Historical context: This article was first published in 2020 and described an exploratory SEO modelling workflow. The account below preserves the method’s purpose and explains its limits. It is not a current Dataiku installation tutorial or evidence that the model can improve rankings.
Table of Contents
What the experiment tried to examine
The original workflow collected pages appearing in search results, joined available page and link measurements, and used Dataiku to model observed positions. It described Random Forest and XGBoost analyses and examined their feature-importance outputs.
That is a predictive modelling question about a collected sample. It is different from identifying Google’s ranking rules. Google uses multiple systems, and search position varies by query, context and time.
The historical workflow
- Collect a defined set of search-result URLs and their observed positions.
- Gather available page, link and performance measurements.
- Join records using consistent URL identities and check missing or duplicate rows.
- Import the resulting dataset into a modelling environment.
- Train and evaluate candidate models.
- Inspect model behaviour and investigate possible explanations.
The original article referred to Ahrefs, Screaming Frog, BatchSpeed and Dataiku. Product interfaces, editions and performance measures have changed. This summary does not verify those historical setup instructions or recommend purchasing that tool combination.
Why feature importance is not a ranking factor
A feature-importance score describes a particular fitted model and dataset. It does not establish that changing the feature will cause a search-ranking change. Correlated inputs, measurement error, selection and information leakage can all affect interpretation.
The original account highlighted referring-domain measurements. That observation does not establish a universal link-building budget rule or justify manipulative link acquisition. Scikit-learn’s documentation explains why a model’s predictive quality should be evaluated before interpreting feature importance.
What a reproducible investigation would need
- The original query set, market, device, observation dates and URL sample.
- Source datasets, field definitions and permitted reuse.
- Cleaning, joins, missing-data treatment and duplicate handling.
- A documented separation between training and evaluation data.
- Model settings, baseline comparisons, uncertainty and sensitivity checks.
- A clear distinction between prediction, association and causal claims.
This page does not provide a reproducible validation package for the historical result. Treat the account as an exploratory method note rather than established campaign proof.
Use the observation to form a question
A model can suggest something to investigate. A practical next step is to inspect the actual pages, verify measurements and consider alternative explanations. Do not convert a model output directly into a claim that a particular change will produce traffic or sales.
For current foundations, see keyword research and ranking measurement. This historical experiment is not an AI-search optimisation service offer.
