GLP-1 and Vomiting: Why Trial Rates Disagree So Widely

Ask how common vomiting is on a GLP-1 medication and most answers give you a single percentage with no source attached. Our own corpus held 15 distinct sources that state a vomiting rate for a GLP-1 or GIP/GLP-1 drug when this article was last restated at 12:50 BST on 10 September 2026 — fourteen of them trials, the fifteenth a narrative review quoting a figure from a trial we hold no other row for (that review's own row has since been withheld as a duplicate now that we hold the trial it quoted directly — see below). That count grows as the pipeline reaches new registry records, so treat every figure below as a dated reading and the live endpoint as the current one. Read straight, they disagree with each other by almost an order of magnitude: from 6.5% to 58%. The 6.5% stands; the 58% did not survive a later re-collection of that trial (see the 13 September correction below), and as of 11:45 UTC on 30 September 2026 the live base has grown to 52 distinct sources and still runs from 0.0% to 45%. This is the seventh piece in our evidence series, after which effects have enough evidence to calibrate a model, the corpus-derived pooled rates, what patients themselves report, gallstones end to end, what is inside a personalised risk number and hair loss end to end — and, like the hair loss piece, it walks through one effect using nothing but primary trial results, each with its arm and its n.

A fixed anchor sample across the size range

The table below is not the full current source list. At 52 sources and growing, a literal row-per-source table stopped being something we could keep both complete and readable weeks ago, and every earlier promise on this page to "re-read the whole table" once registry collection "finished" assumed a ceiling that never arrived (the correction history is below). Instead, this is a fixed sample of 15 studies spanning the current size range, from the smallest posted trial (n=10) to the largest (n=8,803), re-read against the live corpus on 30 September 2026 — for the complete, currently 52-source list, read the live endpoint directly. Every row is read from each trial's own ClinicalTrials.gov results section or its PubMed abstract, not inferred; where a trial reports more than one dosing arm, the rate shown is the unweighted average across its arms and the n shown is its largest arm's (the underlying per-arm figures are on the linked record).

* Corrected 28 September 2026: this row previously cited SURPASS-2's abstract (PMID 34170647), which states vomiting as "6 to 10% and 8%" for tirzepatide and semaglutide 1 mg but gives no arm sizes. We had labelled it with the trial's whole enrolled population (1,879) instead, footnoted as a mismatch. The registry's posted results, collected 27 September 2026, give all four arms directly: tirzepatide 5, 10 and 15 mg at 5.7%, 8.3% and 9.8% (n=470, 469, 470) and semaglutide 1 mg at 8.3% (n=469). The row shows the unweighted mean of those four arms and, like the other multi-arm rows here, its largest arm's n.

That is a real 6.5%-to-45% range within this fifteen-row sample alone, not fifteen readings of the same thing — and the full 52-source base runs slightly wider still, 0.0% to 45% (a 30-participant drug-interaction study posts the true floor; see the corrections below). One row present when this table was first published is gone: the perioperative-safety review quoting SUSTAIN-3 at 7.0% with no sample size has been withheld as a duplicate now that we hold SUSTAIN-3's own registry record directly (semaglutide vs. exenatide ER, NCT01885208, 6.7% at n=405) — the same don't-double-count-a-secondary-source rule applied to a different trial's abstract in our SURPASS-2 head-to-head piece. Two proportional-reporting figures sit outside this table on purpose: FDA FAERS counts how many spontaneous adverse-event reports name each reaction term, and our rows divide a term's count by the sum of that drug's thirty most-reported term counts — a count of term-mentions, not of reports. On that denominator vomiting is 9.1% for semaglutide and 4.7% for tirzepatide, both collected from FAERS on 7 April 2026 and frozen at that date, with similar rows for the other five drugs we track. It is a real, useful ranking signal, but it answers "of the reaction terms named in reports about this drug, how many were vomiting," not "what share of patients vomit," so our eligibility rules keep it out of the incidence pool entirely rather than average it in.

Why the range is this wide

Three real differences drive it, and they are visible in the table above once you know to look.

Sample size and precision. The three smallest studies in this sample (n = 10, 10, 18) sit at 9.7%, 45.0% and 20.2% — Phase 1 pharmacology studies in a handful of participants, where one or two people vomiting swings the percentage by 10 points or more. The two largest studies (n = 6,647 and 8,803) sit at 10.5% and 6.5%. A rate from 10 people and a rate from 8,803 people are not the same kind of evidence, even though both are genuinely "a stated vomiting rate from a GLP-1 trial."

Population and endpoint. SELECT (6.5%) is titled "Semaglutide Effects on Heart Disease and Stroke in Patients with Overweight or Obesity" in its own trial record — a cardiovascular-outcomes trial, not a weight-management one, and its own name says GI tolerability was not the study's focus. STEP 1 (24.6%) and the appetite/eating-behavior study (22.3%) are, by their own stored titles, weight-management or appetite-mechanism studies. The type-2-diabetes head-to-head trials in the table — STEP 2, SURPASS-2, SURMOUNT-1 and tirzepatide vs. dulaglutide (8.0–17.4%) — sit in between.

Drug and mechanism, with a real limit. The single highest-rate row (45.0%) is an early-phase orforglipron study in 10 participants, consistent with the sample-size pattern above. But the next-highest row, STEP 1 at 24.6% (n=1,306), is a large, year-long semaglutide weight-management trial — not early-phase, not orforglipron, not small. The other orforglipron rows in this sample (20.2%, 18.7%, 10.0%, 9.7%) span nearly the same range as the semaglutide and tirzepatide rows around them. So drug and development stage explain this sample's single most extreme value; they do not explain the shape of the rest of it — and we are correcting our own earlier claim that they did (the 10 September version of this bullet said the two highest rows were both early-phase orforglipron studies, which was true of the pre-correction table and is not true of this one).

None of this makes any single number in the table wrong. It means a headline that reports "vomiting rate: 24.6%" without saying STEP 1 and weight-management dosing, or "6.5%" without saying SELECT and a cardiovascular population, is quoting a real number attached to the wrong question.

Our own pooled estimate — and why we show you its width

Pooling all 15 sources with the same weighting the corpus uses for every effect gave 10.0%, from 25 stated rates across these 15 distinct sources, when this section was first written on 10 September 2026 and until later registry batches moved it. As read from the live production endpoint at 20:44 UTC on 30 September 2026: 10.6%, from 192 eligible stated rates across 52 distinct sources — read the current figure live any time at the vomiting endpoint. Its 95% interval is now roughly 4% to 28%, against the 0–72% the 15-source reading above carried. The weighting is each source's own sample size (a source stating none counts as 1) multiplied by a discount for how confident our extraction was in reading the figure. The pooling step also winsorizes at the 5th and 95th percentile once an effect has more than ten sources — at fifteen sources those percentiles landed on the lowest and highest readings themselves, so nothing was clipped in that first reading; at 52 sources they no longer do. The lowest two readings (0.0% n=30, 4.1% n=52) are floored to 4.4% and the highest two (27.3% n=407, 45.0% n=10) are capped at 24.6% before either enters the weighted average — unchanged from the 29 September reading, since no source entered or left either tail this week.

That interval is not a bug and we are not going to round it away. Our published source-diversity label for this estimate now reads very high (it read "high" at the 15-source reading above) — but that label counts only how many distinct, independent sources stand behind the number, and says nothing about how narrow the interval is. The buckets are one source or fewer very low, up to three low, up to nine moderate, up to twenty-four high, and twenty-five or more very high, so 52 sources sits in the top band. Fifty-two sources measuring genuinely different populations at genuinely different doses produce genuine heterogeneity, and an honest interval has to say so rather than manufacture false precision. A single-population number (like SELECT's 6.5% from 8,803 people) is more useful for someone matching SELECT's population than our own pooled figure is — this is exactly why we publish the anchor sample above, not just the headline number.

What patients themselves report

Of the 42 distinct community reports in our corpus — see that article for how the pool's growth is currently constrained — 15 mention vomiting, a 36% reporting frequency (95% CI 23–51%). (Restated 25 September 2026 from 11 of 26, 42% (95% CI 26–61%) when ten GLP-1 communities were admitted to the report pool, taking it from 26 to 42.) This is not an incidence rate and we do not treat it as one: it is the share of people who chose to post about their GLP-1 experience who mentioned vomiting at all, which is a different, and structurally higher, quantity than the share of all patients who vomit. People who post are self-selected toward talking about something notable. Read alongside the trial range above, both tracks agree on the same broad picture — vomiting is common enough to be one of the most-discussed GI effects, and neither number should be read as "the" rate.

If you are on a GLP-1 now

This is educational content, not medical advice. Persistent vomiting, inability to keep fluids down, or signs of dehydration warrant contacting a clinician — GLP-1 dosing is titrated specifically to manage GI tolerability, and a rapid dose increase is a common, correctable cause of a bad reaction. Nothing here should be used to start, stop, or change a medication or its dose without your prescriber.

Every figure in this article was re-read from the live production API at 12:50 BST on 10 September 2026, row by row, when it was last restated (it was first published on 8 September 2026 with 19 stated rates from 14 sources; a retatrutide phase-2 registry record the pipeline reached since then was the fifteenth source, and the 05:17 run on 10 September added a sixteenth — the 36 mg arm of a Phase 3 orforglipron trial. The same afternoon one source was withdrawn: the tirzepatide-versus-insulin-glargine trial, SURPASS-4 (PubMed 34672967), had been carried at 7.0% with n=995, but its abstract states vomiting only as a range — "5-9%" — and 7.0% is that range's midpoint, a figure the paper never states. It was removed under the same verbatim rule that governs every other row here, taking the count back to 15 sources and moving the pooled figure from 9.9% to 10.0% and the interval from 0–71% to 0–72%), and each trial's rate comes from that trial's own posted results or abstract; the pooled percentage and its interval will move as the corpus grows. Query everything behind it — all 15 effects, both tracks, every source with its URL — at magistra.health/en/data-api, with the full eligibility methodology at magistra.health/en/methodology.

Correction, 13 September 2026 — the single highest row in the table above no longer described the live corpus, and two of the paragraphs reading it were affected. The 58.4% row ("A Study of LY3502970 in Healthy Participants", NCT03929744, n=6) was captured while our trial-registry collector still kept only each trial's five highest-percentage arms — a selection rule that ranked on the very quantity being measured, disclosed and fixed on 10 September 2026. Re-collecting that trial's arms in full put it at 10.0% with n=45, the value the table above now carries. The "sample size and precision" bullet's smallest-studies trio and the "drug and mechanism" bullet's claim that the two highest rows were early-phase orforglipron studies were both read off the pre-fix table; both have since been rewritten against current data (see above).

Reading restated 29 September 2026 (superseded by the 30 September table above): NCT06023095 read 45% and was already the single highest live figure, though not yet a formal row in this table. The 30-participant orforglipron drug-interaction study read 0.0%, not the 1.1% an earlier reading of this note had computed by adding three dosing periods together; the trial's own registry record (NCT06370728) posts three periods — orforglipron alone, 0 of 30 vomiting; carbamazepine alone, 1 of 29; orforglipron plus carbamazepine, 0 of 28 — and 0.0% is the orforglipron-alone period, the only one that measures the drug we track.

Restated 30 September 2026, in full — the standing promise above is retired, not fulfilled as originally worded. Every row in the table was re-read from the current corpus at 11:45 UTC; the withdrawn SUSTAIN-3-review duplicate was dropped; NCT06023095 (45.0%, discussed in both corrections above) is now a formal row rather than only a footnote; and the "why the range is this wide" bullets were rewritten where the current data no longer supported the original claim. What is retired is the promise itself: this table will not be "re-read once collection finishes," because registry collection has no finish line to reach — it is now maintained as a fixed anchor sample across the size range, re-verified against the corpus periodically, while the live endpoint remains the only current, complete source list.