Start from your objective, not your software
Most wrong analyses start with the tool rather than the question. Before you open SPSS, R or Python, write one sentence that states what you want to know: are you comparing groups, testing an association, predicting an outcome, or forecasting over time?
The recommender asks for exactly that objective first, because it changes the answer more than anything else. Comparing three or more group means points to ANOVA; predicting a yes/no outcome points to logistic regression; a value measured repeatedly over time points to ARIMA or SARIMA.
Describe your variables honestly
Enter the true measurement level of your variables. A five-point Likert item is ordinal, not continuous, even though your software will happily average it. A count of visits is not the same as a percentage.
Where your outcome is categorical and your predictors are categorical too, the recommender points to chi-square. Where you have many correlated continuous items and want to reduce them, it points to principal component analysis.
Check the assumptions the tool lists
Every recommendation comes with the assumptions that must hold: normality of residuals, homogeneity of variance, independence of observations, linearity, absence of severe multicollinearity, and adequate sample size per cell.
If an assumption fails, use the fallback method listed beside it — a non-parametric equivalent, a transformation or a robust standard error — and report why you switched. Examiners and reviewers accept a justified fallback; they do not accept an unchecked assumption.
Case note: a Port Harcourt retail study
A postgraduate researcher wanted to know whether adopting mobile payments changed monthly revenue for small retailers. The objective was prediction with a continuous outcome and a mix of categorical and continuous predictors, so the recommendation was multiple linear regression with robust standard errors.
Two assumptions failed on the first pass: residuals were skewed and variance grew with revenue. Log-transforming revenue fixed both, and the final model was defensible in the viva.