Classic keyword research is technically dead.

What replaces it

Querymantic is demand intelligence: it maps the full fan of questions behind a topic, not just the phrases a keyword tool can count. It's a suite for Claude Code and Claude Cowork that takes the keyword exports you already have, from Semrush, Ahrefs, Google Search Console, Moz, or Ubersuggest, and turns them into a structured map of demand: what exists, how it clusters, where a client is invisible, and which queries now resolve inside AI search.

Seven modules do the work.

• Fan-Out Radar simulates the sub-queries an AI engine expands a topic into, then measures how much of that fan your content covers.
• Citation Grid scores, cluster by cluster, how ready your content is to be cited in AI answers.
• Entity Web maps the entities you own and the ones you don't, yet.
• Demand Pulse reads demand over time: rising intents, fading ones, seasonality.
• Click Ceiling estimates the clicks you can still win, so priority follows winnable clicks instead of raw volume.
• Live Wire, strictly opt-in, folds observed Search Console and AI-citation data into the picture.
• Output Forge turns one run into client-ready deliverables: dashboard, slide deck, audit document, workbook.

It reads Italian natively, on top of English, French, German, and Spanish.

What GEO actually rewards (the part most pages get wrong)

Generative engine optimization is structure, not a trick. The foundational study on this, GEO: Generative Engine Optimization (Aggarwal et al., Princeton, KDD 2024), measured 10,000 real queries and found that specific, boring content properties move AI citation probability by a wide margin. The biggest levers were not keywords. They were citing your sources, stating concrete statistics, and writing in clear, quotable passages.

Three findings worth knowing, because Querymantic scores against them:

• Citing authoritative sources inline was the single highest-impact change in the Princeton study, associated with large gains in citation visibility. An AI engine reads a page that already cites good sources as more trustworthy.
• Answer-first structure matters. AutoGEO (ICLR 2026) found that putting the conclusion in the first 150 characters after a heading is associated with a meaningful lift, because engines extract from the opening of each section.
• Passage density matters too. Practitioner testing across the GEO community converges on the same pattern: self-contained paragraphs of roughly 50 to 150 words, each carrying a concrete data point, get pulled into AI answers more readily than long, sprawling blocks. This one is field observation, not a controlled study, so treat it as the softest of the three.

These are reported effect sizes from specific studies on specific datasets, not guarantees for your site. Treat them as direction, not as physics. That caution is the whole point: Querymantic exposes the assumptions as parameters you can change, rather than selling them as certainties.

There is also a clear negative result in the same literature. Keyword stuffing, the old SEO reflex, showed no GEO benefit and sometimes hurt, because it degrades the fluency that engines reward. The methods that work and the methods that backfire are both in the open method.

SEO and GEO are one job, not two

You cannot be cited by an AI engine if a crawler can't read you in the first place. Technical SEO is the floor; GEO is what you build on it. Querymantic treats them as one pipeline because the same corpus feeds both: crawlability, structured data, and entity signals decide whether you're readable, and answer-shaped content, citations, and topical coverage decide whether you're chosen.

The framework underneath is dual on purpose. Search engines judge the source: experience, expertise, authority, trust, the E-E-A-T signals Google formalized. AI engines judge the passage: is it clear, is it organized, is it referenceable, does it say something the rest of the web doesn't. Querymantic scores a corpus on both sides, so a recommendation isn't "write more," it's "this cluster is readable but not citable, and here's the specific gap."

Offline, deterministic, traceable

Querymantic makes no external calls and loads nothing at runtime. No API, no account, no telemetry. Your data stays yours, on your machine, which is the same answer I give every client who asks where their data goes.

Same input, same output, byte for byte. Every score that rests on an assumption, like how wide the fan-out runs or what share of searches end without a click, is exposed as a parameter with its provenance recorded in the output. You can challenge it, override it, and watch exactly what changes. A number you can't trace is a number you shouldn't trust, including mine.

One thing said plainly: there is no secret markup and no special file that makes an AI cite you. Anyone selling that is selling a trick. What works is the structure above, applied honestly, on a site a crawler can actually read.

The method is open

The full suite is on GitHub under an MIT license, with the engine, the tests, and the methodology written down module by module. I'm building it in public. Take it, run it on your own exports, check my numbers against the studies they come from. That's the point of the rigor: it survives inspection.

The demand audit

If you'd rather have the analysis than the tooling, that's the work I do. I run the method on your data and deliver the full picture: where your demand lives, what AI search is doing to it, and what to fix first, scored against the GEO and E-E-A-T criteria above. You get the branded deliverables, dashboard, deck, audit document, and workbook, plus a strategy session to turn them into a plan. Fixed scope, five to ten days, the lowest-risk way to test the method on your own market.

FAQ