Home › Methodology

How klarava measures GEO/AEO

GEO/AEO is a young field, full of opinion and little measurement. klarava exists to change that: instead of guessing whether AI understands a website, we measure it, along a defined method with a deterministic, reproducible core. This page explains how we think and score, openly. What it does not do is hand out the exact recipe, and that distinction is deliberate.

At a glance

Every check uses up to three layers: two always run and are deterministic and reproducible (same captured content, same scoring version, same date → same result), plus, where available, an AI analysis for nuance. A score is orientation, not a visibility guarantee. And the two things that would let anyone copy or game the system, the exact weighting model and the ready-made fixes, stay ours.

Measurement, not opinion

Most GEO advice is a checklist someone believes in. That is a weak basis for decisions and for spending. Our starting point is the opposite: a website's AI understandability should be quantified the same way twice, so you can compare it, track it over time and prioritize what actually moves it. Rigor is the point, not a longer checklist.

The three layers we measure

Each layer answers a different question an AI implicitly asks about a site.

Technical / GEO (deterministic). Can a machine fetch, read and extract the site at all? We evaluate signals such as AI-crawler access, server vs. JavaScript-only rendering, structured data, answer structure, sitemap and agent-readable files. These are facts about the page, so the result is reproducible.
Brand clarity (deterministic). Can a complete, current picture of the business be built from structurally verifiable signals? Entity/Organization data (schema), people and team signals, evidence and sources (profiles, review/authority platforms) and freshness (date signals), each judged against defined criteria. This layer scores the structure, not the semantic clarity of the content, that is the AI layer's job.
AI analysis. Beyond the mechanics, this is about the content: how well does a language model actually understand the site, do who/what/for-whom come across clearly, what would a prospect ask an AI, and where are the gaps? This layer adds nuance that pure rules miss, which is why we keep it separate and label it as the model-based part.

Why deterministic and AI

The deterministic layers give you numbers you can trust and repeat, the backbone of any benchmark or monitoring over time. The AI layer catches the things rules cannot: ambiguity, tone, whether the story actually adds up for a reader who is a machine. Used together, you get a score that is both stable and meaningful, rather than one that is only one or the other.

What a score means, and what it does not

A high score means an AI can reliably fetch, read and understand your site, and build a correct picture of your business. That is the foundation of showing up correctly in AI answers. It is not a promise of a specific placement in ChatGPT or Perplexity: those systems change, weigh many signals and are outside anyone's direct control. We treat the score as a decision aid and a direction, and we say so plainly. Beware anyone who sells a guarantee here.

What we deliberately keep private

Transparency about our approach is a matter of trust. But two things stay behind the service, on purpose:

The exact scoring model. The precise weights, thresholds and formulas are our core IP. Publishing them would let the score be gamed and copied, and would turn a measurement into a checklist to fake. You get the result and what drives it, not the machine behind it.
The tailored fixes. Knowing a gap exists is not the same as closing it well. The concrete, site-specific building blocks, the exact structured data, the FAQ and profile copy, the prioritized action plan, are the work, and the value. That is what we do for clients, not a template to copy.

In other words: we are not a black box, but the model is our craft. If a method is fully published, it is no longer a method, it is a commodity.

From measurement to result

Measuring is where it starts, not where it ends. Once the gaps are clear, the point is to close them, and to keep them closed as AI systems evolve. That is where the work lives: ready-to-use building blocks tailored to a specific site, and monitoring that tracks understandability over time instead of a one-off snapshot.

See in seconds how well an AI understands your website, measured, for free.

Run the free check

Frequently asked questions

Is klarava a black box?

No. We explain the three layers, what each one evaluates, and which parts are deterministic versus AI-based. What we do not publish is the exact weighting model, so the score cannot be gamed or copied. Transparent method, private recipe.

Why not publish the exact weights and thresholds?

Because a fully published scoring model stops being a measurement and becomes a checklist to fake. The weighting is our core IP, and keeping it private is what makes the score meaningful and hard to game.

Is a good score a guarantee that AI will recommend me?

No. A good score means an AI can reliably read and understand your site, which is the foundation. Actual placement in AI answers depends on systems that change and weigh many signals. We treat the score as orientation, not a guarantee.

What makes the result reproducible?

Two of the three layers are deterministic: they assess facts about the page against defined criteria. Given the same captured content and the same date, they yield the same result (freshness signals are relative to the check date and so change over time). That is what makes benchmarking and monitoring over time trustworthy. The AI layer is labelled separately as the model-based part.

Can't I just do this myself after reading this?

You can absolutely improve the basics, and this page tells you what matters. What it does not give you is your own measured score, the exact model, or the tailored fixes and ongoing monitoring. That is the part we do.