Method

How we measure.

One method, for any category. Every number on this site links here: the questions, the sessions and the counting rule behind each one, so you can check the work yourself.

Updated 27 October 2026Maintained by Seenova LabsThe method · A client Diagnostic · Case one · Words

One method, any category

Six questions. Three assistants. Three runs.

Each question is asked on ChatGPT, Perplexity and Gemini, three times each, in fresh sessions. Fifty-four answers make one baseline.

The grid shows 54 answers: six questions, each asked three times on each of ChatGPT, Perplexity and Gemini.

Why three runs.

Assistants give different answers to the same question. One run proves nothing. Three runs give a rate you can check.

“Which firms are best for office fit-outs in Dubai?”

Run 1
  1. Firm A
  2. Firm B
  3. Firm C
Run 2
  1. Firm A
  2. Firm C
  3. Firm D
Run 3
  1. Firm B
  2. Firm A
  3. Firm E

Same question. Three runs. Different names. Illustrative, no real firms.

Entry 1 · how a client Diagnostic is measured

Your questions, three assistants, three runs.

  • QuestionsSix, agreed with you by day 2, written the way your buyers ask. Your company's name is never in them.
  • RunsEach question asked three times on each of ChatGPT, Perplexity and Gemini, in fresh sessions, on separate days where possible: 54 answers.
  • By-name checkOne extra question on each assistant: "What do you know about [your company]? Would you recommend it?" Recorded word for word, including what it gets wrong and what it warns buyers about.
  • What is reportedFor each question and assistant, a rate: named in 1 of 3 runs, 3 of 3, 0 of 3. Every competitor named, with its own rate. Every answer in full.
  • Source mapEvery website the assistants cited across the 54 answers, grouped: your own pages, review sites, directories, articles, competitors' pages. This is what decides the fix.
  • The fixesFix 1, by day 12: one page of your facts on your own site, live or handed to you ready to publish. Fix 2, by day 20: changes to your own profiles on the review and listing sites the assistants read in your category (for software, sites like G2 and Capterra), submitted, plus each platform's own review-request link where it offers one.
  • RerunIn monthly work, after the fixes, the same six questions, same assistants, same counting rule, compared against the day-7 baseline. Movement is reported as a rate, not a promise. The Diagnostic itself does not include a rerun.
  • WhoEvery engagement is asked, recorded and counted the same way.
Entry 2 · Case one

Case one: HR and payroll software, India. September 2026.

The source of the "36 answers", "1 vendor in all 36", "10 vanished" and the chart on the home page, and of the Case one report.

  • CategoryHR, HRMS and payroll software bought by companies in India.
  • QuestionsSix. Written the way a buyer asks, with no company named. Listed word for word below. The wording was locked on 17 September 2026 and not changed between runs.
  • AssistantsChatGPT, Perplexity and Gemini, used through their normal web interfaces, not through developer access.
  • RunsRun 1: Thursday 17 September 2026. Run 2: Tuesday 22 September 2026. Each run asked all six questions once on each assistant: 18 answers per run, 36 in total. A third run was planned for 29 September and was not done. This case uses two runs.
  • SessionA fresh, logged-out or private session for each question. Nothing said about testing visibility. No follow-up prompts.
  • Asked fromCase one was asked from India.
  • RecordEvery answer saved in full as a text file, with the sources the assistant showed, in the same folder as this count. The files are kept for inspection on request.
  • Counting ruleA vendor counts as named in an answer if its name appears anywhere in the answer text, in any case. A name that appears only in the assistant's list of sources does not count. Product names of the same company count once for that company. Counting was done by a script against a watch list of 59 vendor names. The script and the list are kept with the answer files.
  • The three hero figures36 = 6 questions × 3 assistants × 2 runs. 1 = one vendor was named in all 18 answers of run 1 and all 18 of run 2. 10 = vendors named at least once on 17 September and in none of the 18 answers on 22 September.
  • The chartVendors A to D were named in 36, 34, 31 and 28 of the 36 answers. Six more vendors were named between 10 and 20 times (20, 19, 16, 14, 11 and 10). Eighteen vendors were named three times or fewer, eight of them exactly once. Twenty-eight vendors from the watch list were named at least once.
  • Names withheldWe do not publish vendor names from this test, because we may work with any of them. The letters A to D are the same vendors on every page.

The six questions, word for word

  1. What's the best HRMS for a company in India?
  2. What's the best payroll software for Indian PF, ESI and TDS compliance?
  3. What's the best HR and payroll platform for a 100-person company in India?
  4. What's the best HR software for Indian startups and SMBs?
  5. Which HRMS platforms are recommended for companies with employees across multiple Indian cities?
  6. What are the most recommended HRMS tools for a mid-sized Indian company?

What we don't claim from this test

  • It covers one category in one country on two days. It does not predict what the assistants say about your category.
  • Two runs is a small sample. The 10-vendor swing between runs is itself the finding: a single answer proves nothing.
  • Assistants change between runs and between accounts. Your own result on the same question may differ.
  • The watch list could miss a vendor we did not think to list. If you are an HR-software vendor and were named in an answer we missed, write to us and we will correct the count.
Words used on this site

Plain definitions.

  • AnswerOne reply from one assistant to one question in one session.
  • RunAsking every question once on every assistant. Two runs means every question was asked twice on each.
  • NamedThe company's name appears in the answer text. Being named is not the same as being recommended; the report shows both.
  • RateHow many answers out of how many named you. "3 of 9" means three of nine answers to one question.
  • Source mapThe list of websites an assistant showed as its sources, across all answers.
  • BaselineThe first full measurement, before any fix. Later runs are compared with it.

Change log

  • 27 Oct 2026 — Corrected the third-run note; added where Case one was asked from.
  • 23 Oct 2026 — The method now comes first, for any category; the September test is Case one. The Diagnostic includes two fixes; asking the questions again happens in monthly work.
  • 6 Oct 2026 — Page published with Entry 1 (the September public test) and Entry 2 (how a client Diagnostic is measured). Client entries will be added here, each with dates and a sample size, only with the client's written permission.

Tell us your category and your buyers' country.

The Diagnostic runs this method on your buyers' questions, with every answer on record.