Generative engine optimization

Less than the volume of writing about it suggests, and the differences that are real are worth knowing.

**The unit is the passage, not the page.** A search engine ranks a document; a model retrieves a chunk. A page that answers its question in one self-contained paragraph is more quotable than one that builds to the answer over six.

**Attribution replaces the click.** A citation is an impression rather than a visit, so the value shows up as brand recall and eventual direct traffic rather than in a session count. Measuring it the way you measure search traffic will make it look like a failure.

**Structure is read literally.** Question-shaped headings, definition sentences of the form "X is a Y that Z", tables and lists — these are easier to extract, and extraction is the whole mechanism.

What is not different: quality, expertise, accuracy, load speed, and being crawlable at all. Most GEO advice is SEO advice with a new label on it.

There is no Search Console for AI answers. No assistant reports impressions, and citation is not observable from your own server.

What you can observe is the fetch — which assistants opened which pages, how often, and whether the request succeeded. That is a leading indicator rather than a result, and it is the honest one: a page assistants fetch repeatedly is a page they consider relevant, whether or not any given answer cited it.

Anyone selling a GEO tool that reports your citation rate is inferring it, usually by querying the assistants themselves and counting mentions. That is a real technique with real limits — the answers vary by phrasing, by user, and by day.

Write the direct answer first and put it near the top. Use headings that match how questions are asked. Publish structured data, particularly `FAQPage` — a marked-up question and answer pair is the most quotable shape available. Keep the technical door open: an assistant that cannot fetch your page cannot cite it, and a robots rule blocking answer fetchers is the most common self-inflicted wound here.

And write things only you can say. A model synthesising from five sources that agree will cite whichever it trusts most; a page carrying a measurement nobody else has is the one that gets named.

This field generates confident numbers faster than anyone measures them, and a few are worth recognising before you plan around them.

Retrieval weighting tables — "semantic relevance 40%, authority 15%" and similar — appear in a lot of GEO guides. No lab publishes these, retrieval differs per product and per query, and the figures trace back to blog posts citing each other. Treat them as a way of remembering which factors exist, not as coefficients.

Citation-rate tools work by asking assistants questions and counting whose name appears. That is a real measurement of a real thing, with two limits worth stating: the answers vary substantially with phrasing, and they vary day to day as models are updated. A single run is a sample of one.

The claim that sitemap priority influences AI citation appears in at least one production codebase and has no source at all.

What is actually established is unglamorous and mostly familiar: content has to be fetchable, structure helps extraction, specific claims outrank generic ones, and a page nobody can reach cannot be cited by anyone.