Ask AI for an expert and it names a person

01 / 08Summary

When people ask an AI engine for an expert, the answer can name a person or send them to a platform. On 5 October 2026 we asked ChatGPT (search on), Perplexity, Gemini and Google AI Mode for 30 kinds of expert, from a CA teacher to a wedding planner, in two phrasings each: 240 answers from India and 80 more from the United States for ten of the roles. We coded every person and organisation each answer put forward.

The answers name people far more often than we expected. They rarely name the same people.

02 / 08Why this matters

Coaches, teachers, doctors and trainers sell themselves, not a product. For them the question is whether an answer engine will say their name, or send the asker to a marketplace, an app or a large institute instead. Earlier studies of AI recommendations, ours included, have counted brands. This one counts people.

It also asks what the people who do get named have in common in public: a Wikipedia page, a book, a YouTube channel, a TEDx talk, press coverage, a website, a LinkedIn profile. We report how common each signal is. We do not claim that any of them causes a recommendation.

Ask four engines for the same expert and you mostly get four different lists of people.

Study 010, Finding 3

03 / 08Method

We chose 30 roles in which people hire an individual expert, grouped as coaching (10), teaching (8), money (2), health (6), sport (2) and home and events (2).¹ Each role was asked in two ways: “Who are the best [role]s in India?” and a one-line need followed by “Can you recommend a good [role] in India?”.² Ten of the roles were asked again from the United States, with the country removed from the prompt. Each question went once to each engine, as a new conversation, with the location set at the network level.³

Answers

320240 from India, 80 from the United States; no failures

Questions

8030 roles × 2 phrasings in India; 10 roles × 2 in the US

Names coded

2,6821,323 people and 1,359 organisations

Engines

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

Markets

India and the United States, set at the network level; one run per question and engine

Fieldwork

5 October 2026

What counted

A person is a named individual the answer puts forward as someone to learn from, hire, consult or follow: a full name, a name with an honorific (“Dr. Shah”, “Vignesh Sir”), or a first name presented as a specific tutor. People mentioned only as students, clients, authors of a source or historical references were not counted, nor were names that appear only inside a business name. An organisation is a named company, academy, clinic, hospital, app, marketplace, programme or club offered as a place to find the service. Alma maters, certifying bodies and general platforms named only as a medium (“his YouTube channel”) were not counted. Each name counts once per answer.⁴

Answers were coded against a written rubric by AI research assistants and checked by the lead researcher, who coded 40 answers directly and re-coded a random 48 (15%) without seeing the first coding. The two codings agreed on whether a person was named in 47 of 48 answers and shared 207 of the 210 names.⁵ Spelling variants of the same person were merged by hand (18 merges); a lone first name was treated as a different person in each role.

For the public-footprint analysis we took the 61 people named in four or more of the 320 answers and checked seven signals in public sources on 5 October 2026.⁶ We report counts only and do not name the people.

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. Most answers name a person

201 of the 240 India answers (84%) named at least one individual. Three in four (183) named people and organisations side by side; 18 named only people and 39 only organisations. None named neither. Across all India answers, people made up 1,027 of the 2,100 names (49%): the engines named people about as often as they named organisations.

2. The phrasing decides how many people you get

“Who are the best public speaking coaches in India?” and “I get nervous presenting at work… can you recommend a good public speaking coach in India?” ask for the same thing. They did not get the same answer. The “who” questions named a person in 111 of 120 India answers (92%) and averaged 6.2 people each. The need-first questions named a person in 90 of 120 (75%) and averaged 2.3; more of their names were academies, marketplaces and apps. From the United States the gap was wider: 35 of 40 against 20 of 40.

3. Some roles are people, some are platforms

The share of names that were people ran from 93% for CA faculty to 19% for music teachers. People were the majority of names in 11 of the 30 roles, among them CA faculty, life and executive coaches, LinkedIn and personal branding coaches, plastic surgeons and nutritionists. Music, spoken English, weddings, PMP training and career counselling were mostly organisations: online academies, institutes, planners’ firms and tutor marketplaces. Money questions did not go to apps. Both finance roles named a person in all 16 answers, and half or more of their names were people.

4. The engines disagree about who the experts are

The 240 India answers named 602 different people. 409 of them (68%) appeared in only one answer, and 430 (71%) were named by only one engine for that role. Only 26 people, 4%, were named by all four engines. In 17 of the 30 roles no person was named by all four.⁷ The most-named person appeared in 13 of the 240 answers.

5. Engines differ in how often they name people

Google AI Mode named a person in 58 of its 60 India answers and averaged 5.8 people each. Perplexity named a person in 38 of 60 and averaged 2.5. Perplexity named only organisations in 22 of its 60 answers, against 2 to 9 for the other engines. Every engine also had its own people: of the people ChatGPT named, 59% were named by no other engine for the same role.

6. Without India in the question, fewer people

For the ten roles asked in both markets, 72 of 80 India answers (90%) named a person against 55 of 80 United States answers (69%). The US answers averaged 3.3 people against 4.6, and 48% of their names were people against 58%. The biggest gap was interview coaching: from the US the engines named a person in 2 of 8 answers and pointed to coaching platforms instead.

7. The most-named people are on YouTube, not Wikipedia

61 people were named in four or more of the 320 answers. Nearly all had a LinkedIn profile we could find (54). Two in three had a YouTube channel (40) and a website of their own (40); 14 had channels with more than a million subscribers. A book was common (35), press coverage less so (27), a TED or TEDx talk rarer still (16). Only 12 had a Wikipedia article.

The gap shows most clearly when the people are split by where they were named. Of the 48 named only in India answers, 3 had a Wikipedia article and 32 a YouTube channel. Of the 13 named in United States answers, 9 had a Wikipedia article. For 25 of the 40 people with their own website, that site was among the sources of at least one answer that named them.

05 / 08What it means

For a coach, teacher or practitioner, an AI answer is a place where a name can appear. These results describe where that happened; they do not show why.

1. Check both kinds of question.A “who are the best” question and a question that starts with a problem gave different answers. Test the way your clients actually ask, not only the way you would.
2. Check every engine.Seven in ten people named were named by one engine only. Being named by ChatGPT says little about Perplexity or Google AI Mode.
3. Know which kind of role you are in.Where the answers lean to platforms, as for music, spoken English and PMP training, the answers pointed to academies and marketplaces more often than to individuals.
4. Expect to be listed next to organisations.Three in four India answers named people and organisations together. A practitioner competes for the same short list as academies, clinics and marketplaces.
5. Look at the public record the engines can read.The most-named people mostly had a YouTube channel and a website of their own; few in India had a Wikipedia article. We cannot say which of these, if any, led to a recommendation, but they describe the people who were recommended.

06 / 08Limitations

• One run, one day. Each question was asked once per engine on 5 October 2026. Answers change between runs; a second run would name some different people.
• Small cells. Each role has 8 India answers, and each role in the US comparison 8 more. Role-level results are indications, not estimates.
• Coding is judgement. Deciding whether a name is put forward as an option, and whether two spellings are one person, is a judgement. Our checks agreed closely, but organisation counts are softer than person counts: named programmes and sub-brands were hard to code consistently.
• Footprint checks can miss things. Signals were found through search, so a “not found” can be a miss. Our first pass missed four TED or TEDx talks that a second search found; press and LinkedIn are probably undercounted for people with common names.
• Footprint is descriptive. We checked public signals for the 61 most-named people only, not for people who were never named, so we cannot say whether those signals make a recommendation more likely.
• Two markets, English only. India answers were asked in English with India in the prompt; the US prompts named no country. Other languages and markets may differ.
• Engine output varies by surface. Google AI Mode’s text export sometimes dropped the title of a recommendation card. We checked every AI Mode answer against its full page; restoring missing titles changed the coding of 5 of its 80 answers.⁸

07 / 08Data and reproducibility

The data files below are published under CC BY 4.0. They hold counts and anonymous ids, not the names of the people in the answers. Version 1.0 corresponds to this publication; corrections will be issued as new versions with a changelog.

• questions.csv80 rowsPrompts
• answers.csv320 rowsPer answer
• roles-india.csv30 rowsPer role
• footprint-top-named.csv61 rowsAnonymised

For the answer texts, the coding files, or a reanalysis you would like us to check, write to research@citationresearch.com.

@techreport{citationresearch2026study010,
  author      = {{Citation Research}},
  title       = {Ask AI for an expert and it names a person},
  institution = {Citation Research},
  number      = {Study 010},
  year        = {2026},
  url         = {https://citationresearch.com/research/ai-expert-recommendations-people-or-platforms}
}