In July I wrote about the supply side of the peptide question, when an FDA advisory committee voted 8 to 6 to let compounding pharmacies make six unapproved peptides, over the unanimous objection of the agency's own scientists, with six of the eight recently appointed members running peptide clinics (The FDA Cleared Six Peptides for Compounding. The Evidence Didn't.). That piece had a clean shape. Evidence lost to interest, and the interest was legible on the roster.

This is the other half, and it turns out the other half got there first.

On September 7 a team at nference posted a preprint counting BPC-157 users in clinical notes across a federated network of 15.2 million patients. On September 15 Reuters covered it. The preprint is careful. The coverage is not, and the gap between them is the actual story.

Six things the July piece could not have said, because until now nobody had counted anything.

1. A denominator exists. Of 15.2 million patients with at least one clinical note between 2020 and 2026, 1,536 (0.01%) had a note mentioning BPC-157. Large language model curation confirmed administration in 1,039, ruled it out in 443, and left 54 unclear. Before this, the honest answer to "how many people are taking this" was a shrug and a screenshot.

2. The supply chain, measured, and it quietly guts the July harm reduction argument. A source was documented for 205 of the 1,039 users. Compounding pharmacies were the single largest channel at 74 (36%), just ahead of gray market peptide vendors at 71 (35%). Wellness and med spa and weight clinics accounted for 23 (11%), retail 19 (9%), online or telehealth 18 (9%). Separately, 74 patients (7.1% of the whole cohort) got it on a clinician's prescription.

Sit with that. The yes voters in July argued that reclassification would pull users out of the gray market and into regulated pharmacies. These data close at the end of the second quarter of 2026, before the vote, and the migration had already happened. The committee was not opening a door. It was papering one that people were already walking through, including by prescription, for a compound the agency had placed on its restricted list.

3. Almost nobody takes it alone. 525 of 1,039 users (50.5%) had a documented co-therapy. Testosterone in 17.0%, NSAIDs 14.7%, TB-500 12.9%, corticosteroids 11.7%, tirzepatide 6.1%, semaglutide 4.3%, ipamorelin 3.6%, CJC-1295 3.5%. Add the non-drug care that came with it, physical therapy in 15.3% and surgery in 12.6%.

Pull that thread. A man takes BPC-157 with testosterone, an NSAID and a corticosteroid, goes to physical therapy, and tells his internist his shoulder feels better. Four of those five have known effects on pain and function. The fifth is the one in the headline.

4. The user looks like a wellness customer rather than a patient population. The male share of new users rose from 57% in 2020 to 70% in 2026. Of 972 users with a recorded race, 932 (95.9%) were White, against 78.3% in the same networks' background population, a 1.23-fold enrichment. Both differences clear their stated tests at P below 0.001.

5. Clinical dark matter, quantified for the first time. Among the 80 patients whose start date could be pinned, use preceded the first clinician note by a mean of 60 days. 603 of 1,039 (58%) have exactly one note that mentions it, ever. Supply route is undocumented in 73.1%. Route of administration could not be determined for 63%. Duration of use appears anywhere in the chart for 12.5%.

6. Reuters added one number the preprint does not contain. The FDA's adverse event database holds 10 reports involving BPC-157 since 2021, all of them serious, and 8 of the 10 were filed in 2026. Ten reports against a documented user base of a thousand in one network says nothing about incidence. The concentration in one year says something about attention.

The 79% is a denominator trick, and it is the number that will travel. Of 1,039 confirmed users, 354 (34%) had any documented directional response at all. Within that 354, improved 279, no change 58, worse 17. Hence 79%. Against the full cohort the same figure reads 27% improved, 6% no change, 2% worse, and 66% not ascertainable.

So "79% improved" means that roughly a quarter of users had an improvement written down by somebody, in a record where nobody was asked systematically, nothing was measured, and no one was compared to anything. Of those 279 improvements, 263 (94%) were reported by the patient and 16 (6%) were observed on examination.

The preprint says all of this, plainly and more than once. It states that the study was not designed to quantify effectiveness. It names selection, confirmation and reporting bias. The authors behaved well. The number still got out, because 79% is a number and "not ascertainable in two thirds of the cohort" is a sentence.

The duration response curve is flat, and nobody mentioned it. Sort the documented improvement by how long patients took the stuff and you get 58% at one to thirty days, 57% at thirty-one to ninety, 62% at ninety-one to one hundred eighty, and 59% beyond that. Trend test P = 0.76. Real drugs usually leave some fingerprint in exposure duration. This one leaves a horizontal line. It is the quietest and most damning panel in the paper, and it appears in no news story about the paper.

The 33-fold rise is a documentation curve wearing a trend's clothes. New confirmed users went from 4 in the first quarter of 2020 to 134 in the first quarter of 2026. Over those six years BPC-157 went from obscure to a podcast staple with a nickname. Mentions in notes rise when use rises, when asking rises, and when clinicians start recognizing the word. The study cannot separate the three, and its own lead author said so in the same article that printed the number.

The headline welds a cabinet secretary to a curve that started without him. Reuters describes "an experimental peptide backed by U.S. Health Secretary Robert F. Kennedy Jr.," with the July vote "encouraged by Kennedy." Kennedy's Joe Rogan comments came in late February 2026. The committee voted on the 23rd and 24th of July. The study's data stop at the end of June. So the entire 33-fold rise predates the vote, and the curve was already climbing hard from 2021 onward. His tenure overlaps the steepest stretch, which is worth reporting and is a different claim from the one the headline makes.

The preprint makes a softer version of the same slip, opening its discussion with the surge in attention "in the wake of the recent recommendation by an FDA committee." Its own data cannot see past that recommendation.

The source is a preprint from a company whose product is the method. Posted to Preprints.org on September 7, not peer reviewed, carrying the publisher's disclaimer. All eight authors are nference employees. No external funding. Five of the twenty-five references are the group's own prior work, including two other preprints running the identical template on retatrutide and on incretins. nference sells large language model curation of de-identified health records to pharmaceutical clients. A paper showing that the method can characterize a population invisible to claims data is a real methodological contribution and a sales brochure at the same time, and it is allowed to be both. Reuters called it "a study."

The July story was easy to tell because the failure was one-sided. This one is harder, because the correction is running on the same grade of fuel as the thing it corrects.

A preprint from a commercial data firm, eight days old, becomes "a study finds." A percentage with a 34% denominator becomes "79% improved," which is now available to be quoted back by every clinic that sells the vial, and it will be. A cabinet secretary gets welded to a curve that had already gone vertical before he was sworn in. The one panel that genuinely cuts against the compound, that flat duration response line, appears nowhere, because no result that fails to reject a null has ever led a wire story.

The sellers will cite the 79%. The critics will cite the 33-fold. Both will be citing the same unreviewed document, and neither will mention that the document says, in its own discussion section, that it was not designed to quantify effectiveness and that its cohort is not representative of anything in particular.

There is a better use for this paper and it is the boring one. It is the first real map of a category of patient the medical record was never built to see. People buying unapproved compounds from vendors, taking them for about two months, telling one clinician once, sixty days late, and then vanishing from the chart. That population will keep growing whatever the FDA does with its bulk substances list, and the instrument that found them is the same instrument we will need for whatever the next one turns out to be.