Notes
What to tell a client while a site is down
17 November 2026
The first message is not the postmortem. I used to wait until I understood the cause, and the client heard about the outage from their own customer. The useful note is the one that says what is broken, what I have actually seen, and when I will write again. The story of why can wait until it is true.
This is not the pager. Alerts to me, health checks, and the box I run are the hosting stack. The uptime block inside the monthly page is the uptime report. A public URL they can open without emailing me is optional and smaller, which is the status page. This note is the words, and the order of them, while the site is still wrong.
Three audiences, three sentences
I am talking to myself in the log. I am talking to the client in mail. I might be talking to anyone who has the status link. Those are not the same sentence. The log can say “Traefik returned 502, container healthy, started 14 minutes ago.” The client mail cannot, not as the first line. The status page cannot include the log, the work, or the secret report URL. If I paste one text into all three places, somebody gets either a lecture or a hole.
Who gets the client text is decided in onboarding, not during the outage. One address, named, plus a backup person if they have one. Not the whole company list. Not the address that files the invoices if that person is on leave. If I never wrote the name down, the incident is also a reminder that the first two weeks were skipped.
When the first note goes out
I do not mail the client on the first failed probe. A single miss is often a restart or a blip, and a mail about it trains them to ignore me. I wait until I believe it is real: two failures from outside the server, or I have opened the site myself and the money path is wrong. The plan should have named the window. If it did not, I use this and I tell them once, in a calm week, so it is not a surprise later: a note when I know the public site is affected, and another when it is back. Not a running commentary every five minutes.
That window is a choice, not a standard I can cite. A contract that says “within a minute” is a different product, and it needs a monitor built for that, which is the distinction on the uptime page. Do not let an incident rewrite the plan into a promise you did not sell.
The first note
Four lines. What is affected. What I have confirmed. What I do not know yet. When I will write again even if nothing has changed. The fourth line is the one I used to skip, and it is the one that stops them refreshing their inbox.
Subject: [site] is failing for visitors
The booking page is returning an error. The homepage still loads.
I have confirmed this from outside our office, starting about 14:10 UTC.
I do not know the cause yet. I have not changed anything.
I will write again by 15:10 UTC, even if it is only to say it is still down.
No cause. No “looks like the host.” No joke. No “just.” If I name a vendor I have not checked, I am writing fiction they may forward to that vendor. I have done that once. The follow-up was worse than the outage.
If I have already fixed it before the first note, the note is the recovery, not a drama in the past tense about a ten-minute gap they did not see. The monthly letter can still mention the gap. The live mail does not need a plot.
If they write first
Sometimes their customer is faster than my probe. The mail arrives as “the site is down,” with a screenshot and no URL. I answer the named person. I do not answer the six people they copied, unless the named person is the one who wrote and asked me to. I still use the four lines. I add what I have checked since their note, and the time of the next one.
I do not open with an apology essay. A paragraph about how seriously I take uptime, sent before I have fetched the URL, tells them I have not looked. If I have looked, the first sentence is the finding. If I have not, the first sentence is that I am looking now, and the time I will write again. Sorry can be the last line, once, when I caused it. It is not the structure of the mail.
I also do not argue with the screenshot. If they say checkout is broken and my fetch says the homepage returned 200, both can be true. I say what I fetched, and I say I am opening the path they named. Telling them the site is fine because the homepage loaded is how a true check becomes a false sentence they will quote back at me.
What you know, and what you are guessing
I keep a private split so the mail does not leak the guess.
| I can say it | I cannot say it yet |
|---|---|
| This URL returned this status at this time. | The host is having an incident. |
| The homepage loads and checkout does not. | No orders were lost. |
| I changed this one setting, and the page loads again. | It will not happen again. |
| I have not changed anything. | It is nothing to worry about. |
“No orders were lost” is the sentence people want to send because it calms the room. I only send it after I have looked. A checkout error does not tell me whether the order was written. Saying it early is how a calm mail becomes a false record. If I have not looked, the sentence is “I have not checked the orders yet. I will.”
If it stays down
I write again at the time I named, even when the update is “still down, still no cause.” A late update feels like hiding. An update that repeats the unknown is still an update. I add only what changed: a vendor status I have now read, a deploy I rolled back, a DNS change I can see.
I do not widen the audience because I am anxious. Adding their whole staff on message three means twelve people have three different versions of the story. I reply to the named person. If they want to forward it, the mail was written so they can.
I also do not start a redesign of the incident process in the middle of the incident. The status page, if they have one, gets the short public line: investigating, or identified, or recovered. It does not get the client’s name, the invoice, or my theory. Our status page is built to leave the work log out. Use that limit. A status page that turns into a diary is a second letter nobody asked for.
The public line, if a status page already exists
Three states, and I pick one. Investigating means I know the path is wrong and I do not know why. Identified means I know a cause I am willing to say in public. Recovered means the path works again, with the time. I do not add the amount of money, the name of the person who clicked publish, or a theory about the host. The status page already leaves the work log out. The outage is when I am tempted to put the log back, because a public URL feels like it needs a story. It needs a state.
Investigating. The booking page is returning an error. The homepage loads. Next update by 15:10 UTC.
Identified. A publish this morning broke the booking page. I am removing that change.
Recovered. The booking page has been loading since 15:02 UTC. It had been failing since about 14:10 UTC.
If there is no status page, I do not create one during the outage. A new URL in the middle of a bad afternoon is another thing for them to learn while they are trying to answer their own customer. The mail is the channel we already named in onboarding. A status page is a decision for a calm week. Alerts to me and notes to them are different pipes, which is the split on the hosting stack. I do not invent a third pipe at 2pm.
The recovery note
Short. When it came back, in UTC or in their timezone if I am sure which one they live in. What visitors would have seen. Whether I changed something. What I still have not checked.
Subject: [site] is back
The booking page has been loading again since 15:02 UTC.
From 14:10 to 15:02 it returned an error. The homepage kept loading.
I rolled back this afternoon’s publish. I have not yet checked whether any booking attempts failed in that window. I will look before the end of the day and write only if something needs you.
Then I actually look. If nothing needs them, I do not send a third mail that says “all clear” unless I promised one. The monthly letter will carry the gap. Another ping at 6pm for a gap they already know about is how the channel becomes noise. The monthly pass is where the 14 minutes belongs, in a sentence, not buried.
One clock, one subject
I pick one clock and I name it. UTC, unless I am sure of the timezone they work in. “3pm” from my laptop and “3pm” from theirs are how a fifty-minute gap becomes a fight about a different afternoon. The monthly letter uses the same clock as the live notes. If I convert later, I convert every time in the story, and I say that I converted. I do not leave one line in UTC and the next in a local time I did not label.
The subject stays stable. First mail: the site is failing for visitors. Next: the site is still failing. Last: the site is back. Each subject starts with the site’s name. I reply on the same thread. A witty second subject is a new conversation, and they will find it tomorrow under the wrong day, or not at all.
If one client has two sites and only one is down, the name in the subject is the site, not the company. I have written “your sites” and then spent the update explaining that the shop was fine and the brochure was not. The name belongs in the first words, so a forwarded copy still makes sense to someone who was not on the thread.
When I call instead
Email is the default because it is a record and because I can write it without performing calm. I call when waiting for them to read would make it worse. Checkout taking money and then failing. A login I do not recognise. Anything that looks like data leaving the site. A domain that will expire inside a week and I cannot renew it myself.
The call still gets a mail afterwards with the same facts. Memory is a bad incident log. I have misremembered a time by an hour and then put the wrong hour in the letter. The mail is the clock.
A made-up afternoon, so the shape is visible
This did not happen to a client of mine. I am walking through it so the templates have a body. A brochure site, Webflow, the contact form starts returning an error at 11:00 UTC. Homepage fine. I confirm it from a network that is not the studio. At 11:20 I send the first note: form only, cause unknown, next note by 12:20. At 11:40 I see a new embed in the footer that their editor published at 10:50. I remove it. The form works at 11:48. The recovery note says the form failed from about 11:00 to 11:48, the homepage did not, I removed an embed published this morning, and I have not changed anything else. I do not name their editor in the mail. If it keeps happening, that conversation is separate, the way the mixed-stack note already says: the letter is not a weapon.
The monthly line, later, is one sentence. “The contact form failed for about 50 minutes on 12 November after a publish. It has worked since I removed the new footer embed.” That is the whole record. No timeline graphic. No blame.
What I write down for myself
After recovery I add three lines to the roster, not a report. Start, end, and whether the cause was ours, theirs, or still unknown. If it is the third, I say unknown. A guessed cause that hardens into the roster will be repeated at renewal as if it were a fact. I would rather have a hole than a myth.
The lines look like this. I am inventing the row so the shape is visible. It is not a client, and the times are not a measurement.
Start: 2026-11-12 14:10 UTC
End: 2026-11-12 15:02 UTC
Cause: a same-day publish, rolled back. Not unknown.
When I do not know, the third line stays short.
Start: 2026-11-12 14:10 UTC
End: 2026-11-12 15:02 UTC
Cause: unknown. The homepage stayed up. The booking page returned an error. I did not change anything.
That third line is what I am willing to repeat in the monthly letter. If I cannot say it out loud without adding “I think,” it does not go in the roster. The monthly pass is where the gap is written for them. It copies these lines. It does not improve them.
If the same failure happens twice, the second recovery note includes what I will change in the plan: a tighter check on that path, or a rule about who may publish embeds, or a project to replace the form. The incident is not the place to sell the project. The next calm week is. I mention it so they are not surprised, and I quote it when the site is up.
A checklist I can run half-awake
- Confirm from outside. One probe is not a client mail.
- Named recipient only. No new audience mid-incident.
- First note: affected URL, what I confirmed, what I do not know, the time of the next note.
- No cause I have not checked. No “nothing was lost” I have not looked for.
- Update at the time I named, even if the update is “still down.”
- Recovery note with the window and whether I changed something.
- Call only for money, strange logins, data, or a domain inside a week.
- Mail after the call. The mail is the clock.
- Monthly letter includes the gap. It does not replace the live notes.
- Three lines on the roster. Unknown stays unknown.
The page they already know how to open is still the letter, not a new portal I introduce while they are annoyed. If they have never seen one, the sample digest is the shape, and the real one that month will have the gap in it. I would rather they see the gap than hear a polished version later and wonder what else I smoothed.
Sources
The timing in this note is a choice I use, not a published response standard. The product limits I am relying on are the ones already described for the status page and the uptime section.
- Client website uptime report. Probes and the uptime section. Not the client conversation.
- White-label status pages. Public state and incidents. Not the work log.
- The monthly pass. Where the gap is written after the fact.
Monthly agency tip
One tip a month.
One practical tip a month on proving your retainer is worth it. No spam.
I email a confirmation link before I add you. I store the address, this page, and the date. Unsubscribe any time. Read the privacy page.