GLP-1 Side Effect Rates: What Our Database's Own Clinical Evidence Shows
14 of the 15 GLP-1 side effects we track have a published rate pooled directly from our own corpus of clinical trial and regulatory sources — not copied from a single paper. Here is every number, with its confidence interval, exactly how many sources it rests on, and the corrections that shaped the eligible base.
Ask "how common is nausea on semaglutide?" and most answers online repeat one number from one trial's package insert. Magistra's real-world evidence database instead pools every clinical trial and regulatory rate we've found that's independently sourced and citable, weights it by sample size and evidence quality, and computes a rate with a proper confidence interval — the same computation our predictor tool runs live. As of 10 September 2026, 14 of the 15 GLP-1 side effects we track clear the bar for this corpus-derived estimate (10 as of 31 August 2026; hair loss and dizziness joined the table on 7 September 2026, fatigue on 8 September 2026, gallstones and pancreatitis on 10 September 2026 — all five are in the table below). The remaining one, emotional blunting, publishes no clinical figure at all: its literature figure was itself uncited and was withdrawn on 3 September 2026, and no citable rate has replaced it. No effect now rests on the static literature-fallback table for a published clinical percentage. (Corrected 8 September 2026: this sentence still counted fatigue as a fallback after it had moved into the corpus-derived table earlier the same day — 13 corpus-derived plus 3 fallbacks is 16 of 15 effects.)
Correction, 31 August 2026. This article originally (as of 22 August 2026) published 13 corpus-derived rates. On 31 August we found that the April-2026 pass that seeded our initial database had written rate rows restating our own static literature table under real trial and study URLs (NEJM, The Lancet and others) that no extraction ever read — rows a URL-based eligibility screen cannot catch. All 150 seed rows are now labelled and excluded from the eligible base, which fell by roughly half (details in our methodology's correction log). Three effects previously shown here with a "very low" confidence rate — dizziness, hair loss and pancreatitis — turned out to rest entirely on those seed rows and have moved to the literature-fallback table below; the table's figures are restated over the cleaned base. Earlier versions of this table should not be cited.
This is the clinical track only — trial registries and regulatory sources, never patient self-reports. We track how often patients themselves mention each effect separately, as a reporting frequency, and never average the two together (a mention rate and an incidence rate measure different things). This article is about the clinical number only; the community track's own numbers are here, with definitions in our methodology.
The 14 corpus-derived rates (all rows re-read at 06:44 UTC on 7 October 2026)
| Effect | Pooled rate | 95% CI | Rates / sources | Source diversity | Live |
|---|---|---|---|---|---|
| Nausea | 20.5% | 6–50% | 246 / 83 | Very high | API |
| Diarrhoea | 14.7% | 5–35% | 233 / 73 | Very high | API |
| Headache | 8.4% | 3–19% | 198 / 59 | Very high | API |
| Constipation | 11.7% | 3–36% | 216 / 63 | Very high | API |
| Reduced appetite | 8.6% | 3–23% | 213 / 62 | Very high | API |
| Vomiting | 10.4% | 4–27% | 232 / 73 | Very high | API |
| Abdominal pain | 5.8% | 2–18% | 168 / 40 | Very high | API |
| Fatigue | 6.5% | 2–18% | 108 / 32 | Very high | API |
| Dizziness | 6.7% | 3–14% | 136 / 40 | Very high | API |
| Hair loss ‡ | 5.0% | 2–11% | 22 / 11 | High | API |
| Acid reflux | 5.1% | 3–10% | 83 / 27 | Very high | API |
| Injection site reaction | 4.4% | 1–15% | 19 / 8 | Moderate | API |
| Gallstones † | 1.7% | 1–4% | 36 / 13 | High | API |
| Pancreatitis † | 0.1% | 0–1% | 13 / 4 | Moderate | API |
Restated at 06:44 UTC on 7 October 2026, 14 of 14 rows — the first capture in just over three days, against this section's own 24-hour aim. Three registry runs happened since the last capture (01:17 UTC, 4 October): the 5 October run added seven new trials, all reporting results posted directly to their ClinicalTrials.gov record — a lixisenatide-before-meals T2D study (2 effects), a 24-week lixisenatide T2D study (5 effects), a long-term tirzepatide (LY3298176) T2D safety study (36 rows, 12 effects — the largest single addition), a dulaglutide (LY2189265) T2D study (3 effects), SCALE – Maintenance (liraglutide, obesity, 9 effects), a semaglutide/knee-osteoarthritis obesity study (5 effects), and a liraglutide-vs-lixisenatide T2D study (4 effects) — 73 rows total, all seven re-verified against the trials' own posted adverse-event modules the same day. The 6 October run added one more new trial, STEP 6 (semaglutide, obesity, 16 rows across 8 effects), plus 2 more headache rows from the liraglutide-vs-lixisenatide study above. The 7 October run added 68 rows, all a reproductive/thermoregulatory effect this table does not yet track (temperature_dysregulation), so it moved none of the 14 rows here. No rule or formula changed. The site-wide eligible base moved from 1,832 rates across 78 studies to 1,923 across 86; every row above moved by 0.3 points or less, and no row crossed a source-diversity tier.
Restated at 01:17 UTC on 4 October 2026, 14 of 14 rows. No collection run took place on 3 October. The 05:17 IST run on 4 October continued the second registry batch. Six trials entered the pool, and one added a row:
All 98 registry rates the run stored match the trials' own posted arm counts, re-read against the registry before this note was written. The site-wide base moved from 1,734 rates across 72 studies to 1,832 across 78. No rule or formula changed. Every rate moved by 0.1 points or less. Injection site reaction's interval narrowed from 1–18% to 1–16% as its sources went from six trials to seven; headache's widened from 4–19% to 3–20%. Two rows crossed a source-diversity threshold: acid reflux (high → very high) and pancreatitis (low → moderate).
Restated at 05:44 UTC on 2 October 2026, 12 of 14 rows. The 05:17 BST collection run that morning continued the second registry batch. Five trials entered the pool, and one added rows:
All 75 registry rates the run stored match the trials' own posted arm counts, re-read against the registry before this note was written. The site-wide base moved from 1,659 rates across 67 studies to 1,734 across 72. No rule or formula changed. Injection site reaction moved most, from 5.1% to 4.5%, and its interval widened from 2–16% to 1–18% as its sources went from five trials to six. Fatigue moved from 6.9% to 6.6%. Every other rate moved by 0.2 points or less. One row crossed a source-diversity threshold: gallstones (moderate → high). Hair loss and pancreatitis are unchanged.
Restated at 05:50 UTC on 1 October 2026, 13 of 14 rows. The 05:17 BST collection run that morning continued the second registry batch. Six trials entered the pool, and one added a row:
The same run also stored 15 rates from NCT06771115, a trial of VTX3232, which is not a GLP-1 drug, given alone or added to semaglutide. They reached the live API at about 05:43 UTC, and we withheld all 15 about five minutes later, before this table was re-read: neither arm measures a GLP-1 drug on its own. Our collector now skips both arms. All 89 registry rates the run stored match the trials' own posted arm counts, the 15 withheld ones included. The site-wide base moved from 1,585 rates across 61 studies to 1,659 across 67. No rule or formula changed. Injection site reaction moved most, from 4.4% to 5.1%, and its interval widened from 3–6% to 2–16% as its sources went from three trials to five. Hair loss moved from 5.5% to 5.1%. Every other rate moved by 0.2 points or less. Three rows crossed a source-diversity threshold: fatigue (high → very high), hair loss (moderate → high) and injection site reaction (low → moderate). Pancreatitis is unchanged.
Restated at 05:53 UTC on 30 September 2026, 13 of 14 rows. The 05:17 BST collection run that morning continued the second registry batch. Our pre-publish gate held it at first, because injection site reaction's row count rose by more than half (5 → 9). Every new registry rate was then re-read against the trial's own record and the batch was released at 05:52 UTC. Four trials entered the pool, a fifth added rows, and one review entered:
All 76 registry rates match the registry's own posted arm counts. The site-wide base moved from 1,508 rates across 56 studies to 1,585 across 61. No rule or formula changed. No rate moved by more than 0.3 points: nausea (20.3% → 20.6%) and fatigue (7.4% → 7.1%) moved most. No row crossed a source-diversity threshold. Pancreatitis is unchanged.
Restated at 05:46 UTC on 29 September 2026, 9 of 14 rows. The 05:17 BST collection run that morning continued the second registry batch. Six trials entered the pool, and a seventh completed:
Every one of those 74 rates was re-read against the registry's own posted arm counts before this note was written, and all 74 match. The site-wide base moved from 1,434 rates across 50 studies to 1,508 across 56. No rule or formula changed. No rate moved by more than 0.1 points. The widest interval narrowed the most: reduced appetite's went from 1–48% to 2–34%. No row crossed a source-diversity threshold. Hair loss, acid reflux, injection site reaction, gallstones and pancreatitis are unchanged.
Restated at 23:45 UTC on 28 September 2026, 2 of 14 rows. No new trial entered the pool, and one trial that had been counted twice now counts once. SURPASS-2 had been counted as two distinct studies for diarrhoea and vomiting: once through its registry record, collected 27 September, and once through its own journal abstract (PMID 34170647), whose semaglutide figures ("12%" and "8%") are the registry's semaglutide 1 mg arm (54/469 = 11.5% and 39/469 = 8.3%) rounded to whole percents. Both abstract rows are now withheld as secondary-source duplicates, the rule applied to a review restating SUSTAIN-3 on 19 September. Diarrhoea moved from 171 rates across 43 sources to 170 across 42, and vomiting from 172 across 44 to 171 across 43; neither pooled rate moved, and vomiting's lower bound fell from 4% to 3%. The site-wide base moved from 1,436 rates across 51 studies to 1,434 across 50. The other 12 rows were re-read at the same time and are unchanged.
Restated at 05:54 UTC on 28 September 2026, 9 of 14 rows. The 05:17 BST collection run that morning continued the second registry batch. Seven trials entered the pool: the two headache arms of ADJUNCT ONE that the previous run's row cap had cut (2 rates, liraglutide 1.2 and 1.8 mg), SUSTAIN 7 (24: six effects across semaglutide 0.5 and 1.0 mg and dulaglutide 0.75 and 1.5 mg), SUSTAIN FORTE (8: four effects across semaglutide 1.0 and 2.0 mg), NCT01029886 (10: five effects across exenatide once weekly and liraglutide once daily; the registry states no adverse-event counting window for this 2011 trial, so those rows carry none), GETGOAL-S (6: six effects from its one lixisenatide arm), SCALE Diabetes (18: nine effects across liraglutide 3.0 and 1.8 mg) and ADJUNCT TWO (6: constipation and diarrhoea across three liraglutide doses, again as an adjunct to insulin in type 1 diabetes; its remaining 15 tracked rows were cut by the run's row cap and arrive with the next run). Every one of those 74 rates was re-read against the registry's own posted arm counts before this note was written, and all 74 match. The site-wide base moved from 1,362 rates across 45 studies to 1,436 across 51. No rule or formula changed. No row moved by more than 0.2 points: constipation 12.0% → 11.8% and diarrhoea 15.2% → 15.0% moved the most; nausea, reduced appetite, vomiting, fatigue and dizziness moved by 0.1; headache and abdominal pain kept their rates and gained sources. One row crossed a source-diversity threshold: dizziness moved High → Very high, its distinct sources having risen from 24 to 26. Hair loss, acid reflux, injection site reaction, gallstones and pancreatitis are unchanged. Thirty-four of the 44 trials are still queued in full, so this table will move again with roughly the next week of daily runs, and it will be restated the same way each time.
Restated at 05:50 UTC on 27 September 2026, 8 of 14 rows. The 05:17 BST collection run that morning began the second registry batch: 44 pivotal trials with posted results that we had never collected, pinned on 27 September and drained at 80 arm rows a run. Its first four entered the pool — SURPASS-2 (24 rates: six effects across three tirzepatide doses and semaglutide 1 mg), AWARD-11 (9: three effects across three dulaglutide doses), PIONEER PLUS (21: seven effects across three oral semaglutide doses) and ADJUNCT ONE (22: liraglutide as an adjunct to insulin in type 1 diabetes, three doses; its two remaining headache arms were cut by the run's row cap and arrive with the next run). Every one of those 76 rates was re-read against the registry's own posted arm counts before this note was written, and all 76 match. The site-wide base moved from 1,286 rates across 41 studies to 1,362 across 45. No rule or formula changed. No row moved by more than 0.3 points: constipation 12.3% → 12.0% and abdominal pain 6.6% → 6.3% moved the most; diarrhoea, headache, reduced appetite and fatigue moved by 0.1; nausea and vomiting kept their rates and narrowed their intervals. One row crossed a source-diversity threshold: abdominal pain moved High → Very high, its distinct sources having risen from 23 to 26. Dizziness, hair loss, acid reflux, injection site reaction, gallstones and pancreatitis are unchanged. Forty of the 44 trials are still queued, so this table will move again with each of roughly the next eight daily runs, and it will be restated the same way each time.
Restated at 17:44 UTC on 26 September 2026, 10 of 14 rows. The collection run that day finished the batch the 25 September restatement said was coming. It added 46 arm-level rates from the rest of SURMOUNT-2 and four cardiovascular-outcome and weight-management trials we had never collected: SCALE Obesity and Prediabetes, REWIND, LEADER and ELIXA. The site-wide base moved from 1,240 rates across 37 studies to 1,286 across 41. No rule or formula changed; this is new evidence entering the pool. Diarrhoea moved the most, 17.9% → 15.3%. Nausea (21.4% → 20.3%) and headache (9.6% → 8.6%) each moved by about a point, and no other row moved by more than 0.5 points. Hair loss, injection site reaction, gallstones and pancreatitis are unchanged. One row crossed a source-diversity threshold: headache moved High → Very high. That batch is now complete; the next registry batch will move this table again, and it will be restated the same way.
Corrected at 12:15 UTC on 25 September 2026, 1 of 14 rows. Injection site reaction read 4.5% (0–38%) from 6 rates and 3 sources. It now reads 4.4% (3–6%) from 5 rates and 2 sources. One of its three sources, NCT05254418 (a 7-person dulaglutide pilot in chronic kidney disease), never posted an injection-site-reaction row. It posts "local injection site erythema", 1 of 7 participants, which is a separate MedDRA preferred term. Our extraction model filed it under this effect on 20 August 2026, before our collector restricted each effect to its own registry term. Every other registry row behind this effect reads "Injection site reaction". The rate is now withheld as a term mismatch (the row stays in the corpus for audit), under the same rule that took a dyspepsia row out of acid reflux on 17 September. We found it by re-reading all 1,233 rated registry rows against ClinicalTrials.gov's live records the same day. Every other row matched its registry cell exactly. The site-wide base moves from 1,241 rates to 1,240, still across 37 studies, because the trial's nausea row is unaffected. No other row in this table changed. The percentage barely moves because sample-size weighting had already discounted the 7-person arm almost to nothing, but the interval narrows sharply: it had been opened by a disagreement between two different measurements, not two readings of one.
Restated at 05:52 UTC on 25 September 2026, 11 of 14 rows. The 05:17 BST collection run that morning read five pivotal trials whose posted results we had never collected: SUSTAIN 6, SOUL, STEP UP, ATTAIN-1 and SURMOUNT-2. That added 75 arm-level rates and took the site-wide base from 1,166 rates across 32 studies to 1,241 across 37. No rule or formula changed; this is new evidence entering the pool. Nausea moved the most, 23.4% → 21.4%, and no other row moved by more than 0.7 points. Gallstones, pancreatitis and injection site reaction are unchanged. No row crossed a source-diversity threshold. The same batch continues on the next run with LEADER, REWIND, ELIXA, SCALE and the rest of SURMOUNT-2, so expect this table to move again.
Restated at 12:48 UTC on 24 September 2026, 12 of 14 rows. The source counts trace mostly to the same-day v5.35 fix described in our methodology: `buildRateBase` had been collapsing two different trials into one source whenever their titles truncated to the same text (SURMOUNT-1 and SURMOUNT-5 both read "A Study of Tirzepatide (LY3298176) in Participants With Obes..." once cut for storage), so SURMOUNT-5's rows were being averaged into SURMOUNT-1's entry instead of counting as their own study. The dedupe key now includes the study URL, not just the display name. Eleven effects (nausea, diarrhoea, headache, constipation, reduced appetite, vomiting, abdominal pain, dizziness, hair loss, acid reflux and injection site reaction) had both a SURMOUNT-1 and a SURMOUNT-5 row and gained exactly the distinct source the fix un-hid; the twelfth (fatigue) has no SURMOUNT-1 entry to collide with — its SURMOUNT-5 row is a genuinely new source, stored when SURMOUNT-5 entered the corpus on 23 September, not an unmerge. The two changes show in different columns: each of the 12 rows carries two more stated rates than on 19 September (SURMOUNT-5's two arms, stored on 23 September), and the fix changed the source count, not the rate count. No effect's pooled rate moved by more than half a point except hair loss. Intervals moved both ways: five narrowed (headache, vomiting, abdominal pain, fatigue, injection site reaction), four held, and three widened by one point at the top (nausea, hair loss, acid reflux). Four rows crossed a source-diversity threshold, all upward: diarrhoea, constipation and reduced appetite moved High → Very high, and hair loss moved Low → Moderate — see the ‡ footnote for what changed in its evidence base.
‡ Read this row's interval against the ones above it, not on its own. Until 14 September 2026 hair loss rested on one trial (3 rates, the three tirzepatide arms of SURMOUNT-1), so its 4–7% was a single trial's sampling interval with the between-study term 0 by construction — narrower than diarrhoea's because it rested on less, not on more — and the API labelled it `intervalBasis: "sampling_only_single_source"` (added 11 September 2026). That morning a second study entered its base (NCT05548647, a semaglutide trial posting alopecia at 4.9% in 72 participants), so the row now reads 5 rates from 2 studies and the API reports `sampling_plus_between_study` for it like every other row. It is still the thinnest row in the table: two studies is not diversity, and its "low" label is the thing to read first. On 15 September 2026 a third study entered its base: NCT05971940, an orforglipron trial posting alopecia in three arms — 0 of 143 (0%), 1 of 137 (0.7%) and 2 of 141 (1.4%) — so the row reads 8 rates from 3 studies, pooled 4.5%, and its interval widened from 4–7% to 2–12% as the between-study term absorbed a source well below the first two. Still "low". On 24 September 2026 a same-day fix (v5.35) stopped SURMOUNT-5 (NCT05822830, tirzepatide vs semaglutide) being pooled into SURMOUNT-1's entry under a truncated shared title — it posts alopecia at 31 of 374 (8.3%) tirzepatide and 23 of 376 (6.1%) semaglutide, shown in our per-source list as the arm average, 7.2%, beside the larger arm's n. The row then read 10 rates from 4 studies, pooled 5.4%, interval 2–13%, and the label moved to "Moderate" — the first time this row had left "low". On 25 September 2026 two more studies entered: STEP UP posts alopecia at 72 of 1,004 (7.2%) on semaglutide 7.2 mg and 8 of 201 (4.0%) on 2.4 mg, and ATTAIN-1 at 30 of 723 (4.1%), 36 of 724 (5.0%) and 39 of 728 (5.4%) on orforglipron 6, 12 and 36 mg. The row now reads 15 rates from 6 studies, pooled 5.3%, interval 3–11%, still "Moderate".
† Rows whose evidence is partly or wholly serious-adverse-event rates from trial registries' SAE tables — a floor on all-cause incidence, not incidence itself. Pancreatitis: 2 of its 4 studies; gallstones: 3 of its 13, as of 7 October 2026 (3 of its 12 from 4 to 7 October). (Corrected 2 October 2026: this note said "wholly or mostly" and "gallstones: 3 of its 7 studies" after the gallstones row above had grown to 9 and then 10 studies. Three of ten is a minority, so the old wording overstated how much of that row is an SAE floor.) See below.
Two rows joined on 10 September 2026, and they are not ordinary rates. ClinicalTrials.gov posts a serious-adverse-event table and a separate "other (not including serious)" table that only lists terms above the trial's reporting threshold, usually 5%. Acute pancreatitis and gallbladder events are adjudicated serious and sit far below that threshold, so for those effects the SAE table is the only place the registry states a number at all. Those rows had been withheld since 7 September, after a fetch briefly pooled every trial's SAE table as incidence and sent constipation from 28% to 5% overnight. They are now re-admitted for these two effects only — never for a term that also has a non-serious row in the same arm, which would be a subset of a rate rather than a rate — and every surface that publishes them says what they are. On the API, rateBase.clinical.sources[].rateKind marks each source "incidence" or "serious_ae", and the pooled figure carries the same disclosure. Read pancreatitis's 0.1% as at least 0.1%, and compare it to the FDA Wegovy label's own 0.2 acute-pancreatitis cases per 100 patient-years — a different unit, on a different denominator, which is exactly why we print the kind of rate beside the number.
Restated again at 12:50 BST on 10 September 2026, four rows. Nausea, diarrhoea, reduced appetite and vomiting each lost one source that afternoon: the SURPASS-4 trial (tirzepatide versus insulin glargine, PubMed 34672967) had contributed 17.5%, 17.5%, 10% and 7% at n=995 to those four effects, and its abstract states each of those quantities only as a range ("nausea (12-23%), diarrhoea (13-22%), decreased appetite (9-11%), and vomiting (5-9%)") — every stored figure was the midpoint of its range, not a number the paper states. The rows were withheld under the verbatim rule described in our methodology; no pooled rate moved by more than 0.2 points.
Restated 10 September 2026 — every row. The 9 September restatement (which itself replaced a 31 August snapshot that had gone nine days stale, with three effects moved by more than ten percentage points — constipation 29.0% to 11.8%, acid reflux 33.3% to 5.3%, reduced appetite 21.1% to 9.8%) is re-run here as a single same-moment capture of all fourteen rows, read from the production API on 10 September 2026, because gallstones and pancreatitis joined the table that morning. A dated snapshot is honest, but a dated snapshot nobody re-runs stops describing the thing it points at; the rule this table now follows is that it is re-captured whenever any row's membership changes. Previous values should not be cited. The two biggest movers have documented causes, both in our methodology's correction log: the pinned fetch of pivotal trials' posted results (7–9 September) added dozens of registry arm rates across nearly every effect, and the same fetch briefly pooled the registry's serious-adverse-event table as incidence before those rows were withheld. Constipation's own journey through that week is walked through end to end in the constipation article.
Restated 15:45 BST on 13 September 2026 — every row, and the rule above now has something that enforces it. Twelve of the fourteen rows had moved in the three days since the last capture, because our re-collection of the trial registry has been running throughout (the corpus went from 1,573 to 2,317 points in a week). Most point estimates barely shifted — the largest were abdominal pain 8.3% to 6.6% and fatigue 7.5% to 6.3% — but the evidence behind them changed a great deal: nausea went from 45 stated rates to 130, abdominal pain from 11 to 110, constipation from 26 to 100, and three source-diversity labels moved up a grade (headache and abdominal pain moderate to high, fatigue low to moderate). Two intervals had been understating our uncertainty, which is the direction that matters: abdominal pain published 3–21% against a live 1–44%, and vomiting 0–72% against 2–44%. The honest reading of that is not that the old numbers were fabricated — each was a true capture of its moment — but that this table promised to be re-captured whenever a row's membership changed and nothing was actually checking. As of today something is: a guard runs with our daily job, compares every row here against the live API, and fails when they disagree, so the promise in the paragraph above is now enforced by a machine rather than by a good intention. Previous values should not be cited.
Restated 14:05 BST on 14 September 2026 — every row, on the guard's first morning. The guard ran with our daily job for the first time and reported ten of the fourteen rows drifted overnight — the expected case, because the trial-registry re-collection is still running (600 of 942 tracked arm rows were stored when that run started; roughly two more runs to go). No point estimate moved by more than 0.3 points (abdominal pain 6.6% to 6.9%, headache 11.2% to 11.0%, fatigue 6.3% to 6.5%), but the evidence behind them kept growing — diarrhoea 62 stated rates to 81, constipation 100 to 120, headache 88 to 104, vomiting 67 to 87, acid reflux 25 to 46 — and three source-diversity labels moved up a grade (dizziness and acid reflux moderate to high; hair loss very low to low, because a second study entered its base, see ‡ below). Vomiting's interval narrowed from 2–44% to 3–30% and fatigue's widened from 3–14% to 2–22%: both are the evidence moving, not a method change. Previous values should not be cited.
Restated 10:39 BST on 15 September 2026 — every row. The guard reported ten of the fourteen rows drifted overnight, again the expected case: the 15 September run stored 120 more registry arm rows. No point estimate moved by more than 1.1 points (headache 11.0% to 9.9%, hair loss 5.3% to 4.5%, dizziness 6.5% to 6.3%); the evidence kept growing (diarrhoea 81 stated rates to 122, reduced appetite 99 to 123, headache 104 to 118, vomiting 87 to 102, dizziness 45 to 59); two source-diversity labels moved up a grade (vomiting high to very high, fatigue moderate to high); and two wide intervals narrowed as more studies entered (reduced appetite 0–83% to 1–61%, diarrhoea 3–58% to 5–42%) while hair loss's widened from 4–7% to 2–12% — the honest direction for a row whose third study posts rates well below its first two (see ‡). Site-wide, the eligible base went from 887 to 1,007 rates, still 29 distinct studies.
Restated 12:59 BST on 16 September 2026 — every row. The guard reported eight of the fourteen rows drifted, this time at midday rather than overnight: the 16 September run stored 126 more registry arm rows at 05:17, but its sync to production failed in a network outage at 08:52, so the rows reached the live corpus only at 12:53, when we re-ran the sync by hand. All 126 come from two orforglipron trials — ACHIEVE-3 (NCT06045221, four arms, two of them oral-semaglutide comparators) and a Phase 1 tablet-versus-capsule study in adults with obesity (NCT06440980), whose 38 posted dose-level arms are why fatigue's entry count jumps from 27 to 65 and dizziness's from 59 to 97 while each still counts as one study. No point estimate moved by more than 0.4 points (fatigue 6.5% to 6.1%, abdominal pain 6.9% to 6.6%, vomiting 9.8% to 9.9%); vomiting's interval narrowed from 3–31% to 3–29% and acid reflux's from 0–59% to 0–56%. Every drifted row gained exactly one study.
Restated 15:58 BST on 16 September 2026 — eight rows, one row of one study withheld. The guard reported eight rows drifted three hours after the midday capture, and this time the corpus shrank rather than grew: reading every arm of the headache base before quoting it in a new article found that a drug-interaction study (NCT06370728) doses its participants with carbamazepine alone for several weeks, and that our collector had stored that period — an anticonvulsant, no GLP-1 on board — as an eligible rate for eight effects. Those eight rows were withheld from both stores at 15:50 BST; the study itself stays in each base on its orforglipron periods. Every one of the eight rows lost exactly one stated rate (nausea 134 → 133, headache 118 → 117, constipation 126 → 125, reduced appetite 123 → 122, vomiting 132 → 131, abdominal pain 117 → 116, fatigue 65 → 64, dizziness 97 → 96), no study count and no point estimate moved, and two intervals widened by a point (headache 4–21% to 4–22%, fatigue 1–23% to 1–24%). The same afternoon the trial-registry re-collection described in the entries above was confirmed complete against the live registry, so this is the first capture of the table whose per-study figures are all means of every posted arm.
Restated 21:50 BST on 16 September 2026 — one row, one rate of one study withheld. The guard reported a single row drifted six hours after the previous capture, and again the corpus shrank rather than grew: reading every source of the abdominal pain base before quoting it in a new article found that the one PubMed abstract in that base (PMID 37062422, a single-ascending-dose Phase 1 study of TG103, an investigational GLP-1/Fc fusion protein, in 24 dosed healthy volunteers) states 12.5% for abdominal distension and no abdominal-pain figure at all; our extractor had stored the distension figure as an abdominal pain rate. Abdominal distension is a separate adverse-event term — the FDA label lists the two as separate rows of the same table — so that row was withheld from both our stores at 21:49 BST (the abstract's nausea 20.8% and decreased appetite 41.7%, which it does state, stay in their bases). Abdominal pain moved from 116 stated rates across 17 studies to 115 across 16, 6.6% to 6.5%, with its 1–42% interval unchanged; no other row moved; site-wide 1,125 → 1,124 eligible rates from the same 30 studies.
Restated 00:58 BST on 17 September 2026 — one row, and this time a guard found it. The correction above closed with an admission: nothing we run checks that the sentence a rate was read from names the effect it is filed under. That check now exists, and its first run over all 1,124 eligible rates flagged exactly one — a registry row in the acid reflux base. NCT05548647 posts dyspepsia at 24 of 72 (33.3%) and posts no reflux term at all; the row had been stored as an acid reflux rate on 18 August, before the collector's tracked-term map gated that path. Dyspepsia is a separate adverse-event term — the FDA Wegovy label lists dyspepsia (9% against 3% on placebo) and gastroesophageal reflux disease (5% against 3%) as separate rows of the same table — and the other 14 registry entries in this base all read "Gastrooesophageal reflux disease", so the row was withheld from both stores at 00:51 BST. Acid reflux moved from 52 stated rates across 14 studies to 51 across 13, 5.2% to 5.0%, and its interval narrowed from 0–56% to 3–9%: the withheld row was six times the pooled rate, and nearly all of the spread between studies was that one row. No other row moved; site-wide 1,124 → 1,123 eligible rates from the same 30 studies.
Restated 03:43 BST on 17 September 2026 — two rows, and this time nothing was withheld except a denominator. Every correction above removed a rate. This one removes a sample size. Reading by hand the eight eligible rates our new term-matching guard reports as "unlocatable" — rows whose stated figure it cannot find in the text we store, which is its own declared blind spot — sent us back to the four PubMed abstracts behind them. Six of the eight are verbatim correct in value, term and denominator. Two are not: SURPASS-2 (PMID 34170647, tirzepatide against semaglutide) reports "diarrhea, 13 to 16% and 12%; and vomiting, 6 to 10% and 8%" — tirzepatide's range first, the semaglutide 1 mg arm's single figure second — and our extractor took the semaglutide values correctly, then attached to them the trial's total randomised population of 1,879, which is all four arms. The abstract states no per-arm denominator anywhere. Because we weight each source by the population it reports, that made this one abstract the third-heaviest entry in both bases, carrying 8.7% of the diarrhoea pool and 8.6% of the vomiting pool on a number four times larger than the arm it describes; it was also published as that source's stated sample size on our API. We will not substitute 1,879 ÷ 4 — that is arithmetic we would be doing, not a figure the paper states — so the sample size is now null and the rates, which are correctly read and correctly attributed, stay. Diarrhoea moved from 16.5% to 17.0% and vomiting from 9.9% to 10.0% as the rest of the evidence took back the weight; both intervals, both rate counts and both study counts are unchanged, as is the site-wide base at 1,123 eligible rates from 30 studies.
Restated 09:48 BST on 18 September 2026 — ten rows, and the first study to enter this table since the registry re-collection finished. The guard reported ten of the fourteen rows drifted. Yesterday's scheduled 05:17 run died mid-collection when this laptop's battery ran flat; relaunched by hand, it completed at 09:16 BST and stored ten rows from one newly pinned pivotal trial whose posted adverse-event table we had never collected — STEP 3 (NCT03611582, semaglutide 2.4 mg as an adjunct to intensive behavioural therapy, 611 randomised). Every one of the ten drifted rows gained exactly one stated rate and exactly one distinct study, all of them that trial's semaglutide 2.4 mg arm (n=407): nausea 237 of 407 (58.2%), constipation 150 (36.9%), diarrhoea 147 (36.1%), vomiting 111 (27.3%), headache 78 (19.2%), abdominal pain 53 (13.0%), dizziness 52 (12.8%), fatigue 52 (12.8%), decreased appetite 35 (8.6%) and gastro-oesophageal reflux 25 (6.1%). Its placebo arm is excluded, as every placebo arm is. Those are high figures for a single arm, so eight point estimates rose and two were unchanged at one decimal place, none moving by more than 0.9 points (headache 9.9% to 10.8%, fatigue 6.1% to 6.9%, abdominal pain 6.5% to 7.2%, nausea 22.5% to 23.2%); six intervals widened, two narrowed, one shifted up without changing width and one did not move. No source-diversity label moved. Site-wide, the eligible base went from 1,123 rates across 30 distinct studies to 1,133 across 31 (current figure). Previous values should not be cited.
Restated 06:55 BST on 19 September 2026 — six rows: one pivotal trial joined, and one rate that was the same trial quoted second-hand left. The guard reported six of the fourteen rows drifted overnight. The 19 September 05:17 run stored the posted adverse-event table of SUSTAIN-3 (NCT01885208, semaglutide 1.0 mg against exenatide extended-release in type 2 diabetes) — twelve rates, both non-placebo arms of six effects: nausea 22.3% and 11.9%, diarrhoea 11.4% and 8.4%, headache 9.4% and 9.6%, constipation 6.4% and 5.2%, decreased appetite 7.9% and 5.2%, vomiting 7.2% and 6.2%, the first figure of each pair being the semaglutide arm (n=404) and the second its exenatide comparator (n=405), which we track as one of our nine drugs. Reading those rows against what we already held then found a double count: a narrative review of perioperative GLP-1 management (PMID 41445996) quotes this very trial — "nausea (22%), diarrhea (11%), and vomiting (7%)" — and those three restated figures had been stored as clinical evidence from a distinct source since 17 September. They are the same three arms of the same trial, to the review's rounding, so they were withheld from both our stores at 06:36 BST as secondary_source_duplicate; the rows stay in the corpus with their excerpts so the exclusion is auditable. Net movement across the six rows: nausea 23.2% to 23.0% (134 rates → 135, 29 studies unchanged), diarrhoea 17.3% to 17.2% (127 → 128, 24), headache 10.8% to 10.7% (118 → 120, 19 → 20), constipation 12.0% to 11.8% (126 → 128, 23 → 24), reduced appetite 9.4% to 9.3% (123 → 125, 23 → 24) and vomiting 10.4% to 10.3% (132 → 133, 27). No point estimate moved by more than 0.2 points, three intervals moved by a single point (headache 4–25% to 4–23%, reduced appetite 1–59% to 1–58%, vomiting 3–32% to 3–33%) and no source-diversity label moved. Site-wide the eligible base went from 1,133 rates across 31 distinct studies to 1,142 across 31 (current figure) — twelve rates in, three out, one study in and one out. Previous values should not be cited.
Amended 22 September 2026 — the rule, not the rows. The 10 September note above says this table is re-captured whenever any row's membership changes, and the 13 September note says a guard now enforces that. Both promise more than we can keep. Our collector runs once a day, early in the morning, and on most days it changes the membership of at least one row; this table is then re-captured by hand, so for some hours on most days it trails the live data by construction — and longer on a day the machine that runs the job is down, which has happened (on 18 September its battery ran flat). The rule this table actually follows is narrower: we aim to re-capture it within 24 hours of a membership change, the timestamp in the heading always tells you how old the capture is, and in between the live API that every row links to is authoritative. The guard is unchanged: it compares every row against the live API with our daily job and fails on any disagreement — it is the trigger for a re-capture, not a guarantee that none is ever owed. When this note was written all fourteen rows matched the live API.
A thin row deserves reading exactly as labelled: hair loss (8 rates from 3 studies as of 15 September 2026; 5 from 2 on 14 September, and until that morning 3 rates, all three arms of one trial, labelled "very low") carries one of the tightest-looking intervals in the table and the weakest evidence in it — the source-diversity column, not the rate column and not the interval width, is the first thing to read. Gallstones sat in the same description until 10 September 2026, when three trial registries' SAE tables took it to four studies.
"Rates / sources" is the eligible rate-point count and, after collapsing to one entry per distinct source (so one paper stating several rates can't out-vote five independent studies), the count our source-diversity label is actually based on. Every figure is served live at the link beside it — the numbers above are a same-day snapshot; check the API for the current value, since the corpus updates daily.
The effects with no corpus-derived clinical rate (5 as of 4 September 2026; 3 since 7 September 2026; 2 since 8 September 2026; 1 since 10 September 2026)
Correction, 8 September 2026. Fatigue left this table on 8 September 2026 and is now in the corpus-derived table above: the pinned registry fetch reached STEP 1 (NCT03548935) and STEP 2 (NCT03552757), whose posted adverse-event tables state fatigue at 8.0%, 4.7% and 6.9% on semaglutide 2.4 mg — 3 stated rates from 2 sources, pooled 7.5% (95% CI 5–11%), confidence "low". Its FDA-label figure (11%) still sits in the static literature table the API serves as clinicalRates, but no surface falls back to it for fatigue any more. Two effects remain below.
Correction, 7 September 2026. Hair loss and dizziness left this fallback table on 7 September 2026 and are now in the corpus-derived table above, not just noted here: a pinned fetch of SURMOUNT-1's posted adverse-event table stored per-arm rates for both (3 stated rates from 1 source each: alopecia 5.2/5.0/5.9% and dizziness 4.9/5.8/4.3% on 5/10/15 mg tirzepatide, n=630–636 per arm). Fatigue, emotional blunting and pancreatitis remain below as described; an earlier version of this section kept listing all 5 rows here after the heading above had already been updated to say 3 — that mismatch is fixed with this correction.
Correction, 10 September 2026. Pancreatitis left this table on 10 September 2026 and is in the corpus-derived table above: the SURMOUNT-1 and SELECT serious-adverse-event rows that had been withheld since 7 September were re-admitted with an explicit serious-AE label, giving it 8 stated rates from 2 studies, pooled 0.1% (95% CI 0–1%). Its FDA-label figure (0.2 acute-pancreatitis cases per 100 patient-years) still sits in the static literature table the API serves as clinicalRates, but no surface falls back to it for a published pancreatitis percentage any more. One effect remains below — and with it, no effect on the site now publishes a clinical percentage from the static literature-fallback table at all.
| Effect | Literature range (fallback) | Why |
|---|---|---|
| Emotional blunting | none — withdrawn 3 September 2026 | 0 eligible clinical rate points in the corpus, and the 2–5% this row showed until 2026-09-03 had no rate-stating source behind it, so no clinical number is published at all |
The row above isn't a corpus figure — the fallback table is the set of fixed, published-literature reference figures the predictor falls back to whenever an effect has no eligible clinical data yet, and as of 10 September 2026 emotional blunting is the only effect still reaching it (and it publishes no number). We label them "Clinical (literature)" everywhere they appear, specifically so they're never mistaken for our own pooled estimate. Full detail in our data explorer. (Updated 3 September 2026 under a founder-approved decision: fatigue's range became a single FDA-label figure because the label states no dose gradient, and emotional blunting's range was withdrawn because it had no rate-stating source behind it. Updated again 4 September 2026 under a second founder-approved decision: the remaining three rows, whose April-2026 ranges had no recorded derivation and whose cited trial pages state no such figures, now each carry one FDA-label figure with its table, arm and N — and pancreatitis's is a rate per patient-year, not a share of patients, which is why it is written out in words rather than as a percentage.)
Why some of these confidence intervals are wide
A few rows above have wide ranges — reduced appetite spans 0–87%, acid reflux 0–72%, nausea 7–52%, abdominal pain 1–44% and vomiting 2–44%. That isn't a mistake or a rounding artefact; it's what happens when independent sources genuinely disagree on the rate (for vomiting, SELECT's 8,803-patient semaglutide arm reports 6.5%, while a 10-participant early-phase orforglipron study reports 45% — both were in the live source list at the API link above when this paragraph was last restated on 13 September 2026, and the full per-source spread is walked through in our vomiting article, which states the source count it was last restated at rather than a current one), and our method — a simplified random-effects model (an unweighted variant inspired by DerSimonian-Laird, not the full inverse-variance-weighted estimator) — widens the interval to reflect that real disagreement rather than pretending the sources agree when they don't. A narrow interval built from 2-3 sources would be the false confidence; a wide one is the honest answer to "how sure are we." As more independent sources accumulate for an effect, the interval narrows — you can watch that happen over time at the live API.
The rule behind every number here
Every rate above passed the same three checks before it was allowed to count: it has to come from an external, linkable clinical trial or regulatory source that our pipeline genuinely extracted a rate from (no self-citations, no news aggregator results, and — since the 31 August 2026 correction above — no April-2026 seed rows, whatever URL they wear); spontaneous adverse-event report shares (FAERS) are tracked separately and never blended into an incidence estimate; and a single paper stating multiple rates collapses to one vote. Full detail on the eligibility rule and which effects don't yet clear the higher bar needed for a fully calibrated statistical model is in our companion piece on evidence strength.
See the complete picture — all 15 effects, both tracks, live counts and every source — at magistra.health/en/data-api.
See your own numbers
Our free predictor estimates your side-effect risk and weight trajectory, with the stated rates and distinct sources shown behind every figure. No signup required.
Open the predictorWorking from the data itself? The dated snapshot behind these figures is available as a one-off purchase, alongside the free public API: Data & API.