ChatGPT
- A
- B
- C
- D
Possible related labels: B-U, C-U, D-U, Q. Not counted.
Research observation
We asked five AI assistants the same question. Here's what they said.
We entered the same hospital question in ChatGPT, Claude, Grok, Perplexity and Google Search, which returned an AI Overview. We kept the first response from each service. The names and suggested starting points differed.
The exact question
Which is the best hospital for heart care in Chennai?
All five responses were recorded on September 27, 2026, in India. We asked no follow-up questions and did not regenerate an answer. This records AI outputs; it does not recommend any hospital.
23 reference labels. Four appeared in most answers.
Across five answers, we recorded 23 name-and-location reference labels. Four appeared in at least three answers; 14 appeared in only one.
23recorded reference labels
4appeared in at least 3 of 5 answers
14appeared in exactly 1 answer
Some labels may refer to the same hospital. These are counts of recorded references, not verified unique hospitals. We kept different branches, unspecified locations and unresolved names separate. Possible matches are marked but never added to the totals.
Explicitly named
Possible match; not counted
Not explicitly named
| AI service | A | B | B-U | C | C-U | D | D-U | E | F | G-U | G-1 | G-2 | H | I | J | K-1 | K-2 | L | M | N | O | P | Q |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ChatGPT | |||||||||||||||||||||||
| Claude | |||||||||||||||||||||||
| Grok | |||||||||||||||||||||||
| Perplexity | |||||||||||||||||||||||
| Google AI Overviews |
Claude
- A
- C
- E
- B
- F
- G-U
- H
- D-U
- I
Possible related labels: B-U, C-U, D, J, Q. Not counted.
Grok
- A
- B
- C
- J
- D
- G-1
- G-2
- K-1
- L
- M
Possible related labels: B-U, C-U, D-U, E, K-2, Q. Not counted.
Perplexity
- A
- P
- F
- C
- K-2
- H
- E
Possible related labels: C-U, J, K-1, Q. Not counted.
Google AI Overviews
- Q
- N
- H
- G-2
- C-U
- O
- B-U
- P
Possible related labels: A, B, C. Not counted.
Response-by-response detail
| AI service | Reference count | Recorded labels | Suggested starting point |
|---|---|---|---|
| ChatGPT | 4 labels in four hospital rows | A, B, C, D | Hospital A for planned adult care. |
| Claude | 9 labels in seven initial bullets | A, C, E, B, F, G-U, H, D-U, I | No single overall choice; suggestions depended on the condition. |
| Grok | 10 hospital or branch labels across the complete answer | A, B, C, J, D, G-1, G-2, K-1, L, M | Hospital A for most adult care, with condition-specific alternatives. |
| Perplexity | 7 name labels in six recommendation entries | A, P, F, C, K-2, H, E | A combined Hospital P / Hospital A entry. |
| Google AI Overviews | 8 labels: seven cards plus one additional name in the lead | Q, N, H, G-2, C-U, O, B-U, P | Hospital Q and Hospital P together in the lead. |
Labels A and C appeared in four answers; B and H appeared in three. No two responses contained the same set of recorded labels. A mention is not a ranking, recommendation strength or proof of clinical quality.
How the labels and counts work
Letters replace hospital names. A–D retain the original ChatGPT coding; the remaining labels preserve additional recorded references. The letter order is not a ranking.
B-U and C-U keep Google's unresolved location details separate from B and C. D-U is Claude's branch-unspecified reference. G-U has no branch specified; G-1 and G-2 are explicit different branches. We do not assign G-U to either branch.
K-1 and K-2 preserve two name variants at the same stated locality. E and J preserve different names whose relationship was not verified. Q is an unqualified reference that may match A. We did not merge these labels.
Perplexity named A and P in one combined recommendation. Both names count as recorded labels; they do not count as two independent recommendations. Google's P appears in the opening sentence, outside its seven cards.
The totals include later named suggestions and exclude citation publishers, source lists, ordinary search results and repeated shorthand. “Not explicitly named” means that exact reference label was not recorded, not that the hospital was definitively absent under every possible alias.
Short excerpts, names withheld
Each excerpt comes from the recorded first response. Hospital labels replace names; the surrounding words are unchanged. These are separate extracts, not a reconstructed conversation.
“For planned adult heart care in Chennai, [Hospital A], would be a strong first choice.”
ChatGPT
“There's no single official "best." Rankings vary by who publishes them.”
Claude
“[Hospital A] is the strongest overall choice.”
Grok
“[Hospital P] / [Hospital A] is the strongest all-round recommendation for heart care in Chennai”
Perplexity · excerpt ends here
“[Hospital Q] and the [Hospital P] are widely considered the best and most comprehensive choices for heart care in Chennai.”
Google AI Overviews · location included in the redaction
How we recorded the answers
One question, one first response per tested service, one calendar day in India. Answers vary. The five captures were made on September 27, 2026; ChatGPT was recorded earlier that day, and the other four were recorded that morning.
- ChatGPT, 12:25 a.m. IST: signed in; Web Search explicitly selected; the interface displayed Pro and reported searching 48 websites. Existing personalization was unchanged.
- Claude, 10:03 a.m. IST: signed in; Opus 5.5 Medium; automatic web search was visibly used. Existing Memory and Manual permission settings were unchanged.
- Grok, 10:03 a.m. IST: signed in; Expert mode; automatic search was visibly used and the response displayed 73 sources. The activity also showed “Explored workspace”; we did not inspect underlying workspace content or alter account settings.
- Perplexity, 10:03 a.m. IST: signed out; Search mode selected; the response displayed Best, “Researched 2s” and 24 sources.
- Google AI Overviews, 10:01 a.m. IST: ordinary Google Search, not AI Mode; signed in. Google displayed “Results are not personalised,” an Erode IP location and Chennai result context. No search setting was changed.
Times are capture times, rounded down to the minute. Exact timestamps, verbatim text and screenshots are preserved privately. Account, model and search conditions differ, so this is a documented observation rather than a controlled performance benchmark.
The four standalone assistants are shown in their relative order in Similarweb’s August 2026 worldwide website popularity rankings, with Grok ahead of Perplexity in the domain rankings. Similarweb’s rank combines estimated visitors and pageviews; it is not total usage or healthcare usage. Google AI Overviews is shown separately because it runs within Search and is not comparable on that metric. Sources accessed September 27, 2026.
What this does—and does not—show
These responses show different hospital shortlists and different recommendation styles for the same question. They do not show which hospital is best, what patients usually see, or whether any patient followed an answer.
We did not independently verify the outputs’ medical claims, hospital services, rankings, ownership or branch aliases. We do not endorse those claims. Sources may overlap between services; five responses are not five independent assessments of clinical quality.
One response per service cannot establish repeatability, a trend, market share or the effect of GEO/AEO work. Accounts, location, personalization, models, timing and search access can change the answer. This is not medical advice or a TarsX client result. The fictional phone illustration and sample monthly report elsewhere on this site are separate from this recorded research.
Check the answer your hospital receives.
TarsX checks how hospitals appear in AI answers, identifies information gaps and gives their teams clear change requests, then verifies the updates.