AI engines agree on the top coaching pick for only 5 of 20 Indian exams

01 / 08Summary

A student who asks an AI answer engine which coaching to join gets a short list of names and a set of sources. We wanted to know whose names, and which sources. On 5 October 2026 we asked ChatGPT (search on), Perplexity, Gemini and Google AI Mode 100 coaching questions, five for each of 20 Indian exams, twice each, from India. The 800 answers named providers 4,021 times and cited 8,605 sources.

Five results stand out. The first is how far one brand reaches; the rest are about how little the engines agree.

02 / 08Why this matters

Choosing a coaching institute is one of the larger spending decisions a student and family make before an exam. The question used to go to seniors, forums and a search box. It now also goes to answer engines, which reply with a shortlist and reasons. A provider that is not on the shortlist has to be found some other way.

Coaching is also a category where the sources are contested: providers publish their own guides and rankings, education portals sell listings, and students argue on forums. We wanted to see which of these the engines lean on, with a sample large enough to say so with numbers.¹

Four engines, twenty exams: the same top pick five times.

Study 001, Finding 2

03 / 08Method

We wrote five questions for each of 20 exams in the words students use: “best coaching for…”, “which online course is best for…”, “top institutes for…”, “is it worth joining coaching or should I self-study, and which coaching if so”, and one question for a group such as droppers or working professionals.² Each question went to four engines, twice, from a network location in India, in a new session each time.³

Questions

10020 exams × 5 phrasings

Answers

800100 questions × 4 engines × 2 runs

Citations

8,6053,841 unique URLs on 798 domains

Engines

ChatGPT (search on), Perplexity, Gemini, Google AI Mode; default models

Market

India, set at the network level

Fieldwork

5 October 2026, 11:12–12:21 UTC

What we counted

A mention is a provider’s name in the answer text, counted once per answer. Before matching we removed the source labels engines attach to sentences, because a label names the source, not a recommendation.⁴ A citation is a URL the engine presented as a source, inline or in its source list, counted once per answer after canonicalisation.⁵

Providers were matched with an alias dictionary built from the answers and checked by hand. Sub-brands were merged into their parent where the answers presented them that way (for example, CA Wallah and PW OnlyIAS into Physics Wallah).⁶ In a random sample of 40 answers the dictionary found 209 of the 223 provider mentions a reader would count (94%) and made one false match; we then added the 13 names it missed.

Each cited domain was coded as a provider site, a portal or listing (Shiksha, Collegedunia, Justdial, Google Maps profiles and similar), a list or guide site, a forum, video, social, an official body, news or other. Domains coded by hand carry 98% of all citations; the rest were coded by rule. Individual teachers named in answers are counted separately and reported only in aggregate.

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 / 08Findings

1. One brand reaches almost every exam

Physics Wallah was named in 339 of 800 answers (42%), for 19 of the 20 exams. Only PMP answers left it out. It was named in 39 of 40 NEET UG answers and 38 of 40 JEE Main answers, but also in 26 of 40 CLAT answers, 25 of 40 GATE CSE answers and 16 of 40 NDA answers. Its own site, pw.live, was the most-cited domain in the study: 470 citations in 241 answers.

Multi-exam platforms such as Physics Wallah, Unacademy, Testbook and Adda247 took 21% of the 4,021 mentions; specialist institutes took 72% and global course platforms 6%. More than half of answers (445 of 800, 56%) named at least one multi-exam platform, and a platform was the most-named provider in 6 of the 20 exams.

2. Each exam has its own small club

The answers named 370 different providers in total, and 99 of them only once. Within an exam the field is narrow. The median exam had 25 named providers, and in 16 of the 20 exams the five most-named took at least half of all mentions. NEET UG was the most concentrated (70%); CMA Intermediate (32%) and MHT CET (35%) were the least, with long tails of local institutes.

Commerce exams behave differently. Answers for CA, CMA and CS often recommended individual teachers by subject rather than an institute, which spreads mentions thinly across many names. We report teachers only in aggregate.

3. Four engines, four shortlists

For each exam we took each engine’s most-named provider across its 10 answers. All four engines shared a top pick for 5 of the 20 exams: JEE Main, NEET UG, UPSC CSE, CLAT and PMP. At least three shared one for 13 exams.⁷ Beyond the top pick, agreement was thin: of the 560 exam–provider pairs, 255 (46%) were named by one engine only and 117 (21%) by all four.

Perplexity stood apart on the “coaching or self-study” question. It named no provider in 35 of its 40 answers to that question, even though each one asked which coaching to join. The other three engines named none in 7 of their 120.

4. Ask again, get a different list

We compared the two runs of each question on each engine: 400 pairs. Only 80 (20%) named exactly the same providers. In 127 pairs (32%) fewer than half the providers were shared. The provider named first was more stable: it was the same in 238 of 374 pairs where both runs named someone (64%). Perplexity repeated itself most (36 of 100 pairs identical) and Google AI Mode least (6 of 100).

Ask twice and, four times in five, you get a different shortlist.

Study 001, Finding 3

5. Providers’ own sites are the most-cited sources

Coaching providers’ own websites took 4,517 of the 8,605 citations (52%). Education portals and listings took 14% and list and guide sites 15%: 28% together. Every engine cited provider sites more than any other type, but the mix behind them differed. Gemini cited the fewest sources (3.2 per answer) and 73% of them were provider sites. Perplexity leaned most on portals and listings (23%). ChatGPT was the only engine to cite official bodies much (14%), such as the ICAI, PMI and exam conducting bodies.

Google AI Mode behaved most like a local search page. 376 of its 2,243 citations (17%) were Google Maps business profiles, in 26 answers, and it was the only engine to cite Quora (181 citations). Some answers were tailored to the city of the network location, listing branches in Bengaluru, Gurugram or Chandigarh that the student never asked about.⁸

The cited provider pages were mostly about courses. In a hand-coded random sample of 100 provider-site citations, 65 were course, batch or pricing pages, 15 were advice articles, 13 were pages ranking or comparing several providers and 7 were homepages.

Being cited and being named went together. When an answer cited a provider’s own site, it named that provider in 1,560 of 2,138 cases (73%). In answers to the same exams that did not cite the site, the provider was named in 17% of 9,302 cases. This is an association: an engine may look up a provider it has already chosen, or choose one it has found.⁹

05 / 08What it means for providers

The answers are not one list. They are four engines’ lists, and each changes between asks. That shapes how a provider should read its own position.

1. Check every engine, not one.For 15 of 20 exams the engines did not share a top pick, and almost half of the providers named for an exam appeared on one engine only.
2. Measure with repeats.A single answer is an anecdote: four in five repeat pairs changed. Two runs per question is a floor, not a standard.
3. Keep course pages current and readable.Provider sites were the largest source type, and most cited provider pages were course, batch and fee pages. Those pages are what the engines read.
4. Do not ignore portals and maps.Perplexity drew 23% of its citations from education portals and listings, and Google AI Mode 17% from Google Maps profiles.

06 / 08Limitations

• One day, two runs. All 800 answers were collected on 5 October 2026. Engines change without notice; a later sample may differ.
• Location we did not choose. Requests came from India, but the city varied with the network route, and some answers were tailored to it. A student in another city may see other local names.
• A mention is not an endorsement. We counted names, not tone. A provider named as one option among several, or with a caveat, counts the same as a first recommendation.
• Matching is not perfect. The dictionary found 94% of provider mentions in our hand check before we added the names it missed. Small local institutes written in plain text are the most likely to be missed.
• English questions only. Students also ask in Hindi and Hinglish; those answers may name different providers.
• We observe, not explain. The link between being cited and being named is an association. We cannot see why an engine chose a source, or whether a page was paid for.

07 / 08Data and reproducibility

The question list, the answer log with the providers named in each answer, and every coded citation are published below under CC BY 4.0. Version 1.0 corresponds to this publication; corrections will be issued as new versions with a changelog. Raw answer text is available on request: it names individual teachers, whom we report only in aggregate.

• questions.csv100 rowsCSV
• answers.csv800 rowsCSV
• citations.csv8,605 rowsCSV
• providers.csv370 rowsCSV
• domain-types.csv798 rowsCSV
• exam-summary.csv20 rowsCSV

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

@techreport{citationresearch2026study001,
  author      = {{Citation Research}},
  title       = {AI engines agree on the top coaching pick for only 5 of 20 Indian exams},
  institution = {Citation Research},
  number      = {Study 001},
  year        = {2026},
  url         = {https://citationresearch.com/research/ai-coaching-recommendations-india}
}