AI Survey Analysis in India: How the Poseidon Engine Works

Analysis starts when a study is marked complete. Responses are collected through the SuperJ app, and the engine ingests them once the survey ends.

The typed model is why routing matters. If only smokers saw question 12, the base for question 12 is smokers, and the engine knows that from the survey structure. A generic tool reading a CSV does not.

Every number in an answer comes from a SQL query, not from model text. The model drafts the query and the wording. To do that it is given column names and types, summary statistics computed in SQL, and up to 10 sample rows. It is not given the full response table, but those sample rows are real responses, so ask us whether free-text answers can appear in them and what the model provider's data terms are.

Three details show what that means in practice:

• Multi-select questions. Answers are stored as one cell such as "A;B;C". The engine splits them so counts are selections, and it uses respondents as the denominator for percentages. Mixing the two is a common error in hand-built tables.
• "Don't know" codes. Codes such as 98 or 99 are set to null on rating and NPS columns, so they do not drag a mean down.
• Checked column names. Before a query runs, every column it names is checked against the real schema. A wrong name is refused with a suggestion, not guessed.

Only the result goes back to the model. This separation is what makes the checks below possible, because a number from SQL can be recomputed and a number from text cannot.

Each answer also carries a confidence score, shown as a low, medium or high badge. It combines the verification results with how well the answer matches your question, the statistical power, the data quality and whether execution succeeded. Treat it as a prompt to look harder, not as a guarantee.

We publish no accuracy percentage for these checks, and you should distrust any vendor who does without a method and test data. Run your own test: ask for five numbers and recompute them from the export.

The formula is chosen by question type. These are the ones we can state precisely:

Its formula library also lists chi-square, t-tests, ANOVA, non-parametric tests and confidence intervals. For when each test applies, see our significance testing guide. For the design side of MaxDiff and Kano, see those guides.

Ask for a cut, a base and a method in the same sentence. Our bank, by job:

Then ask the follow-up the engine suggests, and one more of your own: "Is that difference statistically significant, and what is the smallest group in this comparison?"

When the study is complete and its data is ready, a multi-stage report pipeline can write a full report. Your research brief steers it. The stages plan the sections, plan the analyses, run them, group open-ended answers into themes, draw charts, write insights, then audit the draft for consistency and safety before finalising.

A report follows a fixed order: executive summary, methodology, a section for each goal, data quality, respondent profile, question findings, cross-question analysis, derived analytics, themes, recommendations, conclusions and an appendix last. The drafting rule is that each paragraph says what the finding means for the decision, then gives the evidence. You can download the report as Markdown, HTML or PDF. Our help FAQ also lists Excel and PPTX downloads, which the engine's own documentation does not; treat Markdown, HTML and PDF as the confirmed formats.

Ask for a theme question in the report, such as the themes in open-ended answers on price, and read the raw answers behind each theme. Read the executive summary last. Check the data-quality section first: it flags completeness, outliers and small bases.

• No Gabor-Granger analysis. You can design a Gabor-Granger study in our survey builder, but Poseidon does not compute the demand curve. Analyse it yourself from the data. See our Gabor-Granger guide.
• Van Westendorp is a template only. Poseidon has a starting template, not a full executor. Check the price points by hand, as in our Van Westendorp guide.
• No choice-based conjoint. The survey builder does not offer it either. See conjoint alternatives.
• It supports analysts, it does not replace them. It does not know your market, and it cannot tell you whether a sample suits your decision. Read the verbatims yourself and decide what to act on.
• Weighting. The engine can compute effective sample size from weights. Ask whether your study's results are weighted to population targets before you rely on them.
• Panel limits. Data comes from smartphone users on the SuperJ app, so households without a smartphone are under-represented.

Our public pages do not describe running Poseidon on responses collected elsewhere. Ask first if that is your plan.

If your job is a consumer survey in India with automatic analysis, this is a test you can run. Field a short survey in two cities where you already know one answer differs, then check that the cross-tab shows it, every figure in the report matches a table, and the themes fit the raw answers.

Pro is sold in bundles of 400 to 1,000 responses and costs 20% less per response than Plus. Plus lists an "Advanced analytics engine"; the plan list does not name Poseidon. Whether the Free plan includes Poseidon reports is not described on our public pages, so ask. Enterprise plans come with a dedicated research team and are priced on request. See pricing.

The SuperJ panel has a 25M+ consumer footprint. The verified, survey-eligible panel is a subset, and availability depends on your targeting. Members complete Aadhaar-based KYC once and each study receives a zero-knowledge proof, not identity documents. Surveys run in as many Indian languages as the study needs, and speeders, straight-lining and inconsistent answers are removed before analysis. Details are on the panel methodology page. The limit is the one above: smartphone users only.