Blog
Field notes on measurement
What we learn building AI answer collection: where the data lies, what a number has to carry before an agency can defend it, and what the surfaces outside the West actually do.
Seven AI visibility reporting mistakes that cost agencies credibility
Composite scores, missing intervals, mixed markets, silent parser changes — the reporting errors we see most often and the fix for each.
Read post →How to package an AEO service line agencies can actually sell
Scope, pricing, deliverables and the three promises not to make when adding AI visibility to an agency's offer.
Does llms.txt do anything? We tested it
We added llms.txt to half of a controlled page set and measured citation outcomes for 60 days. The effect was inside the noise band.
Which AI crawlers to block, and which blocks quietly cost you money
Training crawlers and answer indexers are different agents. Confusing them is how sites accidentally remove themselves from ChatGPT.
The screenshot problem: why one AI answer proves nothing
Single-reading evidence is the most common form of AI visibility reporting and the least defensible. Here's what it costs when a client checks.
When a provider API doesn't return what users see
OpenAI's web search tool, Perplexity's Sonar and Gemini's grounding metadata can return a different answer than the interface. Here's when that happens and how to collect around it.
One reading is not a measurement
AI answers are non-deterministic. Reporting a single sample as a client's visibility score is reporting noise with a decimal point on it.
How to measure AI visibility properly: a working method
A step-by-step method for measuring brand presence in AI answers, including sample sizes, prompt design, market handling and what to do when a client disagrees.
The twelve AI answer surfaces that matter, and how they differ
AI Overviews, AI Mode, ChatGPT, Perplexity, Copilot, Gemini, the Chinese five and Naver. Each has its own access model, citation structure and failure mode.
AI Overviews vs AI Mode: two surfaces, one merge on the horizon
One appears whether the user wants it or not. The other is a tab they choose. They cite differently, and Google says they're merging.
AEO vs GEO vs LLMO: three names, one discipline
The terminology fight is a distraction, but the terms do carry different emphases. Here's what each community means and which one to use with clients.
What answer engine optimisation actually is (and what it isn't)
AEO is not SEO with new words, and it is not a checklist of schema tags. It's the practice of being one of the few sources a model synthesises from.
Tracking Chinese AI search: a different web, different rules
DeepSeek, Qwen, Doubao, Kimi and Baidu ERNIE are not translated ChatGPTs. Western AEO advice mostly inverts there.
What AEO agencies should actually report to clients
Citation share with intervals, evidence on demand, competitor entry alerts and per-surface history. Not a vanity score.
The economics of scraping AI answers
Bandwidth is the cost line. Residential proxies, resource blocking, stream terminators and query pooling decide whether the unit economics work at all.
How to build an AI visibility dashboard on top of an API
A practical shape for agencies: project per client, prompt sets as templates, nightly runs, alerting off intervals, and white-label reporting.