Writing

Notes from the audits

What we find scoring real B2B sites for AI visibility — the patterns, the mistakes, and the findings worth acting on. Every claim is sourced or is our own measured data. No news roundups.

15 September 2026 · Research

AI crawlers: training vs retrieval — why most “blocked” findings are cosmetic

We scored 137 UK B2B sites for AI crawler access. Almost every blocker blocks the harmless class — and the audits selling that finding cannot tell the difference.

15 September 2026 · Mechanics

How an AI assistant decides who to cite

Retrieval, extraction, attribution — three gates, three different failures. Semrush found only 6–27% of the most-mentioned brands also rank as top sources.

15 September 2026 · Contrarian

llms.txt: what it actually does, and what Google says about it

Google says it is not needed for AI Overviews. Anthropic recommends it. Both are right, because they are answering different questions.

15 September 2026 · Evidence

AI traffic converts better than organic — and you are undercounting it

0.5% of visitors, 12.1% of sign-ups. What five published datasets agree on, and why your analytics shows less than the truth.

15 September 2026 · Method

Measuring AI visibility without fooling yourself

Ask a model the same question three times, get three answers. The protocol that survives a client re-running it themselves.

15 September 2026 · Case study

Entity clarity: how a model decides who you are

Google described us as a Polish local-business tool and mixed us up with laboratory glassware. What an entity collision looks like from the inside.

Reference

Start here

Definition

What is Answer Engine Optimization?

A plain definition, the three things a model must do before it will cite you, and what the work consists of.

Comparison

AEO vs SEO: what actually differs

Where the two overlap, where they diverge, and which one your problem belongs to.

Terminology

AEO vs GEO: is there a real difference?

Two acronyms, one discipline — and the four questions to ask a provider instead.

Protocol

How we measure AI visibility

Twenty fixed questions, three runs per engine, and the limits of what any of it proves.

Want this run on your site?

Send your domain. The check comes back free, without a call.