Retention automation — Hijama Associates

Workflow analysis

Where the money was leaking

Before

Customer booksBy message, by hand

→

Treatment happensThe part they were good at

→

⚠ Bottleneck Nothing No thank you, no review ask, no reason to return. The relationship ended with the appointment

After

Customer booksAcuity, self-serve

→

✓ Automated Confirm, prepare, remind Three touches before they arrive, none of them typed by a person

→

✓ The new part Review, then return One review ask, then nudges at 3, 5 and 8 weeks until they rebook

The insight the client had not priced

They were not short of customers. They were short of second visits. Hijama has a natural repeat cycle of a few weeks, so every lapsed customer was a booking the diary had already earned and then forgotten to collect.

Notes

What shaped the decisions

Why the revenue number is the honest one

December is their slow month, and the comparison runs November to December. The lift happened against the season, not with it.

The one rule the whole thing rests on

A new booking cancels every sequence that customer is sitting in. Without it, someone who rebooks on day 2 still gets “we miss you” on day 21, and the practice looks like it is not paying attention.

Requirements

What it had to do, and why it mattered

Five, and the last one is the reason it still runs.

01

One person, however many bookings

Customer records matched on phone number, not name or email.

A regular who books three ways becomes three people otherwise. Duplicate records are how automations send the same person the same message three times.
Identity
02

Every timer starts from their last visit

Not from a campaign date, not from enrolment.

A shared send date treats a customer who came yesterday the same as one who came in March. Relevance is the entire difference between a nudge and spam.
Timing
03

A cancellation is a beginning, not an end

Acknowledge, offer a rebook link, then prompt at 2 and 6 weeks.

Cancelling is not the same as leaving. Most cancellations are a diary clash, and nobody was following up on any of them.
Recovery
04

Ask for the review once

A tracked link sets a flag; anyone carrying it never sees the ask again.

The second ask costs more than the first one earned. Reviews come from people who feel appreciated, not pursued.
Reputation
05

Rebooking silences everything immediately

A new booking or reschedule cancels every sequence in flight.

This is the requirement that makes it survivable. One “we miss you” to somebody who booked last week undoes every message before it.
Credibility

Method

Design the states first, then the messages

Most retention campaigns are a list of messages on a calendar. This one is a state machine, which is why it never contradicts itself.

Step 1

Name every state

Booked, rescheduled, cancelled, done, reviewed. A customer is in exactly one, and the system always knows which.

Step 2

Clear before you set

Every workflow wipes prior state before writing its own. Stale state is what makes automations send nonsense.

Step 3

One exit, everywhere

A new booking drops the customer out of every running sequence. One rule, enforced in all five.

Findings

Three questions, answered in order

01When is a customer most likely to leave a review?

What it showed

Within the hour of finishing, while they still feel the result. The ask fires 60 minutes after the appointment ends, and only once. A customer who has already reviewed gets a different message with no ask in it, because asking twice is how you lose the third.

02When does a lapsed customer actually come back?

What it showed

Hijama has a natural repeat cycle, so the nudges sit at 3, 5 and 8 weeks from the customer’s own last visit rather than on a shared campaign date. The 8-week message carries an offer with a 7-day expiry, because a win-back without a deadline is just a newsletter.

03Why was a five star practice invisible on Google?

What it showed

The site led with the business name, not the thing people search for. Titles, meta descriptions and social tags were rewritten to open with the category term and the outcall angle, and to name the specific treatments people actually type. Competitors were ranking on exact-match domains; the fix was to match their title placement without owning their domain.

Architecture

Split the job in two, and it stops breaking

The automation decides who someone is. The messaging platform decides when to speak. Neither does the other one's job.

The automation owns state

Read the booking data
Decide booked, done or cancelled
Write that state onto the contact
Clear it the moment it changes

The messaging platform owns time

React to a state change
Hold the 3, 5 and 8 week timers
Send, and mark who reviewed
Drop anyone who rebooks

Why the split matters

Timers are the fragile part of any retention system, and they are the part that has to survive somebody changing their mind. Keeping every timer in one place, with exactly one escape hatch, is what makes the whole thing predictable.

The failure that shaped it

The first version broke in January

An earlier build let state live in more than one place, and long-running timers had no reliable way out. People who had already rebooked kept receiving messages written for people who had not.

The fault

State in two places

Two systems each believed they knew whether someone had rebooked. When they disagreed, the timer won.

The fix

One owner per fact

Exactly one system writes state. Everything else reads it. Disagreement stops being possible.

The guard

One escape hatch

Every waiting sequence checks the same single condition. A rebooking pulls the customer out of all of them.

What the rebuild carried in

391 lapsed customers were backfilled into the return sequence on day one, out of a base of 511. They had all been treated, none had rebooked, and until then nothing had ever gone back to them.

The messages

Eleven messages, each one earning its send

Written, not generated. Every one has a job, and the two that look similar are deliberately different.

The full ladder · the trigger is always something the customer did
When	What it does
Booked	Confirms the slot and tells them how to prepare
Rescheduled	Confirms the new time and repeats the prep
Cancelled	Acknowledges without guilt, offers a one-tap rebook
Morning of	Checklist and what to expect, sent at 8am
Done, first time	Aftercare, then the review ask on a tracked link
Done, already reviewed	Aftercare only. The ask is removed, because they already gave it
3 weeks	Why the cycle matters, and an open door
5 weeks	Social proof, other people are back
8 weeks	A dated offer that expires in 7 days
2 weeks post-cancel	Still want to come in?
6 weeks post-cancel	Final touchpoint, then we stop asking

The two that look identical and are not

A customer who has already left a review gets aftercare with no review request in it. It costs one extra branch to build and it is the difference between a practice that pays attention and one that runs a mailing list.

They were selling hijama. The page said their name

A five-star practice was invisible on the term its own customers type. The title tag opened with the business name, which nobody searches for.

Title tag before

Opened with the business name
Category term buried mid-string
No mention of outcall or home visits
Treatments not named individually

Title tag after

Opens with the category term
Location second, brand last
Outcall stated, the real differentiator
Each treatment named for long-tail search

The competitor read

The clinics ranking above them held exact-match domains, which cannot be taken. What could be taken was title placement, because Google reads a title left to right and so does a person scanning results. Meta descriptions and social tags were rewritten to match, then redeployed.

What shipped

What produced each result

Produced the 18–20%

The return sequence

Nudges at 3, 5 and 8 weeks from each customer’s own last visit, ending in a dated offer. 511 customers carried into it at launch.

Produced the 1,800%

The one-time review ask

Fires an hour after the appointment, with a tracked link that remembers who already reviewed so nobody is asked twice.

Produced the Ranked #2

The search rewrite

Titles, meta descriptions and social tags rebuilt to lead with the category term and the outcall angle, then redeployed.

What I would do differently

Instrument the sequence properly from day one. The revenue movement is real and the owner confirmed it, but the attribution is a month-on-month comparison rather than a controlled read. Holding back a small control group would have turned a believable number into a defensible one, and it would have cost nothing to do at the start.

Patients came once and were never contacted again

What was achieved

There was no follow-up after a treatment, so every month started with an empty diary and the same patients had to be won again.

What I personally did on this

Nobody sends these. The calendar does.

The proof, not the claim

Where the money was leaking

What shaped the decisions

What it had to do, and why it mattered

Design the states first, then the messages

Name every state

Clear before you set

One exit, everywhere

Three questions, answered in order

Split the job in two, and it stops breaking

The first version broke in January

State in two places

One owner per fact

One escape hatch

Eleven messages, each one earning its send

They were selling hijama. The page said their name

What produced each result

The return sequence

The one-time review ask

The search rewrite

How it was actually done