AlphaCitation / Research & Insights
What AI says. What the evidence shows.
Original research, field investigations and perspectives on how AI systems represent medicines and brands, interpret evidence and influence decisions.
13 publicationsLatest October 19, 2026
What we publish
Research Papers
Structured studies of a defined research question, with documented methods, findings, supporting evidence and limitations.
Field Notes
Focused investigations of a particular observation, experiment, anomaly or emerging pattern, with their evidence and scope documented.
Perspectives
Evidence-informed analysis, frameworks and arguments about AI, information integrity, measurement and decision-making.
Featured research
Research Paper 01
Does New Medical Evidence Change AI Treatment Answers?
Whether evidence presented at EASD 2026 reached what four AI engine APIs told people about three diabetes and obesity medicines, in the UAE, Singapore and Saudi Arabia, from 22 August to 11 October 2026. None of the seven registered trial readouts was verified in its own series' answers by day 10 to 12.
- Study
- Longitudinal study
- Subject
- Diabetes and obesity
- Market
- UAE, Singapore, Saudi Arabia
- Engines
- 4 AI engines
- Observed
- August 22, 2026 – October 11, 2026
Latest publication
Perspective 06
Who Gets to Audit AI?
Argues that the right to evaluate an AI system is incomplete without a proportionate right to keep the evidence needed to verify, challenge and correct the findings, and proposes an Audit Clause providers can adopt. Retained Perplexity answers, which changed what one of AlphaCitation's own published results could be taken to mean, are the worked case.
- Theme
- AI governance and independent audit
The Register
Every publication, newest first.
Showing 13 of 13: All publications.
Perspective 06
Who Gets to Audit AI?
Argues that the right to evaluate an AI system is incomplete without a proportionate right to keep the evidence needed to verify, challenge and correct the findings, and proposes an Audit Clause providers can adopt. Retained Perplexity answers, which changed what one of AlphaCitation's own published results could be taken to mean, are the worked case.
- Theme
- AI governance and independent audit
Research Paper 01
Does New Medical Evidence Change AI Treatment Answers?
Whether evidence presented at EASD 2026 reached what four AI engine APIs told people about three diabetes and obesity medicines, in the UAE, Singapore and Saudi Arabia, from 22 August to 11 October 2026. None of the seven registered trial readouts was verified in its own series' answers by day 10 to 12.
- Study
- Longitudinal study
- Subject
- Diabetes and obesity
- Market
- UAE, Singapore, Saudi Arabia
- Engines
- 4 AI engines
- Observed
- August 22, 2026 – October 11, 2026
Field Note 06
The Medicine Was Approved. ChatGPT Ruled It Out Anyway.
Asked one fixed head-to-head question 13 times through OpenAI's API (gpt-4o-mini, not the consumer ChatGPT app the title names), the engine ruled out a weight-management medicine registered in the UAE in 10 runs: six answers said it was not approved, four that it was not available. Across 3,104 answers from four engines, the regulator's live drug register was never a source.
- Subject
- Obesity
- Market
- United Arab Emirates
- Engines
- 4 AI engines
- Observed
- August 22, 2026 – September 25, 2026
Field Note 05
AI Can Choose the Medicine Without Naming the Brand
A medicine can win the treatment decision and still lose the brand credit. One engine named the product in 26 of the 28 Alzheimer's decisions in which it chose the molecule (93%), and in none of the 30 lung-cancer decisions in which it chose the molecule; a single visibility score cannot tell the two apart.
- Study
- Four pre-specified studies
- Subject
- Obesity, Alzheimer’s disease and oncology
- Engines
- 3 AI engines
- Observed
- September 27, 2026
Field Note 04
The Answer Moved Before the Evidence Did
In the month before EASD 2026, one engine's naming of a GLP-1 medicine in the UAE stepped from 36% to 52% of answers as a Dubai clinic's comparison page entered its sources, while the four-engine total moved three points. A congress effect needs a band and a source record, not a before number.
- Subject
- Diabetes and obesity
- Market
- United Arab Emirates and Singapore
- Engines
- 4 AI engines
- Observed
- August 22, 2026 – September 23, 2026
Field Note 03
We Tested an Advantage a Marketed Product Did Not Have
A counterfactual on four engines tested what an approved cardiovascular indication would do to a weight-management medicine's standing in AI treatment decisions. Favourable decisions rose 8.8 points, in the same direction on every engine.
- Study
- Counterfactual experiment
- Subject
- Weight management
- Engines
- 4 AI engines
Field Note 02
Where the Answer Stops Being Global
Saudi Arabia registered a haemophilia A prophylaxis no other regulator has registered. One instrument, put to four engines in two Gulf markets, found the part of a treatment pathway that read the same in both markets and the part that did not.
- Subject
- Oncology
- Market
- Saudi Arabia and United Arab Emirates
- Engines
- 4 AI engines
- Observed
- August 31, 2026 – September 2, 2026
Field Note 01
The Decision Point Moved Earlier Than the Answer
Saudi Arabia's regulator approved a therapy 112 days before the FDA. In one pass across four engines before the US approval, the earlier blood-test trigger appeared in all four answers to a question naming the product and in none of the four to the unbranded clinical question.
- Subject
- Oncology
- Market
- Saudi Arabia
- Engines
- 4 AI engines
- Observed
- August 31, 2026 – September 2, 2026
Perspective 05
Being Mentioned Is Not the Same as Being Chosen
Presence, consideration, recommendation and decision survival are four different states. Most AI-visibility measurement reports the first and leaves executives to infer the other three.
- Theme
- Measurement framework
Perspective 04
What AI Leaves Out About Your Brand
A brand can be highly visible and still sound, in AI answers, as if it had one strength instead of five. Why narrative coverage and risk framing need measuring, not just mention counts.
- Theme
- Brand narrative
Perspective 03
An AI Engine Said a Medicine on the UAE Register Was Not Approved There. A Mention Count Would Have Scored It Present.
Five AlphaCitation investigations of AI engine APIs. In one, the UAE regulator’s register contradicted 6 of 13 answers from OpenAI’s gpt-4o-mini to one head-to-head, and five answers from Gemini and Claude about a second medicine. The other four set out what a mention count misses: question, market, readout verification and brand credit.
- Theme
- Measurement framework
Perspective 02
We Don't Guess If AI Said Your Name. We Test the Reading.
How mention detection was tested in July 2026 against a reference set labelled by people: zero errors, which puts accuracy at 97.3% or better at 95% confidence on that set. Sentiment, the harder call, is graded by an independent AI judge: 85.4% agreement.
- Theme
- Measurement accuracy
Perspective 01
How AI Brand Visibility Is Actually Measured (And Why Most Tools Don't Bother)
Why a trustworthy AI-visibility score needs confidence intervals and independently measured uncertainty, rather than a single number from a prompt wrapper.
- Theme
- Measurement method