Nexus reads from Google Sheets, BigQuery, Amazon S3, and direct file upload. Platform data — Google Ads, Meta, GA4, your CRM — comes in through whichever of those you already use: a scheduled export to Sheets, a table in your warehouse, or a file. There are no native API connectors to the ad platforms today.
For example, a team can export Google Ads and GA4 data into Google Sheets or BigQuery, connect that data to NEXUS, and bring in CRM information through the same process. NEXUS then helps organize and prepare the data for analysis.

Yes. Uploaded data is normalised and column-mapped on arrival, joins across sources are proposed and reviewed rather than guessed, and naming conventions are mapped to a shared vocabulary so “Prospecting” and “prospecting_us” are the same tactic. Validation runs before modelling, not after: multicollinearity (VIF), channel co-movement, and statistical tests on the series themselves, with the findings shown to you rather than silently passed through.

NEXUS does not currently provide an embedded SQL or Python editor. Instead, the platform is designed to simplify analytical workflows through automated data processing and natural-language instructions. Users can work through the platform's built-in analytical capabilities without needing to write code.

Yes, and the diagnostics ship with the result. Models are fit with PyMC-Marketing. Every run reports R-hat and effective sample size, and is marked converged or not against explicit thresholds — a model that didn’t converge says so rather than presenting its numbers as fact. Sampling uses a fixed seed, so the same inputs reproduce the same posterior. R² is reported alongside.

Yes. Specialists propose tests, managers approve them, and results are recorded against the original hypothesis — so a test that didn’t work stays on the record. Separately, Nexus measures the lift around calendar events and promotions using a pre-period counterfactual, and labels that estimate as non-causal, because a before-and-after comparison isn’t a randomised test and shouldn’t be sold as one.

Every figure in an AI-written explanation is checked against the numbers the model was given. If the reply contains a figure that wasn’t supplied, the whole answer is discarded rather than shown. The arithmetic is done in code; the model’s job is to say what the numbers mean, not to produce them.