Questions employers and candidates ask before choosing a recruitment agency: the interview protocol, and why N is still zero

The honest form of this answer is a fraction reported separately for each population: the most frequent employer question, stated as X of N employers, with candidate answers reported in their own fraction and never pooled with it. We have conducted no interviews, so N is zero as of 29 September 2026 and no fraction appears on this page. What does appear is the whole study: the sampling frame, the consent and recruitment method, the interview script for each population, the coding rules, the two coder check, and the reporting form. The thirty five questions below, twenty from employers and fifteen from candidates, are a checklist we wrote, and they are labelled as a checklist wherever they appear.

The answer, first

Two rules make a study like this worth reading, and both are usually broken.

Never pool the two populations. An employer is buying a service and worrying about cost, replacement and time to fill. A candidate is not paying and is worrying about whether the vacancy is real and whether their details will be shared without permission. One combined "most asked question" is a number that describes neither group.

Never report a frequency without N. "Employers most often ask about fees" is unfalsifiable. A named question, with the count of employers who raised it, the total number of employers interviewed, the two dates between which they were interviewed and the frame they were recruited from, can be argued with, which is what makes it worth publishing. The clearest statement we hold about over claiming from a thin sample came from a machine: in batch after batch of our September 2026 audit, the assistant declined to name a competitor as winning most often, saying its own previous answers had not produced a comparable live tested result to support such a claim.

How it was measured

Step one: define the two populations and the frame. Employers means the person who signs off a hire, named by role, in companies of a stated size range, in stated cities, who have used or considered an agency in the last twelve months. Candidates means people who have looked for a role in the same cities in the same period. Write the inclusion rules before recruiting anybody, and publish how many you approached, how many agreed and how many withdrew.

Step two: recruit without buying the answer. Approach through professional networks, your own past enquiries with permission, and industry groups. Do not recruit only your own clients, because your clients have already chosen you and will tell you what you want to hear. If you offer an incentive, state what it was.

Step three: consent, recorded. Explain what will be published, that it will be counts rather than quotes unless they agree to quotation, and that their employer will not be named. Keep the consent record separate from the coded data. India's data protection law applies to what you collect, so hold no identifiers in the dataset you analyse.

Step four: a semi structured script, thirty minutes, with a fixed opening. The opening question is the same for everybody and is deliberately open: "Walk me through the last time you chose, or decided against, a recruitment agency." Let them answer for several minutes before prompting anything. Only then work through the prompts, and record which questions came up unprompted, because that distinction is the whole value of an interview over a survey.

Step five: code in two passes. First pass is open coding of what people actually said. Second pass maps those codes onto a published codebook, with an "other" code that must never be the largest category. Count one question once per person, however many times they raised it.

Step six: two coders, and publish the disagreement count. Two people code the same transcripts independently, a third resolves disagreements, and the number of disagreements is published beside the results.

Step seven: report in the only defensible form. For each population separately: the question, the count who raised it, N, and the interview window as two dates. Publish the script, the codebook and the anonymised coded rows, and state how many people were approached and refused.

Step eight: state what unprompted means. Report two numbers per question where you can: raised unprompted, and raised after prompting. The gap between them is the most useful thing in the dataset and almost nobody publishes it.

The question sets: twenty employer, fifteen candidate

Employers, twenty questions.

• What do you charge, and is it a percentage of salary or a fixed fee.
• When is the invoice raised, on offer or on joining.
• What is the replacement guarantee if the person leaves in three months.
• Is there any refund if we do not hire anybody.
• How long will this take, realistically, for this role in this city.
• How many candidates will we see, and how are they screened.
• Have you filled this exact role before, and where.
• Who will work on our mandate, and how many other mandates do they carry.
• Is this exclusive, and what changes if it is not.
• Do you approach people already working at our competitors.
• Will you represent our company accurately to candidates.
• What happens if we hire a candidate you sent, for a different role.
• Do you verify qualifications and employment history, and how.
• What is your offer to join drop rate.
• Will you handle salary negotiation, and on whose behalf.
• Do you place contract staff as well, and on whose payroll.
• What do you need from us to start, and how fast can you start.
• Where did your last three placements for a company like ours come from.
• What happens to our candidate data after the mandate closes.
• Which mandates do you turn down.

Candidates, fifteen questions.

• Is this vacancy real and currently open.
• Who is the employer, and when will you tell me.
• Will you send my details anywhere without asking me first.
• Do I have to pay you anything, at any stage.
• What is the salary range for this role, honestly.
• How many rounds are there and how long will it take.
• Have you placed people at this company before.
• Will you tell me why I was rejected.
• Will you help me prepare, or only forward my CV.
• Is this a permanent role or a contract through you.
• If it is a contract, who employs me and who pays me.
• Will you keep my search confidential from my current employer.
• How long will you keep my CV on file, and can I have it removed.
• Do you handle overseas roles, and are you licensed for that.
• What should I do if the employer's offer differs from what you told me.

Note the asymmetry, because it is the finding a pooled study would destroy. Employers ask mostly about cost, speed and risk. Candidates ask mostly about truthfulness, consent and payment. Only two of the thirty five appear in both lists in substance.

What the numbers were

Interviews conducted: zero, as of 29 September 2026. N is zero for both populations. There is no most frequent employer question, no share, and no interview window, and there will be none on this page until there are transcripts behind them.

35. Twenty employer questions and fifteen candidate questions, enumerated above. That is a count of a checklist, complete as printed, and it is the only count this page can lead with.

The counts we do hold, and why they are not this. 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, ChatGPT confirmed on 27 July 2026 that it had run no live web search for any of 78 blind questions, and on 18 August 2026 that company was the top source on exactly one of eighteen blind commercial questions. Those are counts of machine answers. This page would be a count of human answers. They are different populations and this page keeps them apart rather than borrowing a number across the gap.

What this cannot tell you

• It cannot tell you what employers or candidates ask. Thirty five plausible questions is a checklist and nothing more.
• It cannot rank the questions. The order is our judgement of what decides a mandate.
• Interviews cannot measure frequency in a population. Even when run properly, forty interviews tell you what forty people said, and the honest report says so in the same sentence as the count.
• It cannot avoid the interviewer's effect. People tell an agency's interviewer different things than they tell a neutral one, which is why the recruitment method has to be published.
• It cannot be compared with a survey. An interview study and a questionnaire produce different data, and a page that presents them as the same thing is wrong in a way that is hard to see.

Sources and change log

Sources. The thirty five questions are our own, assembled from the commercial questions that decide mandates and the consent questions candidates raise. The dated counts quoted are from the aiknowsus.com audit of September 2026 captured to geo-audits/aiknowsus-com/, and the clawlaw.in programme of July to September 2026 recorded in Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md and Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md. For overseas placement, read the Ministry of External Affairs material on recruiting agent licensing and eMigrate at the source and record your date.

Change log. 29 September 2026: first published with the interview protocol, the two question sets, and N stated as zero for both populations. When interviews are conducted, this section will carry N for each population, the counts per question, the interview window, how many were approached and refused, the coder disagreement count, and links to the script and codebook.

Common questions

Why publish an interview study with no interviews?

Because the protocol is usable today and the number would be invented. An agency that runs twenty employer interviews properly will know more about its own market than any published study can tell it, and it will have a page nobody else has.

Can we interview our own clients?

Interview them, and report them as a separate group with its own N, because they have already chosen you. A study made only of your own clients is a satisfaction survey wearing a market research label.

How many interviews are enough?

Enough that you will print N next to every count. Twenty per population is enough to be useful if you say it is twenty. The failure is not a small N, it is a hidden one.

Should we publish quotes?

Only with explicit consent for that quote, with the employer unnamed, and never as the main evidence. Counts with N are the evidence. A quote is an illustration of a count you already published.

Is it safe to hold candidate data for a study?

Strip identifiers before coding, keep consent records separately, state your retention period, and delete on request. India's data protection law applies to what you hold, and a study of candidates' consent concerns that mishandles their data is not a study worth publishing.

What can we publish now, before any interviews?

Answer the thirty five questions yourself on your own site, clearly, one page per cluster: fees and invoicing, replacement and guarantees, time to fill by role, and how candidate data is handled. Those pages are useful to both populations and they are the ones an assistant can quote.

What to do first

Book five employer conversations and five candidate conversations for next month, use the fixed opening question, and record which of the thirty five come up before you prompt anything. Ten interviews is not a study and it will still change what you publish next, and when you are ready to publish a count you will already have the script, the codebook and the consent process in place.