .org | 848,956 | 1,497 | 0.1763% |
| .co | 528,352 | 775 | 0.1467% |
| .com | 1,400,294 | 1,171 | 0.0836% |
| .info | 428,996 | 0 | no visibility |
.link refuses at 16 times the rate of .com, and carries 43% of every refusal on 6% of the volume. The .info zero is the visibility gap again, not a clean TLD.
The objection to a fleet table is that different TLDs carry different clients, and so different copy and different lists. Inside one client both are constant.
| Client | .com | .co | .org | .link |
|---|---|---|---|---|
| Client A | 0.0007% | 0.0484% | 0.3795% | 2.9120% |
| Client B | 0.0003% | 0.0098% | 0.0773% | 0.0000% |
| Client C | 0.0730% | 0.0070% | 1.0952% | — |
Inside Client A, the same copy to the same list refuses 4,160 times more often from .link than from .com. Inside Client C, .org refuses 15 times more often than .com.
That is what rules out copy and audience. Both are held constant inside a client, and the gap is still there. Copy still explains which campaigns draw recipient-side spam blocks — a separate and larger problem, below. It does not explain this verdict.
This is where it parts company with the SURBL finding.
| Domains | Count | Sends | Replies | Reply rate |
|---|---|---|---|---|
| Clean | 3,366 | 3,281,994 | 44,232 | 1.348% |
| Carrying a refusal | 127 | 130,878 | 1,002 | 0.766% |
A domain carrying this verdict replies 43% below a clean one. In the window, 130,878 sends went out from burned domains; at the clean rate those would have produced about 1,764 replies rather than 1,002. Roughly 762 replies a month sit behind the gap.
One caution on causation. A domain whose reputation has already slipped will plausibly show both a lower reply rate and the refusal, so this gap is not proof that the refusal causes the loss. It is good enough for what the number is used for, which is deciding which domains to retire, and not good enough to be called a cost of the verdict itself.
Every cause measured the same way over the same window: delivery lost as a share of all sends.
Two of those four are content problems rather than infrastructure ones. Bad addresses are list quality, and they cost 4.6× what this verdict costs — start with the email verifier and the CSV list cleaner before anything here. The gateway blocks are what the copy and the offer draw, and they are 2.3× the size of it; the segmentation playbook covers those.
The SURBL figure is computed from that study’s own published table rather than from this one, so the two pages agree. It is the smallest of the four by a factor of 7, and it is the one that cost no replies at all.
The five heaviest rates of 34 measured. 6 sat at exactly zero.
| Client | Sends | Refusals | Rate |
|---|---|---|---|
| Client D | 23,426 | 192 | 0.820% |
| Client E | 20,237 | 164 | 0.810% |
| Client F | 9,575 | 76 | 0.794% |
| Client G | 81,871 | 538 | 0.657% |
| Client A | 427,001 | 1,544 | 0.362% |
Client A carries 25% of every refusal in the fleet in absolute terms, which is why it reads at first as that client’s problem. It is not. Its rate is middling and its .com stock is among the cleanest in the estate; it simply sends the most, and it holds the largest block of .link.
We stopped buying the TLDs that draw it
No new .link or .info. The policy is set, but the estate is not unwound: 68 domains carrying 1,387 .link mailboxes are still placed with live clients, and they are retired on a schedule now rather than one at a time as each one burns.
Detection moved to a daily scan
The scan reads the bounce text across the whole estate every morning and retires anything past the 5% line. It retired 535 domains inside this window, and the fleet rate fell from 0.445% of sends a month earlier to 0.177% now, a factor of 2.5.
The visibility gap is not closed
979,576 sends a month go out through infrastructure that never returns this verdict to us. We cannot say those estates are clean, only that we cannot see them. It needs a platform answer, not a process one, and it has been open since July.
Placement is still uninstrumented
Refusals are loud and we now count them. The 98% of sends that produce neither a bounce nor a reply are not instrumented at all, and inbox, spam folder and silent discard are indistinguishable inside that.
If you are buying sending domains, the TLD is the cheapest decision you will make and the one that predicts this verdict better than anything else we measured. Buy .com. The ones that look like a bargain are priced that way because of who else is on them.
If you already have an estate, the work is finding which domains are carrying it before they take the campaign down with them. Check expiry and ownership across the whole list at once, and check what else is flagged against them.
Read next
Google, Microsoft 365 and Azure inboxes with free prewarmed domains and DNS configured before you log in. Prewarmed inboxes send the same day you order.
Get Prewarmed & Fresh Inboxes