Drug development starts with a receptor and a reason to believe it matters. The alcohol work on this drug class did not begin with patients noticing they had stopped ordering a second beer. It began with the observation that the GLP-1 receptor is expressed in brain regions that handle reward, and that putting an agonist directly into those regions changes how a rat drinks. Everything downstream — the registry hazard ratios, the trials, the press coverage — is an attempt to find out whether that laboratory finding survives contact with people.
A separate question travels under the same search terms and is answered elsewhere: whether a glass of wine is safe on a weekly injection is a labeling matter, set out in the page on drinking while treated. This one is about the indication.
The receptor sits inside the reward circuit
The founding experiment gave mice exendin-4 at a dose with no effect on its own and then gave them alcohol. Alcohol-induced locomotor stimulation was attenuated and so was dopamine release in the nucleus accumbens. Conditioned place preference for alcohol — the memory that the drug was rewarding — was abolished by both acute and chronic treatment. In rats, intake fell and so did responding for alcohol on a progressive-ratio schedule, which measures how hard an animal will work for it.[1]
The obvious objection is that a gut hormone analog could produce all of that from the periphery, by way of nausea and slowed stomach emptying. Two lines of work speak to it. Fluorescently labeled semaglutide has been imaged inside the nucleus accumbens of alcohol-drinking rats of both sexes, and in male mice the same drug blunted the rise in accumbal dopamine that alcohol otherwise produces.[5] The drug reaches the tissue, and the tissue responds.
Infusing one site at a time is where it gets complicated
A 2019 study infused exendin-4 bilaterally into four candidate regions and compared them. Into the nucleus accumbens shell, the agonist blocked alcohol-induced locomotor stimulation, blocked the memory of alcohol reward and reduced intake — the full set. Into the anterior ventral tegmental area, it did none of those things. Into the posterior ventral tegmental area it blocked locomotor stimulation only, and into the laterodorsal tegmental area it reduced stimulation and intake but left reward memory alone. Accumbal receptor expression was higher in rats that drank heavily than in rats that did not.[2]
A separate group reached a different verdict on the midbrain. Infusing exendin-4 into the ventral tegmental area of male rats did reduce alcohol self-administration, most visibly in the heaviest drinkers, at a dose that changed neither food intake nor locomotor activity. It did not, however, reduce reacquisition after extinction or the motivation to work for alcohol.[4] Two careful laboratories, one region, two answers.
The brainstem result is the one that most directly addresses central versus peripheral action. Exendin-4 delivered into the nucleus of the solitary tract suppressed locomotor stimulation, accumbal dopamine release and reward memory in mice, and cut intake dose-dependently in rats. Then the causal step: pharmacologically suppressing the receptor in that nucleus prevented systemically administered exendin-4 from blocking the locomotor response.[3] A drug given under the skin was acting through a receptor population in the brain.
A 2026 synthesis of the preclinical literature accepts the reward hypothesis as the leading explanation and then names the alternatives it cannot exclude: altered gastric emptying, nausea and heightened stress.[6] Nausea in particular is not a trivial confound in an animal that cannot report it, and its human time course overlaps the weeks in which drinking outcomes are measured.
Human genetics points the same way and muddies the channel
A 2015 translational study tested a functional receptor variant, 168Ser, in a case-control sample of 908 people and found a nominal association with alcohol use disorder (P = 0.004), partially replicated in males of a 3,803-person confirmation sample (P = 0.033). Carriers of the Ser/Ser genotype self-administered more alcohol and reached higher breath alcohol levels in an intravenous laboratory task of 81 social drinkers.[7] Variation in the receptor tracks variation in drinking, which is what a real target looks like.
A 2025 multi-ancestry drug-target Mendelian randomization study went further and produced the most useful complication in the file. Genetic proxies at the GLP1R and GIPR loci were associated with less binge drinking, at β −0.44 (95% CI −0.72 to −0.15) in the primary data and −0.13 (−0.22 to −0.04) on replication, and with lower odds of heavy drinking accompanied by psychiatric illness versus low-risk drinking, at OR 0.62 (0.45 to 0.85). Estimates for tobacco, cannabis and opioid use were consistently null.[8]
The complication is in the instrument. Those proxies model body-mass-index lowering and HbA1c lowering — metabolic effects, not reward effects — and the same analysis found the variants shifted food preference, reducing liking for fatty food at β −1.58 (−2.01 to −1.14). The authors read that as implicating appetite and metabolic regulation in the alcohol finding.[8] A genetic result consistent with a reward mechanism is also consistent with an appetite one, and the overlap with mood and behavior outcomes is not resolved by this design either.
The registry numbers are large, and they move with the comparator
The widely quoted United States figure comes from a 2024 retrospective cohort of 83,825 patients with obesity, in which semaglutide was associated with a 50% to 56% lower risk of both incidence and recurrence of alcohol use disorder over twelve months against other anti-obesity medications, with the pattern repeating in a type 2 diabetes population of 598,803.[9]
A 2026 emulation of four target trials in 40,703 adults who already had the diagnosis shows why that headline needs a denominator of its own. Against sulfonylureas the hazard ratio for an alcohol-related emergency visit or admission was 0.74 (95% CI 0.62 to 0.89); against other diabetes medicines, 0.78; against other anti-obesity medicines, 0.68. Against medications for alcohol use disorder it was 0.37 (0.29 to 0.46) in the diabetes arm and 0.35 (0.26 to 0.47) in the obesity arm.[10]
One exposure does not have two effect sizes. A comparison does. The figure roughly doubles when the control group is people who were prescribed naltrexone or acamprosate, and the most economical explanation is that those people had more severe disease, not that the drug works twice as well in their presence.
Class membership does not travel either. A 2026 emulation drawn from a records network covering more than 120 million patients compared four agents against DPP-4 inhibitors in type 2 diabetes with no prior diagnosis. Tirzepatide (n = 7,165) reached HR 0.47 (0.29 to 0.75) and semaglutide (n = 20,198) 0.68 (0.52 to 0.89) for first-ever alcohol use disorder. Liraglutide and dulaglutide reached no significant reduction at all, despite dulaglutide contributing 19,061 matched patients.[11]
Then the result that reverses the reading. A Medicare analysis of adults aged 65 and over with type 2 diabetes and a substance use disorder found a lower risk of substance-related admission against DPP-4 inhibitors, HR 0.76 (0.67 to 0.86) — but the alcohol-specific endpoint was not significant at 0.76 (0.53 to 1.08), and against SGLT2 inhibitors there were no significant differences at all.[12] Same drugs, same database era, a null.
The randomized record keeps missing its own primary endpoint
A 2026 phase 2 trial randomized 50 treatment-seeking adults with moderate to severe alcohol use disorder to oral semaglutide, 3 mg daily for four weeks and then 7 mg daily for four weeks, or placebo. The primary outcome was laboratory cue-elicited craving at week six. It did not separate from placebo, and neither did drinks per day. Heavy drinking days did, at b −0.580 (95% CI −1.012 to −0.148), as did drinks per drinking day at −1.177 (−2.307 to −0.047) and naturalistic craving at −2.195 (−4.174 to −0.216).[13]
That is the shape the randomized literature keeps producing: the prespecified primary target holds, the measures of how much people drink on a drinking occasion move, and the number of drinking occasions moves less. The largest randomized separation published to date is 13.7 percentage points of heavy drinking days over 26 weeks in 108 people who had both the diagnosis and obesity.[14] Real, and smaller than a hazard ratio of 0.35 implies.
The benchmark is naltrexone, not placebo
Alcohol use disorder already has approved pharmacotherapy, and a new entrant is judged against it. A 2023 systematic review pooled 118 randomized trials and 20,976 participants. The number needed to treat to prevent one person returning to any drinking was 11 (95% CI 1 to 32) for acamprosate and 18 (4 to 32) for oral naltrexone at 50 mg daily; for return to heavy drinking, oral naltrexone reached 11 (5 to 41). Injectable naltrexone produced 4.99 fewer drinking days across a 30-day period.[15]
No GLP-1 receptor agonist has a number needed to treat for this condition, because no trial program has been run at the size that would support one. Comparing a single 26-week result against a 118-trial evidence base is comparing a hypothesis to a standard of care.
One head-to-head attempt exists in records rather than randomization. A single health system followed 1,946 patients carrying both alcohol use disorder and metabolic dysfunction between 2017 and 2025, of whom 274 took a GLP-1 receptor agonist for at least six months against 1,272 on naltrexone, 232 on acamprosate and 168 on disulfiram. One-year relapse favored the GLP-1 group at IRR 0.55 (0.42 to 0.73). The same table reports that the group had a higher body-mass index, 35.5 against 30.1, and type 2 diabetes in 66% against 14%.[16]Propensity matching cannot fully absorb a difference that large in who gets which prescription.
What is being sold, and what is not
The FDA has licensed nothing in this class as a therapy for drinking. A prescription written for that purpose is off-label everywhere in the country, and most of what moves through cash telehealth is compounded product, which is not FDA-approved and is not reviewed by the FDA for safety, efficacy or quality before it is dispensed.
The scope problem is sharper than the regulatory one. Every human study above enrolled people who had been diagnosed, and most of them delivered the drug alongside therapy or standard care. A questionnaire that establishes a height, a weight and a payment method has not screened for a substance use disorder, has not asked about withdrawal risk and has no route to the counseling that sat underneath the trial results. Which questionnaires go that far is one of the things the seller write-ups record, on the criteria set out in the methodology. Any marketing that leans on the drinking effect belongs next to the red-flag list.
The honest summary is that the target looks real and the size of the human effect does not yet have a settled value. Anyone whose drinking has become a problem has a conversation to have with a clinician, and two approved medicines with published numbers are already on the other side of it.