Answer engine basics
How AI assistants pick their sources
The four steps between someone asking a question and a page getting cited, and where you can actually intervene.
Jonathan BoshoffLast updated 16 September 2026
Read this postLearn
Guides for people doing the work
How assistants pick their sources, how to structure a page they can use, and how to measure whether any of it moved. Written alongside the research, so the advice and the evidence stay in step.
Content structure
Writing a page a retriever can use
A retriever never reads your page. It reads a slice of it, so the slice has to stand on its own.
Jonathan BoshoffLast updated 17 September 2026
Read this postMeasurement
Measuring AI visibility honestly
One probe is a coin flip. Here is the smallest setup that produces a number you can defend in a meeting.
Jonathan BoshoffLast updated 18 September 2026
Read this postAnswer engine basics
How assistants search, read and decide, in plain language.
How AI assistants pick their sources
The four steps between someone asking a question and a page getting cited, and where you can actually intervene.
Fan-out queries, explained
One prompt becomes several searches. Those searches decide which pages are even in the running.
Mentioned versus recommended
Appearing in a paragraph and appearing in the shortlist are not the same result, and only one of them is worth a budget.
Why the assistants disagree with each other
ChatGPT, Claude, Gemini and Perplexity rarely agree on a shortlist. The disagreement is usually where the work is.
Content structure
Writing and shaping pages so a retriever can actually use them.
Writing a page a retriever can use
A retriever never reads your page. It reads a slice of it, so the slice has to stand on its own.
Headings that match the question
A heading that asks the searcher's question is one of the most consistent signals we have measured.
Giving a model something to quote
A model can only credit what it can attribute. Vague copy gets read and then skipped.
The freshness trap
Real updates move citations. A new date on an untouched page mostly moves nothing.
Measurement
Counting what happens in AI answers without fooling yourself.
Measuring AI visibility honestly
One probe is a coin flip. Here is the smallest setup that produces a number you can defend in a meeting.
Control prompts, and why you need them
Leave a few prompts untouched. When the tracked ones move and the controls do not, the work is yours.
Reading a confidence band without a stats degree
A rate of 60% with a band of 45% to 74% is a different result from 60% with a band of 58% to 62%.
Working with agents
Handing the work to an agent without losing the judgement.
Handing work to an agent without losing the plot
An agent is good at the work and bad at deciding which work matters. That split is the whole method.
Briefing an agent with evidence, not adjectives
An agent asked to "improve the page" will improve something. Point it at the four pages that beat you instead.
Logging what changed, and when
Work that is not logged against a date cannot be connected to the number it moved.
The research behind the guides.
Every claim in this section traces back to a study you can read, with the method and the limitations written down beside the numbers.
Read the research