Original research · 2026-07-13
Can AI answer engines cite peptide research correctly?
We put 180 questions — frozen and hash-published before any engine was queried — about 30 research compounds to four AI answer engines, plus a smaller-model control, and checked every source they cited. A ten-compound subset was then repeated in five further languages. Across 5,130 citations, roughly 2.3% of the peer-reviewed papers the engines cited were not actually about the compound they were cited for — in one case a study of an antibiotic-resistant bacterium, cited as evidence for a peptide. Most answers also disregarded the compounds' research-only status, and outside English the share of citations drawn from peer-reviewed literature fell from 24% to between 6% and 11%.
This is a study of how the engines cite sources. It is not a claim that any compound is safe, effective, or suitable for any use.
What we found
Across every answer the four engines gave about compounds we sell — with our public per-batch lab reports freely available — we were cited zero times. This benchmark measures the engines, not us; we report our own absence for transparency.
Asked about compounds supplied strictly for laboratory research, the engines answered as though the question concerned human use — presenting preclinical findings as applicable to people, or returning an administration regimen. We report the rate at which this happened. We do not reproduce, summarise, or endorse what they returned.
Every one of these URLs resolved correctly. Retrieving and reading each source revealed papers about a different compound entirely — in one case a study of an antibiotic-resistant bacterium cited as evidence for a peptide — and a single general review passed off as compound-specific evidence for three unrelated compounds.
In German, French, Italian, Spanish and Portuguese the share of citations that are actual peer-reviewed literature falls by roughly three to four times versus English. Commercial and user-generated sources fill the gap.
How the study was run
Thirty research compounds were each asked six standard question types — research findings, safety, half-life, dosage, available studies, and molecular identity. Each question type targets a specific way a citation can fail.
The prompt set was frozen before any engine was queried, and its SHA-256 hash published, so questions could not be selected after seeing results. English set: 5ce10f34cf033b81ee8b82b174446ca59fd396a78ced68b3022ddd5a9f6c319d. Multilingual set: 93566d613b614ed0bd93d96a71ce048acc31b66f5f78a097f08682bc3773372c.
Every cited URL was then fetched to confirm it resolves. For every PubMed or PMC citation, the source's actual abstract was retrieved from NCBI and compared against the compound it was cited for, with a list of accepted synonyms supplied so that PT-141/bremelanotide or TB-500/thymosin beta-4 count as the same compound rather than as errors. Claim-level dimensions were scored against a fixed published rubric by Claude Haiku 4.5; every misattribution reported here was reviewed by a human.
899 English answers and 274 non-English answers were scored, covering 5,130 citations and 444 unique primary sources.
Results by engine
899 answers across five engine configurations (180 each; one engine returned 179). “Links resolve” = the cited URL loads. “No human dosing” = the answer avoided giving a human protocol or presenting preclinical findings as applicable to people.
| Engine | Citations / answer | Links resolve | No human dosing | Identity correct | Leans on |
|---|---|---|---|---|---|
| ChatGPTgpt-5-chat-latest | 10.5 | 86% | 46% | 89% | mixed / commercial |
| Perplexitysonar | 8.9 | 79% | 39% | 88% | mixed / UGC |
| Geminigemini-2.5-flash | 31.8 | 85% | 31% | 92% | UGC |
| Claudeclaude-sonnet-4-5 | 9 | 84% | 47% | 87% | mixed / UGC |
| ChatGPT (smaller model)gpt-4o-mini | 4 | 87% | 42% | 91% | commercial |
The final row is a smaller model from the same vendor, included as a control for whether model size explains the differences. It does not: the smaller model cited fewer sources but scored comparably on accuracy.
Research-only status is not respected
One of the six question types probes how an engine handles a compound's research-only status. Between 23 and 29 of the 30 answers per engine treated it as a question about human use and returned an administration regimen. We report the rate as a measured property of the engines. The content of those answers is deliberately not reproduced, summarised, or endorsed anywhere in this report or its dataset.
When the link works but the paper is wrong
A resolving link is not the same as a correct citation. Of 444 unique peer-reviewed sources cited across the study, 10 (2.3%) were not about the compound they were cited for — after accepting all known synonyms. Every one of those URLs loaded correctly, so no link-checking approach would catch them.
Verification step: 14sources were flagged on the first pass, and each was then re-checked against the article's authoritative title. 4 of those turned out to be artefacts of our own retrieval — some archive records return journal metadata instead of an abstract — and were reinstated as correctly cited. Only the remaining 10 are reported as errors.
- KPV — cited to hypervirulent Klebsiella pneumoniae — a bacterial pathogen, not the peptide.ChatGPT
- Selank, Semax and MK-677 — cited to a single general review of therapeutic peptides, cited as specific evidence for all three.ChatGPT
- GHK-Cu — cited to skin toxicity of copper compounds.ChatGPT
- Thymalin — cited to haematopoietic stem cell therapy in COVID-19.ChatGPT
- Ipamorelin — cited to a trial of CJC-1295, a different peptide.Perplexity
- AOD-9604 — cited to a general obesity pharmacotherapy review.Gemini, Perplexity
The science thins out in other languages
The same questions were asked again in five European languages. Non-English answers cited more sources, and more of those links resolved — but the share drawn from peer-reviewed literature fell sharply. Commercial and user-generated pages made up the difference.
| Language | Answers | Citations / answer | Links resolve | From peer-reviewed literature |
|---|---|---|---|---|
| English | 120 | 10.1 | 82% | 24% |
| German | 56 | 14.1 | 93% | 6% |
| French | 56 | 13.2 | 91% | 7% |
| Spanish | 54 | 12.7 | 90% | 11% |
| Italian | 54 | 11.4 | 92% | 6% |
| Portuguese | 54 | 12.9 | 93% | 9% |
What the engines actually cite
Across the four primary engines, 23% of citations came from peer-reviewed literature and 77% from commercial pages, vendor blogs, encyclopaedias, and user-generated platforms. The most-cited domains were:
Data and reproducibility
The frozen prompt sets with their hashes, every scored answer with its full citation list and resolution status, the source-level misattribution checks, and the scoring rubric and code are all published below under a CC BY 4.0 licence — so any figure in this report can be re-derived independently. Each run is versioned; we plan to repeat the study quarterly.
- English answers, scored (899 rows, JSONL) — every citation, resolution status and score
- Multilingual answers, scored (274 rows, JSONL) — German, French, Spanish, Italian, Portuguese
- Primary-source misattribution checks (JSONL) — per-source NCBI verdicts with matched synonyms
- Frozen prompt set — English (v1.0) — verify against /data/prompts-v1.0-en.sha256
- SHA-256 hash — English prompt set — published before any engine was queried
- Frozen prompt set — six languages (v2.0) — verify against /data/prompts-v2.0-multi.sha256
- SHA-256 hash — multilingual prompt set — published before any engine was queried
- Scoring rubric (verbatim judge prompts) — every dimension exactly as applied
- Scoring code — judge pass — collection, entailment and analysis scripts alongside it in /data/method/
One deliberate omission
The dataset contains every citation, resolution result, and score, but not the verbatim answer text. Those answers frequently contain human dosing protocols for compounds sold for laboratory research only — the precise failure this study documents — and republishing them would propagate the problem we are measuring. Every claim in this report concerns citation behaviour, and all the evidence for those claims is included. Researchers who need the raw answer text for verification can request it.
Limitations
- — This is a measurement of how answer engines cite sources. It is not a claim about whether any compound is safe, effective, or suitable for any use.
- — Claim-level dimensions were judged by a language model (Claude Haiku 4.5) against the fixed rubric published with this report, not by human experts. Every misattribution reported was reviewed by a human against the source title before publication.
- — The misattribution check tests whether a cited source is genuinely about the compound it was cited for. It does not verify every individual sentence against its source.
- — The multilingual comparison covers a ten-compound subset across two engines, not the full corpus, and the per-language samples are small (54–56 answers each). Those percentages should be read as indicative of direction, not as precise population estimates.
- — The prompt sets were frozen and their hashes published before any engine was queried. That is not the same as preregistration: there is no third-party registry entry and no pre-committed analysis plan.
- — Engines are non-deterministic and their indexes change. Results describe the run dated in this report and are versioned accordingly.
Declared conflict of interest
CertaPeptides (CERTALAB S.R.L.) sells research peptides and therefore has a commercial interest in how AI answer engines describe this product category. This study measures the engines' citation behaviour, not our products. Our own domain appeared in zero of the 5,130 citations analysed, and we report that result unchanged. The prompt sets were frozen and hash-published before the runs, and the scored dataset, the scoring rubric and the scoring code are all released so the work can be checked — or contradicted — independently.
Questions about this study
What does this benchmark actually measure?
Whether AI answer engines cite real, relevant, correctly-attributed sources when answering questions about research peptides — not whether their conclusions are right, and not whether any compound works.
How were the questions chosen?
Thirty compounds were each asked six standard question types (research findings, safety, half-life, dosage, available studies, molecular identity). The full prompt set was frozen and its SHA-256 hash published before any engine was queried, so questions could not be selected after seeing results. This is a transparency measure, not formal preregistration — there is no registry entry or pre-committed analysis plan.
How is a citation judged 'misattributed'?
For every PubMed or PMC citation, the source's actual abstract is fetched from NCBI and compared against the compound it was cited for. Known synonyms are supplied to the judge, so PT-141/bremelanotide or TB-500/thymosin beta-4 are treated as the same compound, not as errors.
Why does citation quality differ between languages?
In the five non-English languages tested, answers cited more sources and more of those links resolved, but the share drawn from peer-reviewed literature fell from 24% to between 6% and 11%. Commercial and user-generated pages made up the difference.
Is the underlying data available?
Yes. The frozen prompt sets with their hashes, every scored answer with its citations, and the source-level misattribution checks are published alongside this report.
All content is for research purposes only. Nothing in this report is medical advice, and no compound described is offered for human use.