How we researched GLP-1 trials with ClinLit AI: questions, queries and checks

Our report Beyond diabetes and obesity: where GLP-1 drugs are being tested in 2026 was researched with ClinLit AI. This post shows how: the questions we asked, what it did with them, where we had to be more precise, and how we checked the answers. If you research trials yourself, you can follow the same steps for your own drug or disease.

We asked 15 questions in all. Each answer took between 14 seconds and under 3 minutes, using between 4 and 22 tool calls: SQL queries on the registry and PubMed, reads of protocol PDFs, and web searches.

Step 1: ask the open question, then look at how it was answered

We started with the question we actually had:

How many interventional trials on ClinicalTrials.gov test a GLP-1 receptor agonist, including dual and triple agonists such as tirzepatide, retatrutide and survodutide? Of those, how many study conditions other than type 2 diabetes, type 1 diabetes and obesity?

ClinLit AI answered 1,449 trials, 442 of them with another condition. It also explained exactly what it had done: which drug names, brand names and codes it matched, and how it decided what counted as “another condition”. It listed its own doubts too:

Borderline inclusions in the 442: 21 trials whose only outside term is the vague umbrella “metabolism and nutrition disorder” … and “prediabetes”/“metabolic syndrome”/“glucose metabolism disorders” counted as other.

Reading that, we saw two problems with our question, not with the answer. Its drug list missed some newer agents, such as maridebart cafraglutide and pemvidutide. And “another condition” counted a trial of semaglutide in healthy volunteers, or one listing a vague metabolic term, the same as a trial in heart failure.

The lesson: an open question gets a reasonable answer to a vague question. For a number you will publish, decide the definitions yourself.

Step 2: give it a definition you can publish

We wrote the definition out in full: the drug names, brands and codes to match, and an ordered list of disease areas with their keywords, where each trial goes into the first area it matches. The question is in the first “Reproduce this” box of the report, with the SQL ClinLit AI wrote from it.

It found 1,543 trials, produced the area table we published, and added notes on the edges of the definition that we would not have spotted:

• Searching the registry’s list of other names for each intervention as well would add 67 trials, mostly where a brand name such as “Ozempic” appears only there.
• Keywords decide the edges: a trial in intracerebral haemorrhage lands in “other” because “stroke” isn’t in its title or conditions.
• The type 2 diabetes area includes 4 gestational diabetes and 6 prediabetes trials, because “diabet” matches them.

We kept the definition as it was, so that anyone pasting it gets the same numbers, and mentioned these limits in the report.

Step 3: ask follow-ups in the same conversation

Once the definition was set, follow-ups in the same conversation reused it: trials by start year as a chart, sponsors by area, and which drugs the newer trials use. One follow-up produced a caveat we added to the report’s chart:

83 of the 166 2026 starts are ESTIMATED start dates rather than ACTUAL ones … treat the 46% as an upper bound.

A conversation keeps its last 10 messages. For a long piece of research, start a new conversation for each topic and paste the definition again.

Step 4: read the protocols, not just the registry

The registry says what a trial measures. The protocol says why, and how that changed. ClinLit AI can read the protocols, statistical analysis plans and consent forms that sponsors post on ClinicalTrials.gov. We asked, for example:

In the SUMMIT trial of tirzepatide in heart failure with preserved ejection fraction (NCT04847557), how did the primary endpoints change between protocol amendments, and what reason does the protocol give? Cite the protocol pages.

It listed the trial’s documents, searched the protocol for the amendment summary, read pages 4 to 13, the document history at the end and the rationale on pages 50–51, and quoted the sponsor’s reasons with page numbers. We asked similar questions about SURMOUNT-OSA, SYNERGY-NASH, the eligibility criteria of five trials, and the sample size calculations of two alcohol use disorder trials. Each answer took 14 to 27 seconds.

These answers went further than our own first reading of the same PDFs. ClinLit AI found:

• the second reason SUMMIT gave for its new endpoints: the win-ratio composite was dropped because it presents “significant complexity in analytics and interpretation”;
• an earlier SURMOUNT-OSA amendment, in September 2022, that first added the absolute change in AHI “to align with recommendations from health authority”;
• that SYNERGY-NASH added its FAST-score pre-screen after “early blinded biopsy data” from the trial itself;
• that STEP-HFpEF lowers its NT-proBNP thresholds for people with a BMI of 35 or more.

Step 5: check the citations

Every protocol answer cites PDF pages, and we opened each one. All the quoted passages were there, with one slip. In the SUMMIT answer, ClinLit AI read the protocol’s document history and gave amendment (c) the date of amendment (b), 21 January 2022. The protocol’s own approval stamp shows amendment (c) was approved on 14 February 2024. One look at the cited page caught it.

That is the point of citations. ClinLit AI, like any AI, can make mistakes; because every claim points to a record, a page or an article, checking it takes seconds rather than a fresh search.

Step 6: the literature

For PubMed, we again gave an exact definition: the substances that count as GLP-1 drugs, and the MeSH topics for each area. ClinLit AI checked every name against PubMed’s lookup tables before counting, and corrected one of ours:

‘Depressive Disorder, Major’ does not exist as a descriptor in this data — the descriptor is ‘Major Depressive Disorder’.

It also explained what such counts can miss: the newest articles may not be indexed yet, and a drug mentioned only in the abstract isn’t counted.

Step 7: beyond the data, to the web

The registry snapshot and the PubMed baseline end at fixed dates. For results since then, such as Altimmune’s July 2026 alcohol use disorder trial and Novo Nordisk’s Alzheimer’s trials, ClinLit AI searched the web, limited to the companies’ and newswire sites, read the announcements, and said clearly that these results came from company announcements, not peer-reviewed papers. It checked the registry and PubMed first, to confirm there was no published paper yet.

The cross-check

Before using ClinLit AI, we had written our own SQL against the same database. It found 1,555 trials rather than 1,543, and area counts within a few trials of ClinLit AI’s (heart and blood vessels, the largest gap, 99 against 87). The differences come from definitions, not errors: our query also matched MeSH condition terms, and its keyword lists differed slightly. The literature counts matched ours exactly, apart from the MeSH name ClinLit AI corrected.

We published ClinLit AI’s numbers, because those are the ones you can reproduce by pasting the same questions.

What we learned

• Start open, then define. The first answer shows you what choices the question involves. Then make those choices yourself.
• Paste the definition, keep the conversation. Follow-ups reuse it; a new topic is a new conversation.
• Ask for page numbers. “Cite the protocol pages” turns a summary into something you can check.
• Read the caveats. ClinLit AI says where a count may be too high or too low; those notes belong in what you publish.
• Open the cited pages before you rely on a detail. It takes seconds.

Try the questions below, or take the definition from the report and change the drug list to your own.