AI answer engines get Indian exam facts right, except the ones that just changed

01 / 07Summary

We asked four AI answer engines 40 factual questions about 11 exams taken in India, three times each, and graded every answer against the exam body’s own published figure, checked the same day: 480 answers in all.

The engines were right far more often than their reputation suggests. Where they slipped, it was mostly on facts that had recently changed.

02 / 07Why this matters

Students now ask answer engines the questions they once took to a coaching centre or a results page: what the fee is, when the exam is, whether they are still eligible. A wrong date or an old eligibility rule can cost a candidate an attempt, and India’s exam bodies change their rules often. In 2026 alone, ICAI moved the CA Final from three sittings a year to two, ICSI rebuilt the CSEET, PMI replaced the PMP exam, and NTA cancelled and re-held NEET UG.¹

Claims that “AI gets facts wrong” rarely say how often, on which facts, or against what. We wanted a count, with the official source for every fact on the record.

The engines rarely invented exam facts. When they slipped, they repeated an answer that used to be true.

Study 002, Findings 1 and 2

03 / 07Method

We first established ground truth. For 11 exams we collected 60 candidate facts from the exam body’s own pages and notices, recorded the URL and a verbatim quote for each, and checked them on 5 October 2026.² We kept 40 that had exactly one correct answer, balanced across fees, dates, eligibility, exam pattern, attempts and result statistics, and dropped any where official pages disagreed with each other or the date was only tentative. Only then did we run the engines.

Each question named the exam and, where it mattered, the cycle (“CAT 2026”, “CMA Final June 2026”), as a careful student would. Each was asked three times on each engine, as a new conversation, from India; no prompt named the country.³

Answers

48040 questions × 4 engines × 3 runs

Official facts

40on 11 exams, each with URL and quote

Citations

3,4511,357 unique URLs

Engines

ChatGPT (search on), Perplexity, Gemini, Google AI Mode

Exams

CA, CMA, CS, CAT, IPMAT, JEE Main, NEET UG, CSIR UGC NET, IIT JAM, MHT CET, PMP

Fieldwork

5 October 2026, 17:02–17:40 IST; facts checked the same day

How answers were graded

We wrote the grading rubric before collecting any answer. Each answer was graded on the one key fact the question asked for: correct (the current official value), partly correct (the core fact right but a material part wrong, or the right value given alongside a conflicting one), outdated (a value that had been official in an earlier cycle), wrong (any other value, including saying a fact had not been announced when it had), or no answer. No answer fell in the last group.⁴

Every answer was read in full where the key fact was not in its opening lines. As a mechanical check, every answer graded correct was confirmed to contain the official value, and correct answers that also mentioned a known earlier value were re-read.

What counted as a citation

A citation is a URL the engine presented as a source, inline or in its sources list. We counted each URL once per answer after removing tracking parameters. An answer “cited the official site” if at least one citation pointed to the exam body’s own domain or its document host.⁵

Disclosure: some brands that appear in this study are clients of Citation Research. Their results were produced by the same questions, engines and rules as every other brand’s.

04 / 07Findings

1. Ninety-six per cent fully right

Of 480 answers, 461 stated the official fact correctly. Twelve were partly correct, five were outdated and two were wrong. Put another way, about one answer in 25 was not fully right, and one in 69 gave a wrong or out-of-date key fact. None declined to answer.

The engines differed. Google AI Mode was fully right on 119 of its 120 answers. Perplexity was not fully right on 9 of 120, the most of the four, and all three of its outdated answers were the same question. ChatGPT made the most outright errors (4 of 120), and in three of those four the sources it listed had nothing to do with the question, such as dictionary pages and programming documentation. Gemini made no outright errors but five partly correct answers.

2. The misses were mostly old facts

Twelve of the 19 answers that were not fully right carried a value that had once been official. Five gave it as the answer: four said NEET UG 2026 was held on 3 May, the date of the exam NTA cancelled, rather than the 21 June re-examination the results were based on; one gave IIM Indore’s 2025 birth-date cut-off for IPMAT 2026. Seven more were right on the main number but mixed in an old detail: the 200-question NEET paper used up to 2024, or PMI’s eight-year experience window, replaced by ten years in July 2026.⁶

The remaining seven misses were of two kinds. Five CSIR NET answers had the right age limit, 30, but gave the wrong reference date, hedged it, or left it out. Two ChatGPT answers said exam dates had “not yet been announced” when ICAI and IIT Kharagpur had already published them.

3. Fees were safe; eligibility was not

All 96 answers about fees were right, as were all 72 about result statistics, all 36 about attempts and all 12 about which IIT organises JAM 2027: 216 answers without a miss. Eligibility rules were the weakest area, with 50 of 60 answers fully right, followed by dates, with 78 of 84. Those two areas held 16 of the 19 misses. By exam, the misses sat in NEET UG, CSIR UGC NET and PMP; five exams had none.

4. Citing the official site did not make an answer right

Fewer than half the answers, 226 of 480 (47.1%), cited the exam body’s own site. Those answers were fully right 94.7% of the time (214 of 226); answers that did not cite it were right 97.2% of the time (247 of 254). Five of the seven wrong or outdated answers listed the official site among their sources. All three of Perplexity’s outdated NEET answers, for example, listed NTA’s NEET site while giving the date of the cancelled exam.⁷

The official site was a minority source overall. Of 3,451 citations, 19.7% pointed to the exam body’s own domain. Education portals and aggregators took 33.9%, and one of them, careers360.com, accounted for 15.0% of all citations on its own. The engines leaned on official sites to very different degrees: 37.2% of ChatGPT’s citations, against 4.3% of Perplexity’s.

5. The same slip, run after run

Asking again did not shake out the errors. Of the 160 engine–question pairs, 149 were fully right on all three runs and 7 were mixed. Four pairs gave the same miss every time, and those four produced 12 of the 19 misses. ChatGPT’s four misses were all one-offs, each in a different question.

Asking again did not fix it. Most misses came back on every run.

Study 002, Finding 5

A side test: dropping the year

Our questions named the cycle, which is how a careful student asks. In a small side test we put five of the questions to each engine once without the year, as a hurried student might.⁸ Two of the 20 answers were outdated: one gave the PMP’s old 230-minute length and one described the 200-question NEET paper. None of the 60 answers to the same five questions with the year was outdated. Twenty answers are too few to generalise from, but the direction matches the main result.

05 / 07What it means

For exam bodies, coaching institutes and education publishers, the risk is less that engines make facts up and more that they keep repeating facts that have changed.

1. Announce a change on its own page, with a date.The misses clustered on rules that had just changed. A dated page that says what changed and from when gives an engine a current source to find.
2. Mark what is cancelled or replaced.Four answers repeated a cancelled NEET date. Pages that still carry an old figure should say plainly that it no longer applies.
3. Put the qualifier next to the number.Five CSIR answers had the age limit right but its reference date wrong or missing. Keep conditions such as “as on 1 July 2026” in the same sentence as the figure.
4. Do not read a citation as an endorsement.Five of seven errors cited the official site. Check what the engine says, not only whether it links to you.
5. Test with repeats and with vague wording.Misses came back run after run, and the side test hints that undated questions fare worse. One screenshot proves little.

06 / 07Limitations

• Careful questions. Each question named the exam and the cycle and asked for one fact. Students who ask more loosely may get worse answers; our side test of 20 answers points that way but is too small to measure it.
• Verifiable facts only. We kept only facts with one clear official answer and dropped those where official pages conflicted. Engines may do worse on exactly the facts we had to leave out.
• One day, one market. All answers were collected on 5 October 2026 from India, with default settings. Engines change without notice, and results on another day may differ.
• Grading is judgement. One researcher graded every answer; a second graded 15% blind. Agreement was high, but partly correct and outdated are calls a different coder could make differently at the margin.
• Citations are not reasoning. We see which pages an answer listed, not which one the engine used for the fact. Perplexity’s list includes the search results it showed, not only pages it quoted.
• Facts move. Several answers depend on the latest published result or schedule. They were right or wrong on 5 October 2026; some will change as new results appear.

07 / 07Data and reproducibility

The questions with their official sources, every graded answer with its full text, and every citation are published under CC BY 4.0. Version 1.0 corresponds to this publication; corrections will be issued as new versions with a changelog.

• questions.csv40 rowsOfficial answer, URL, quote, date checked
• answers-graded.csv480 rowsGrade, grader note, full answer text
• citations.csv3,451 rowsOne row per cited URL per answer
• domain-types.csv405 rowsSource type for every cited domain
• second-pass.csv72 rowsBlind second grading
• side-test-no-year.csv20 rowsSide test without the year

For questions about the data, or a reanalysis you would like us to check, write to research@citationresearch.com.

@techreport{citationresearch2026study002,
  author      = {{Citation Research}},
  title       = {AI answer engines get Indian exam facts right, except the ones that just changed},
  institution = {Citation Research},
  number      = {Study 002},
  year        = {2026},
  url         = {https://citationresearch.com/research/ai-exam-facts-accuracy-india}
}