What people ask AI before choosing an accountant: a published prompt study in India, and the count we do not yet have

The honest form of this answer is a fraction: X of N observed questions fell into the most frequent category, collected between two named dates. We do not have that fraction, because we have not collected a prompt dataset from people choosing an accountant in India, and we will not print a category share without a denominator behind it. What this page publishes instead is everything needed to produce it: the sixty question set we would test, the five lawful ways to collect real prompts, the unit of analysis, the codebook, the two coder rule, and the reporting form. It also states what is already known from our own dated work, which is about answers rather than about questions, and keeps the two apart.

The answer, first

From the shape of the work rather than from a count, people choosing an accountant are asking six kinds of question: what the deadline is and what the penalty is for missing it, whether they need a chartered accountant at all or an online service will do, what it costs, how a specific filing works, how to pay less tax legally, and how to tell whether the person in front of them is real. That ordering is our judgement and it is labelled as judgement, not as frequency.

The six categories below carry ten questions each, so the list can be used as a checklist today while the frequency work is still undone. That is the trade this page makes: a usable enumeration now, and no invented share ever.

How it was measured

The unit of analysis. One question, asked once, by one person, before choosing an accountant. Not a keyword, not a topic, not a search volume. Fix this before anything else, because every disagreement later traces back to it.

Five lawful collection sources, and you name the ones you used.

• Your own enquiry channels: the contact form, the email inbox, the messages people send on a business messaging number, with the personal details removed before coding.
• Your own first meeting notes, recorded with the person's consent, capturing the questions they asked before engaging.
• Your own site search log and support queue, which shows the questions people type when they think nobody is watching.
• Your Search Console query report, kept in a separate bucket because a search query and a question typed to an assistant are different behaviours and should never be pooled.
• A recruited panel of small business owners and salaried people who agree to paste the questions they typed into an assistant, with consent recorded and the collection window fixed in advance.

What is not a dataset, stated so nobody mistakes this page for one. A list of questions an assistant produced when asked what people ask. Purchased prompt data with no stated origin. Scraped conversations. A competitor's chart with no method. Our own sixty questions below are in the first of those categories: they are a plausible checklist, and they are evidence of nothing about frequency.

The codebook: the six categories below, plus a seventh for anything that fits none of them. A question that spans two categories is coded where the asker would put it, with the second category recorded in a notes field. Publish the codebook before the counts.

Two coders, independent, and the disagreement count published. A third person resolves disagreements. If you will not publish how often your coders disagreed, a reader is right to assume it was often.

The reporting form, which is the whole point. Never "the most common question is about cost". Always the category name, the count of questions coded to it, the total number of questions coded, the collection window as two dates, and the sources used. Then publish the coded rows so anybody can recompute it. State exclusions and how many were excluded.

How to test the questions against the assistants afterwards. Once you have real questions, run them blind, with no firm named, on three assistants with three repetitions, and score whether any firm is named, whether a firm's own domain is cited, and whether the answer points at an official portal instead. Log the visible source list, and separately ask the assistant how many questions it searched for, storing that as a claim. In three separate batches of our September 2026 audit the assistant withdrew its own claim to have searched, saying it could not honestly substantiate it.

The sixty question set, in six categories of ten

Deadlines and penalties.

• What is the last date for filing this year's return.
• What is the penalty if I file late.
• What happens if I have not filed for two years.
• When are GST returns due, and which ones apply to me.
• What is the due date for advance tax.
• What is the deadline for a company's annual filing.
• Is there any relief if I missed a deadline for a genuine reason.
• Can I revise a return after filing it, and until when.
• What notices can I get, and how long do I have to reply.
• Do I have to file if my income is below the limit.

Do I need a chartered accountant at all.

• Can I file my own return, and when is that a bad idea.
• What is the difference between a chartered accountant and a tax consultant.
• Is an online filing service enough for my situation.
• Who is allowed to sign an audit report.
• Do I need a chartered accountant to incorporate a company.
• Is a bookkeeper enough if I already use accounting software.
• At what turnover do I need an audit.
• Can the same person do my books and my audit.
• Do I need someone local, or is remote fine.
• What can a company secretary do that a chartered accountant cannot.

Cost.

• What does it cost to get a return filed for a salaried person.
• What does monthly GST filing cost for a small business.
• What does a statutory audit cost for a company of my size.
• What does incorporation cost in total, including government fees.
• Is it charged monthly, per filing or annually.
• What is charged extra beyond the professional fee.
• What does it cost to fix two years of neglected books.
• Why do two quotations for the same work differ so much.
• Is there a cheaper option for a business that has not started earning.
• What will this cost me in total in the first year.

How a specific filing works.

• How do I register for GST, and what documents do I need.
• How do I claim input tax credit correctly.
• How do I deduct and deposit TDS, and by when.
• How do I close a company that has no business.
• How do I change my registered office address.
• How do I add a director or a partner.
• How do I report foreign income.
• How do I account for income from freelancing abroad.
• How do I handle cash sales properly.
• What records do I have to keep, and for how long.

Paying less tax, legally.

• Which deductions apply to someone like me.
• Should I choose the old regime or the new one.
• Is it better to take salary or dividend from my own company.
• Should my business be a proprietorship, an LLP or a private limited company.
• What expenses can I legitimately claim.
• Is the presumptive scheme suitable for me.
• How do I plan for capital gains on a property sale.
• What are the tax effects of paying family members.
• Which of these ideas will get me a notice.
• What should I be doing before the year ends rather than after.

Trust and verification.

• How do I check that someone is really a chartered accountant.
• What is their membership number and can I verify it.
• Who will actually do my work, the partner or an article assistant.
• Will I get a written engagement letter.
• Who holds my login credentials for the portals.
• What happens to my data and documents if I change accountants.
• How quickly do they reply when there is a notice.
• Have they handled a business like mine before.
• Can I speak to a current client.
• What went wrong on their last difficult engagement.

What the numbers were

Prompts collected: none, as of 29 September 2026. There is no N on this page, therefore no X, therefore no most frequent category and no share. The six categories are printed in our judgement of urgency and that is stated wherever they appear.

60. Six categories of ten, enumerated above. That is the count of a checklist we wrote, and it is the only count this page can honestly lead with.

The denominators we do hold, kept separate on purpose. In the September 2026 audit of our own domain, aiknowsus.com, we recorded 24 batches and 72 conversations, in which the phrase recording that we were not cited appears 161 times in the assistant's own audits of its answers, and in one batch of six questions the assistant reported that it had not cited or recommended us in any of the six. In the clawlaw.in programme we hold a reading of 78 questions on ChatGPT on 27 July 2026 with no live search behind any of them, eighteen blind commercial questions on 18 August 2026 with the company the top source on exactly one, and six questions on Claude on 17 September 2026. Every one of those is a count of answers we collected. None is a count of questions a buyer asked. Mixing the two would produce a frequency that looks real, which is the exact failure this page exists to avoid.

One dated result that tells you where these questions end up. On 17 September 2026, on questions about how to use an official portal, Claude placed official portals above every commercial product and said that no commercial product should rank above the official portal for a question about using it. Four of the six categories above are procedure questions, so expect the official answer to lead, and write the page that explains the procedure and links the portal rather than one that competes with it.

What this cannot tell you

• It cannot tell you what anybody asks. Sixty plausible questions is a checklist, not evidence.
• It cannot rank the six categories. Ten per category is for evenness.
• It cannot capture how people really type. Real prompts are shorter, misspelt, mixed between English and another language, and often about one specific notice rather than about a category.
• It cannot be compared with a published prompt study that does not state its denominator, its codebook and its coder agreement. Most do not.
• It is not tax or compliance advice, and nothing in the question list should be read as an answer.

Sources and change log

Sources. The sixty questions are our own. The dated counts quoted come from the aiknowsus.com audit of September 2026, captured to geo-audits/aiknowsus-com/, and from the clawlaw.in programme of July to September 2026, recorded in Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md, Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md and Tier_1/claude_response_17_09_audit.md. For professional conduct questions, read the Chartered Accountants Act 1949 and the ICAI Code of Ethics at the Institute's own site and record your date.

Change log. 29 September 2026: first published with the sixty question set, the collection protocol, the codebook rules and the statement that no prompts have been collected. If a collection is run, this section will carry N, the X for each category, the collection window, the sources used, the exclusions and the coder disagreement count, and this entry will stay above it.

Common questions

Why publish a prompt study with no prompts in it?

Because the method is the part worth copying and the number is the part that would be invented. A firm that runs this properly on its own enquiries will get a better answer for its own market than any published study could give it.

Can we use our Search Console queries as the dataset?

Use them, and keep them in their own bucket. A typed search and a question asked to an assistant are different behaviours, and pooling them produces a number that describes neither. Report both, separately labelled.

How many prompts would make this worth publishing?

Enough that you are willing to print the denominator next to every count. A study of 120 coded questions with the method published is more useful than one of 5,000 with the method hidden, and the honest way to publish a small one is to call it small.

Is it legal to collect our clients' questions like this?

Collect from your own channels, strip personal details before coding, take consent for anything recorded, and do not collect from anywhere you do not own. India's data protection law applies to the personal data you hold, so keep the coded dataset free of identifiers and say so in your method.

What should we write first, without any of this data?

Take the ten questions in the deadlines category and the ten in the cost category and answer them on your own site, one page each, with the answer in the first paragraph, the thresholds and dates stated, and the official portal linked. Those twenty pages are useful whatever the frequency study eventually says.

Will an assistant even name an accounting firm for these questions?

Often it will not, and that is worth measuring rather than assuming. On our own domain in September 2026 the assistant reported that it had not cited or recommended us in any of six questions about our own category, which is the baseline most firms are starting from.

What to do first

Open your enquiry inbox, take the last hundred messages, and write down the question each one was really asking. That is your first dataset, it costs you an afternoon, and it will beat this page for your own firm. Then answer the ten most frequent, one page each, and keep the count with its date so your second hundred can be compared with your first.