Genlytics

A decade of SEO tooling exists to answer one question: where do I rank for this search? That question is quietly becoming the wrong one. When someone asks ChatGPT which CRM to use for an early-stage startup, there is no results page and no position ten. There is one answer, a handful of brands named inside it, and a list of sources it drew on.

Genlytics measures that. You give it the prompts your customers actually type, and it tells you whether you appear in the answers, where you sit when you do, how you are described, and which pages the model leaned on to get there.

A run is a matrix, not a query

The core object is a run: every tracked prompt crossed with every enabled platform. Prompts on one axis, models on the other, one cell per pair.

Modelling it explicitly rather than as a flat list of results is what makes the rest of the product tractable. A cell can be re-run on its own when one model times out. A whole column can be re-run when a provider ships a change worth re-measuring. A whole row can be re-run when you reword a prompt. Without the matrix, all three of those are "run everything again" — which is slow, and expensive in a way that scales with the number of prompts you care about.

One reading is not a measurement

Ask the same model the same question twice and you can get two different answers. Ask it from a different country and you can get a third. Neither is a bug; it is what a sampled, non-deterministic system does.

So Genlytics samples on a schedule. A daily cron pass re-runs every workspace that has it enabled, and each workspace pins the country it browses from — the answer a customer in the UK gets is not the answer a customer in the US gets, and averaging the two hides the thing you wanted to see. What you read is a trend line across repeated samples rather than a single reading dressed up as a fact.

Every engine is a different problem

Some platforms have an API. Some only have a web interface, which means driving a real browser through a residential proxy in the right country. Some have both, and they disagree.

Each one sits behind an adapter with the same shape, and a registry decides which adapters actually exist at runtime — a platform is simply absent when its credentials are not configured, rather than present and failing at the moment someone asks for a report. Browser-driven platforms run on a separate worker so a slow session cannot hold up the batch that a serverless function has to finish inside its timeout.

The uncomfortable part of this problem is that the ground truth moves. Providers change models, change how they cite, and change what they will answer at all. The adapter boundary is what keeps that a contained problem rather than a rewrite.