Insights · Start here

How do AI answer engines choose health sources? What our study found

Edited by Abhi BasuPublished October 19, 2026Last reviewed October 19, 2026

Disclosure: Signal Fidelity Group is the research company founded by Abhi Basu, who also founded CommsRx. The study is not peer reviewed.

What did the study test?

The question was simple: if two people ask an AI system the same health question, and one of them says they are a patient or a caregiver, does the system send them to different sources?

The design, as published:

• 120 questions: three indications (oncology, type 2 diabetes, and obesity with obstructive sleep apnea), five decision stages from clinical relevance to action and access, and eight questions per cell.
• Three search-enabled systems, called through their APIs: Perplexity's Sonar, Gemini 2.5 Flash with Google Search, and GPT-5 with web search.
• Two conditions per question: the plain question, and the same question with patient or caregiver context added.
• Three repetitions of everything. Of 2,160 planned responses, 2,149 were captured.

Collection ran on September 13, 2026. The revised record was posted on September 14 and the public summary on September 21.

What changed when patient context was added?

In 105 of the 357 eligible question–system comparisons (29.4%), adding the patient or caregiver context changed the set of cited sources by more than the study's operational threshold. "Threshold" here is a statistical cut-off, not a judgment of clinical importance. More broadly, the cited sources were less alike across the two conditions than across repeat runs of the same condition in 280 of 357 pairs (78.4%).

Put plainly: in these tests, adding who was asking and why often changed where the system sent them, even when the question itself was identical.

Did every engine behave the same?

No, and the differences were large:

One system changed its sources for patients in nearly half of comparisons; another almost never did. A communications team that checks only one engine, or checks only as an anonymous user, is seeing one slice of what its audiences see.

Where in the patient journey did sources move most?

An analysis made after the data came in, and so treated as a lead rather than a finding, suggests the effect was stronger for questions about on-treatment risk and adverse events, and about taking action or getting access (36.4%), than for the three earlier stages: clinical relevance, evidence and options, and eligibility (24.8%), an odds ratio of 1.91. The study itself labels this hypothesis-generating and notes it was largely driven by Sonar.

In an exploratory comparison by therapy area, oncology questions moved most (39.0%), then type 2 diabetes (27.5%), then obesity with sleep apnea (21.8%). The study warns against reading this as an oncology effect: the patient context written for oncology questions was more than twice as long (49.2 words on average, against about 21), and once length was added to the model the difference was no longer distinguishable (the study adds that this does not prove length was the cause).

What does the study not show?

The limits matter as much as the headline, and the study states them:

• It is not peer reviewed.
• Its pre-specified hypotheses were not supported. The source changes were real, but they did not follow the pattern the researchers predicted in advance.
• It did not test medical accuracy. A changed source is not a wrong answer, and an unchanged source is not a right one.
• No patients took part. The patient context was written into the question.
• These are API observations, not consumer apps. Signed-in apps can add memory and personalization that the study did not reproduce.

What does this mean for pharma communications teams?

Three practical consequences follow, each an inference from the study rather than a finding of it:

1. Monitor as your audiences, not as yourself. If patient context changes sources, a monitoring program should ask questions the way a patient, a caregiver and a physician would ask them. See How do you monitor what AI says about a drug or a disease?
2. The source layer is the lever. Answer engines draw on pages they can find and trust. Other research points the same way: an Ahrefs analysis found that only 37.9% of URLs cited in Google's AI Overviews also appeared in the first ten result blocks, and a University of Toronto preprint found AI search weighted its results toward earned media in every category and region it tested (in US software-product queries, 72.7% earned against 45.4% on Google). Which patient-facing sources carry your disease area matters as much as your own pages.
3. The later stages of the journey deserve more attention. If the lead holds, questions about side effects and access are where answers are most likely to move, which is also where accuracy matters most.

Where can I read the full study?

The public summary, the study record and the revised results are on the Signal Fidelity Group research page, with a plain-language summary on commsrx.com/evidence.

Sources

• Signal Fidelity Group, The Disclosure Effect, public summary September 21, 2026; revised results September 14, 2026. https://signalfidelitygroup.com/research/the-disclosure-effect (checked October 19, 2026)
• CommsRx, Evidence. https://commsrx.com/evidence/ (checked October 19, 2026)
• Ahrefs, Louise Linehan, March 2, 2026, "Update: 38% of AI Overview Citations Pull From The Top 10." https://ahrefs.com/blog/ai-overview-citations-top-10/ (checked October 19, 2026)
• Chen, Wang, Chen and Koudas (University of Toronto), Generative Engine Optimization: How to Dominate AI Search, arXiv 2509.08919, preprint, September 2025. https://arxiv.org/html/2509.08919v1 (checked October 19, 2026)