Where supplement-subscription churn actually goes: the defection map

The headline: 9,030 exits, typed and mapped

dfmchn's defection map, built from 9,030 typed defection edges across 9.1 million public posts, finds 63% of supplement-subscription exits landing where no retention spend can follow. When dfmchn ran the same map on coffee subscriptions, only 13% of exits were unrecoverable. Supplements lose their churners at nearly five times that rate, because the destinations are different in kind: bulk raw ingredients, whole food, or out of the category entirely.

The seasonal story that is not there, and the price regime that is

Supplement subscriptions are assumed to be a resolution business: sign-ups cluster in January, so the leaving should cluster in the months right after. The chart below stacks each year's defection density on the same calendar so that claim can be tested directly. dfmchn found no January-resolution churn wave in supplement subscriptions: across 2022 to 2026, February to April defection density averages 1.1 times the rest of the year, and a step fit improves on a flat line by only 18%.

There is one price-driven regime in the data, and it is not seasonal: dfmchn measured a step-up of about 45% in supplement price-complaint density at May 2025 that stays elevated through the end of the corpus, moving independently of the calendar. The sharpest single months (December 2024, October 2025, March 2026) do not recur across years the way a seasonal wave would. They are events, and the March 2026 spike, at three times baseline, is the largest in the series. A metric that looks calendar-flat while events of that size land inside it is the signature our field note why your dashboard turns red too late is about.

data table · defection density by calendar month

defection edges per 1,000 posts, by calendar month and year

month	2022	2023	2024	2025	2026
Jan	0.85	0.72	0.76	1.21	1.20
Feb	0.74	0.71	0.74	1.15	1.90
Mar	0.86	0.88	0.76	1.27	3.16
Apr	0.87	0.94	0.83	0.96	1.06
May	0.93	0.55	0.83	1.17	1.11
Jun	0.85	0.62	0.74	0.89	1.26
Jul	0.72	0.70	0.69	1.03	1.50
Aug	0.80	0.92	0.92	0.99
Sep	0.94	0.63	1.46	1.03
Oct	0.95	0.69	1.04	2.33
Nov	1.00	0.89	1.05	1.15
Dec	0.80	0.75	2.18	1.87

“At this point, Huel is not a viable solution to that anymore for me. I used to order every couple of months because I knew what was available and what I needed. Now, it’s maybe once a year.”

The friction map

Before asking who leaves, we let the corpus say what people complain about at all. Clustering the whole category's conversation, never a single brand's (the reasoning is our field note you cannot cluster a brand, only a category), produced 103 clusters, of which 75 survived relevance review. The schema below is what the category itself considers worth arguing about; every brand is scored against this one map.

data table · friction schema by tier

category friction schema by tier

tier	clusters	docs
Taste / tolerability	22	12,279
Subscription mechanics	11	4,752
Efficacy skepticism	13	4,718
Advocacy (positive)	7	3,621
Price / value	9	3,200
Trust / marketing fatigue	5	1,353
Displacement	2	677
Habit fatigue	3	556
Other	3	424

“as soon as I stopped Huel and switched to 4-5 eggs in the AM, my body fat and total weight both went down. Sorry to say, I’m done for good. With this post about a shady owner, really glad I’m done.”

Where defectors go

Every exit sentence names or implies a destination, and each defection is parsed as a full sentence: source, destination, trigger (the structure of our field note churn has a grammar). In this category the exits skew away from rotation: people who leave a supplement subscription mostly do not shop for another one. They buy the raw ingredients in bulk, replace the product with actual food, pick up a cheaper one-off at a store, or stop entirely.

data table · defection destinations by scope

defection destinations by scope

	Another subscription	Bulk raw ingredients	Whole food instead	Grocery or retail	Quit the category	Other
Huel	41%	2%	6%	4%	43%	3%
Unbranded / category level	16%	5%	9%	8%	59%	3%
Other named subscriptions	41%	4%	2%	13%	37%	3%
AG1	48%	7%	11%	4%	28%	2%
Ka'Chava	46%	2%	6%	6%	37%	3%
Seed	33%	0%	10%	3%	50%	3%
Transparent Labs	58%	15%	5%	0%	17%	5%
Legion	49%	2%	18%	17%	15%	0%

“I stopped my subscription. The bars are trash and the new sizing / cost per calorie changes just sealed the deal. I initially looked at this as a meal replacement to simplify my life but I don't need shrinkflation and lead in the mix too.”

Brand fingerprints

Each brand's mentions are scored against the category schema as a share of that brand's own conversation, so bigger brands do not look worse by volume (the unit-of-measure argument is our field note density, not presence). Mentions that fit no category cluster are kept aside as residuals and read by hand; they are usually the most brand-specific signal. Cluster names are induced from the category conversation, so one brand's fingerprint can include a cluster named for a competitor: comparison discourse crosses brand lines.

“I took AG on for three months and went to the doctor for my general physical and blood work and my liver enzymes are through the roof. My doctor told me to stop drinking this shit immediately!”

One brand that died, and one too quiet to read

Care/of shut down inside this corpus's window, and the corpus saw it happen: its closure lands cleanly at June 2024, where its users describe the shutdown email in real time. Persona Nutrition is a different lesson. It is still operating (it returned to its founders in 2024 and now sells mostly business to business), but you could not learn that from this corpus: it has only 4 mentions in four and a half years, they thin out during 2023, and the last one asks whether the service is worth trying. A brand can be too small on public forums for its state to register either way; that silence is a resolution limit of the instrument, and it is reported here as one.

Because the timeline is preserved, Care/of's defection profile can be read in the period before shutdown and compared with the surviving brands on the same map. The highlighted row is the pre-shutdown window of a dead brand, and an honest caveat comes with it: three edges from 21 documents is an anecdote, not a fingerprint. The row is shown for the timeline, not as a measurement.

“I disputed the charges and got a new credit card. There are several posts on here from people who were still charged even though the card they have now is not the same as the one they used for their subscription.”

Fragility summary: what triggers the exit, and what stays winnable

The ratio that matters for retention teams is threatening vs already gone: a still-threatening post is a customer who can still be saved.

data table · exit triggers

exit triggers: completed vs still threatening

trigger	completed	threatening
Taste / tolerability	3,026	710
Trust / marketing	776	366
Price / value	782	341
Efficacy	841	191
Other	664	323
Mechanics	542	265
Habit fatigue	143	60

Not all churn is equal. An exit to another subscription is a rotation, recoverable with a better offer. An exit to bulk ingredients, whole food, a grocery shelf, or out of the category entirely is structural: the budget line is gone and no retention spend follows it there. Across the category, 63% of defection edges are the second kind. Huel alone contributes more than half of all edges; excluding it, the pooled unrecoverable share of the remaining scopes is roughly 73%, so the pattern is not an artifact of the corpus's largest brand.

data table · recoverable vs gone-for-good by scope

recoverable vs gone-for-good by scope

scope	edges	unrecoverable	rotation
Unbranded / category level	2,690	82%	16%
Seed	34	65%	32%
Huel	5,103	56%	41%
Other named subscriptions	719	55%	42%
Legion	24	54%	46%
Ka'Chava	71	46%	51%
AG1	317	45%	54%
Transparent Labs	28	36%	61%

Friction also splits by durability. Price and habit-fatigue complaints decay; mechanics, trust, efficacy, and tolerability complaints accumulate. The structural share below is each scope's fraction of defection triggered by the durable kind. Edges whose trigger fell in the Other tier are counted in neither column, so the columns do not sum to the edge totals.

data table · structural vs perishable friction by scope

structural vs perishable friction by scope

scope	edges	structural	perishable	structural share
Whole category	9,030	6,717	1,326	74%
Huel	5,103	3,679	772	72%
Unbranded / category level	2,690	2,216	254	82%
Other named subscriptions	719	537	113	75%
AG1	317	162	126	51%
Ka'Chava	71	43	24	61%
Seed	34	27	6	79%
Transparent Labs	28	14	11	50%
Legion	24	12	6	50%

“Finally cancelled today after bombarding them with emails with no reply, after threatening to go through my bank with a charge back they conveniently responded and finally cancelled… awful company, avoid at all costs.”

Method note

The analysis used 9.1 million cleaned public posts and comments (January 2022 through July 2026) and no brand internal data.

• Corpus: 9.1M cleaned public posts and comments from twelve fitness, nutrition, and subscription communities. 79,718 documents passed the relevance gate.
• Huel weighting: two of the twelve communities are effectively Huel's home venues, and Huel accounts for 57% of both documents and defection edges, so category-level aggregates lean on Huel's mass. The ex-Huel robustness figure in the fragility section is reported for that reason; brand fingerprints are unaffected because each brand is scored against its own corpus.
• Exclusions: multi-level-marketing supplement discourse (954 rows) was excluded before analysis. The distributor relationship is a different structure and its frictions would contaminate the category schema.
• Schema: a 103-cluster category schema was induced on the whole corpus at once, never per brand; 75 clusters survived relevance labeling.
• Edges: 9,030 defection-bearing documents were parsed into 9,030 typed edges (source → destination → trigger, completed vs threatened, reversible vs final) by a constrained-output LLM pass with a calibration prompt.
• Density, not presence: brand shares are population frequencies of each brand's own corpus, so bigger brands don't look worse by volume.
• Decay: complaints decay by trigger type (price: ~10.5-month half-life; habit fatigue: ~15; mechanics, trust, efficacy, and tolerability: none; structural friction doesn't age out).
• Dead brands: Care/of is scored on its pre-shutdown window, anchored at its shutdown month, so its fingerprint stays comparable to the survivors'.
• Mention floor: brands above the ~150 usable-mention floor carry brand-level claims: Huel, AG1, Ka'Chava, Transparent Labs, Seed, Legion. Scored against the schema but below the floor, so excluded from brand-level claims: Ritual, Momentous, Bloom, Care/of (before shutdown), Gainful, Persona Nutrition.
• Quotes: verbatim and unedited; receipts are chosen only from those that do not identify the source platform.

Related dfmchn field notes: the half-life of a complaint · the cheapest customer to save · your churn model predicts the past.