The short version
In October 2025 I took over the product and engineering team at a weight-loss and wellness telehealth company. This is an honest account of the thirteen months since: how the market moved, what we shipped, how the team changed, and which numbers did not go my way.
Delivery clearly got better. We merge more than twice as many changes a month, our automated test count went from a few dozen to nearly a thousand, and releases run through a candidate, a preview and an automatic rollback instead of a push to main. The commercial side is harder: new-account conversion roughly halved, the business now runs mostly on renewals, and our biggest onboarding rebuild moved people forward without moving purchases.
Where things stood in October 2025
The product worked and patients were being treated, but most of what makes a product team fast and safe was missing or manual. The retention work from my first months is in the head of product case study. This is the delivery picture.
- One onboarding flow. It lived inside the patient app. Our first platform for testing onboarding versions went live later that month.
- Releases on push. The API deployed on every push to main, tags were occasional, and release notes were posted by hand.
- Almost no automated testing. 36 unit tests, no end-to-end tests and no QA automation.
- Tracking we could not fully trust. About 69 server event definitions, checkout events that sometimes lacked key properties, and no campaign source on 91% of that month's new accounts.
- A thin team around the engineers. About one product manager, three to four engineers merging code in a given month, design shared with marketing, and a website run by an outside team.
Commercially, new patients mattered more then: about 21% of payment events were first purchases. By August 2026 that share was about 12%, and lower again in a partial October.
How the market shifted
Three forces reshaped weight-loss telehealth this year. Branded GLP-1 prices fell and new direct channels opened. Regulators moved from warning about compounded GLP-1s to citing how telehealth companies market them. And peptide and longevity treatments opened up on footing that can still be withdrawn. Everything in this table is public.
| When | What changed | Why it mattered |
|---|---|---|
| Oct to Nov 2025 | Warehouse and big-box pharmacies added branded GLP-1 injections at cash prices. Federal pricing deals with both GLP-1 manufacturers set cash prices near $350 a month. Google allowed drug names in US ad text. | The price anchor telehealth sold against started to fall. |
| Dec 2025 | A federal order against a telehealth GLP-1 membership business over hidden costs, unsupported results claims and fake reviews, requiring easy cancellation. 35 state attorneys general pressed Meta on AI-generated weight-loss ads. The first oral GLP-1 for obesity was approved. | All-in pricing and easy cancellation became the floor, and paid social creative came under pressure. |
| Jan to Feb 2026 | The first GLP-1 obesity pill launched at $149 to $299 a month. A government cash-price drug site launched. FDA said it intended to restrict ingredients used in non-approved compounded GLP-1s. | A branded pill at a cash price reset what a fair monthly price looked like. |
| Mar 2026 | 30 FDA warning letters to telehealth companies over "generic" or sameness claims and telehealth brands on product labels. A leading telehealth brand stopped advertising compounded GLP-1s. | Copy and imagery became compliance surfaces. |
| Apr 2026 | A second GLP-1 weight-loss pill launched, with same-day online pharmacy delivery. FDA proposed keeping semaglutide, tirzepatide and liraglutide off the list outsourcing facilities can compound from (not final). | More branded, cheaper, faster options. |
| Jun to Jul 2026 | 25 more FDA letters, again citing labels carrying the telehealth brand. A Medicare GLP-1 bridge opened with a $50 copay. An FDA advisory panel narrowly backed compounding six peptides. | Brand imagery on products became a direct risk, and peptides moved toward the mainstream. |
| Aug to Sep 2026 | That same telehealth brand reported a sharp fall in gross margin and plans for peptide and longevity products. Telemedicine ad-certification applications were up 101% year over year. Large employers signalled GLP-1 coverage cuts for 2027. | The leader was diversifying, paid channels got more crowded, and more people would pay cash. |
| Oct 2026 | HHS said it would not enforce against compounding those six peptides under set conditions, a revocable policy. | A real but fragile opening for new lines. |
Sources: White House pricing deals, Google ads policy, FTC order, first oral approval, FDA, February, FDA, March, FDA, April, Medicare GLP-1 Bridge, HHS peptide policy.
Our own numbers pointed the same way. Paid social was getting more expensive, and first purchases were a shrinking share of our payments.
The way I came to put it: we were not a growth business with a retention problem, but a retention business with an acquisition funnel attached. The roadmap had to say so.
What we shipped, quarter by quarter
The team opened roughly 895 pull requests in thirteen months. Merged changes ran at about 30 to 40 a month through July, then about 80 to 90 a month from August. One copy of the repo history I used is thin for February to April, so read that stretch as possibly under-counted rather than quiet.
Q4 2025: experiments, tracking and upsell
- Onboarding testing platform. Weighted version splits from settings, an A/A test of the split itself, URL variants and an admin editor. From then on, the major onboarding versions could run side by side as split tests.
- Upsell across three apps. An add-on banner for patients, add-on pricing and discounts for admins, pre-filled add-on requests for doctors, an upsell step in onboarding, and a one-time-purchase flow.
- Server-side tracking. Patient traits on events, the onboarding version on checkout and profile events, and ad click identifiers from four ad platforms, plus cookies from a fifth.
- Platform. Database-backed job queues, structured logging, and an automatic notice for every production deploy.
Q1 2026: data safety, payments, pharmacy operations and identity
- Data safety. Database backup and restore pipelines and an environment sync tool.
- Payments. Billing rebuilt to run on a second payment processor, so we no longer depend on one provider.
- Pharmacy operations. Product-to-pharmacy mapping with validation, and an outbox for pharmacy submissions.
- Identity. One anonymous ID across our domains, so a marketing-site visitor and the same person in onboarding can be joined up.
Q2 2026: new lines, retention events and an onboarding service
- New lines. A new pharmacy partner's catalogue, a standalone onboarding for our first add-on line, and a visual redesign of the main flow.
- Retention events. Card expiring, cancellation-flow steps, active subscriptions on the profile, a "No Decision" event, failed-payment value and a first-time flag on payments. How we structure these is in the event properties guide.
- Support. An audit log for onboarding versions and a triage router into the support chat.
- Architecture. Onboarding moved into its own service, and every pull request got its own environment.
Q3 2026: rebuilt patient surfaces and a release system
- A rebuilt patient portal (August). Redesigned follow-ups, treatment history, order and shipment tracking, identity re-verification, and scheduled treatments shown before the first charge.
- Onboarding reliability. Onboarding completes from the payment provider's webhook rather than the browser, and its events go through a durable outbox.
- Rebuilt onboarding (late August). A separate app, first on a small share of traffic and later on about three quarters. The results are in rebuilding onboarding: what the test really showed.
- Fewer stalls in onboarding. In late September we took a physician consult messaging step out of onboarding after reviewing a sample of consult conversations: about three in four patients stalled after long doctor messages.
- Integrations. API and webhook integrations with more pharmacy partners replaced polling for orders and shipments.
- Releases. From late July, lane-based releases with immutable tags, a QA candidate, a production preview, smoke tests, automatic rollback and parallel hotfix trains. We shipped to production on 11 separate days in August.
October 2026: the existing patient becomes the product
- In-portal qualification. Existing patients add a treatment without redoing onboarding, and doctors review their answers.
- Retention and billing. One-time retention discounts in the cancellation flow, failed cards that no longer block future charges, and automatic payment retries.
- New lines and retail. New onboardings for two of the newer lines, and onboarding versions for an in-person retail channel.
The platform underneath
| Area | October 2025 | October 2026 |
|---|---|---|
| Apps and packages | 4 apps, 5 packages | 9 apps, 10 packages |
| Release model | Push to main deploys; ad hoc tags | Lane releases with candidate, preview, smoke tests and rollback |
| CI workflows | 4 | 18 |
| Environments | Local and staging | Per-pull-request, QA and local cloud emulation |
| Automated tests | 36 unit tests | 972 tests in the main repo, including end-to-end journeys, plus 126 QA-repo tests across about 11 device profiles |
| Server event definitions | About 69 | 86 |
| AI in the repo | One agent guide, 6 commands | 4 agent guides, 12 agent skills |
More tests are not yet a safer pipeline. This autumn the suites still did not gate merges or releases, and making them do so is the next job.
The release pace this produced is in a release every four days. That average hides a ramp: early in the year we deployed on only a few days a month, and the pace picked up from April and especially from July.
How the team changed
The team grew less than the output did. What changed most was how work entered, moved and left.
- October to December 2025: take over delivery. I opened a dedicated releases channel in week one and ran a daily product standup, a weekly KPI review and a sync with lifecycle marketing. A first PM took on new products and support, we hired an in-house web developer, and contractors started QA automation. Output stayed flat. Changing how a team works is slower than changing what it works on.
- January to May 2026: stabilize. A second PM took conversion and published a framework for it. In January an engineer proposed a fix to our review flow, an early sign that the team was starting to own its process. A third PM lane followed in April. Early in the year, releases were still fragile.
- Late May to August 2026: build the operating system. One metrics dashboard on server-side events, a daily anomaly monitor, tiered design and copy sign-off, and a reorganized tracker in which 5,100+ tasks in 49 unstructured lists became three PM-owned groups. One intake form routes each request to a PM in about ten minutes, and an approved project gets a seven-step kickoff bundle, with a compliance review for anything patient-facing. Two engineers and a senior copywriter joined. Tracked items closed per month roughly doubled from June, though that count is noisy and includes one-off cleanups.
- August to October 2026: communicate and govern. A three-tier weekly update, a tracking check on every release, and an engineer proposing eleven new production monitors.
| Role | Late 2025 | October 2026 |
|---|---|---|
| Product management | One PM lane | Three PM lanes: conversion; portal and project coordination; new products and support |
| Engineering | Three to four active a month | About four a month, six across the year |
| QA | A lead and manual testers | A lead, about four manual testers, two automation contractors |
| Design and copy | Shared with marketing; about one writer | Two product designers; a copy manager and a senior copywriter |
| Website | An outside team | An in-house developer |
By mid-year I led a team of 16; how I approach that is in building and mentoring a team of 16. One weakness remains: output and launch-night load are still concentrated in a few people.
How we adapted product and channels
Most of what we shipped answered one of four pressures.
Cheaper branded options: earn more from the patients we have
Once branded pills and cash prices reset what a fair price looked like, new patients got more expensive to win and easier to lose. So we leaned harder on the existing base, which we had started doing with add-ons in December.
- Add-ons first. Most buyers of our first add-on line were already patients.
- Qualification in the portal. In the weekly read after launch, add-on purchases from the portal card rose about 70% week over week, on a small base. In the first month, 43% of enrollment starts came from the portal card and 14% from campaigns.
- Price flexibility. A one-time-purchase option and single-use retention discounts.
Enforcement: treat copy and imagery as product
- Imagery. When public warning letters started citing product imagery, we audited every surface that shows medication and made neutral imagery the default for new work.
- Copy. Compliance copy went live on the website in August, and in October we applied our banned-claims list to the product descriptions shared across apps.
- Ads. We took weight amounts and percentages out of ad claims, and new-product creative goes through a weekly compliance review.
The general lesson: copy, imagery and screening are compliance surfaces, and they deserve the same review as code.
Dearer, weaker paid social: look beyond it
Paid social's share of new accounts rose from about half in late 2025 to close to three fifths through September. It was also our lowest-converting channel, and its account-to-purchase rate fell to about three fifths of its late-2025 level, while search held about four fifths of its own.
- Search and new sources. A second social platform and affiliate or native sources, tested from July and August, reached about 2% and about 1.4% of new accounts by October. Search rose to 13% of new accounts in October, but mostly because paid social volume fell, not because search grew.
- An online shop. We learned that a shop works better as a brand and margin channel than as part of the care path when its products cannot join a patient's treatment.
- In-person retail. An in-person retail channel, live from July, which so far contributes a small share of new accounts.
Wellness opened up: new lines, launched carefully
New GLP-1 formats launched from July made up about a fifth of labelled checkouts by October; non-GLP-1 lines were still under 3%. One new line first launched only through doctors in the portal, with no landing pages or paid campaigns. With the evidence for many wellness and peptide treatments still thin, and the rules around them changing month to month, that caution seems right to me.
What the numbers say
Every figure is a ratio or an index, mostly from server-side events. October 2026 is a partial month.
| Measure | Then | Now | Reading |
|---|---|---|---|
| New account to purchase, 14 days | Index 100 (Q4 2025) | 52 in Q3, 62 in early October | Roughly halved, partly recovered |
| Checkout started to completed | 64% | 44% (Q3) | Most of the decline |
| Purchase to first payment, 65 days | 82% | 62% (Jul to Aug cohort) | A drop before the first charge, cause not yet known |
| First to second, second to third payment | 84%, 72% | 78%, 67% | Down about 5 to 6 points each |
| Payment events that are not first purchases | 79% (Oct 2025) | 88% (Aug); 94% in a partial October | The business runs on the base (October is inflated by fewer new checkouts) |
Acquisition. New-account volume peaked in May and fell through the autumn. Part of that fall is real: in late October a separate session tool also showed paid social traffic falling. But new accounts per session also fell sharply, which points to a tracking or routing change as well, so I would not read the size of the autumn drop as settled.
Conversion. About 81% of the decline was lower conversion inside each channel and about 19% was the shift toward paid social. Most of it sits at the payment step. Two onboarding changes also cost us conversion: in matched windows, a May version converted about 9% worse than the one it replaced, and a June test arm that forced a treatment-type choice converted about 12% worse, with all of the June arm's loss at payment. The rebuilt onboarding lifted treatment choice by about 11 points (browser events, directional) and, on server records, sent about 15% more new accounts to checkout (about 21% against 18%). But about five points fewer checkout starters paid (about 40% against 46%), so purchases per new account ended roughly flat.
After the purchase. First payments within 65 days fell from about 82% to about 62% of buyers, and the later renewal steps slipped about 5 points each. Patients accepted per checkout stayed flat, so it does not obviously look clinical. I do not know the cause yet, and part of it may be measurement: the flag that marks a first payment was still rolling out, and a completed checkout is not the same thing as a billed first invoice. It is the number I most want explained. These are step rates between payments for each purchase cohort; they are not the same measure, or the same period, as the 90-day retention figure in the head of product case study.
Retention. After the portal rebuild, the number of people logging in each month, relative to paying patients, rose from about 1.0 to as high as 1.3. Pages per visit rose too, but I could not see a clear change in what patients did. Medication pause, which lets a patient stop without cancelling, rose from about 13% of paying patients in its first month to about 20% from May. Cancellation events per paying patient fell from about 27 to 31% a month to about 22.5% in a partial October, and weekly cancellations fell about 30% from late August to mid-October. I am not calling that a win yet: a partial month biases this ratio down, and with fewer new patients, fewer people are in their first three months, when half of first-year cancellations happen. Self-serve cancellation requests also fell after the portal launch, but much of that was cancellations moving to the support team, which is why I count cancellation events instead.
Billing retries. In their first week, recovery within 72 hours without a card change rose from about 1.5% to about 10%, card updates within 72 hours fell from 42% to 32%, and total 72-hour recovery was flat (41.0% against 40.7%). It is early and not a causal finding, but so far retries look like they are replacing card updates rather than adding to them.
- Caveat: tags started late. In the data I analyzed, onboarding version tags on new accounts start in May 2026, and about a third of accounts still have no campaign source.
- Caveat: two copies of every key event. Browser copies ran at 78 to 107% of server events; see building click-funnel events you can trust.
- Caveat: moving splits and new markets. The rebuilt onboarding's traffic share swung week to week, and a burst of traffic from a new territory, which scrolled less than half as far as our core market, distorted one week's funnel.
What I would do differently
- Guard the payment step from day one. Checkout completion fell quarter after quarter while we optimized the steps above it. Every test should have used server-side purchases as the primary measure and checkout completion as a guardrail.
- Fix the split before the test. A moving traffic share made the first readout of the rebuilt onboarding flattering, and it took several passes to correct.
- Instrument new funnels before launch. In October, our newest products' funnels showed no progress past the questionnaire even though people were starting them.
- Make QA automation a gate. Suites that run on every merge and every release, covering every live onboarding, are worth more than a bigger test count.
- Shorten the queues. Tickets waited about 11 to 12 days in each of two ready-to-deploy states. Waiting, not testing, was the slow part.
- Change process sparingly. Every change to how work flows costs the team time to relearn it, so each one has to earn that cost.
- Keep humans in the release loop. As AI writes more of the code and launch dates press, every change still needs a human reviewer, and any planned downtime needs a checklist, a realistic window and an all-clear post.
- Alert on volume, not only on errors. An automated message that quietly stops sending raises no error, so a drop in send volume should page someone.
The takeaway
The delivery system is the part I am proudest of, because it outlasts any single feature: a team that can test, release and roll back safely can answer a market that changes every month. But shipping more did not, by itself, sell more. This year the market moved the value from the front door to the patient who had already paid, and the work that mattered most followed patients after checkout.
Two questions are still open for me: why fewer buyers reach a first payment, and how much of the autumn fall in new accounts is real. If I had to compress the year into one rule: whenever you improve a step, measure the one after it.