Where an AI answer gets its sources.

A question becomes an answer in four steps

Each stage narrows the field. Dozens of pages become a handful, a handful becomes one answer, and that answer carries two or three links. Your job is to work out which narrowing removed you.

The question you were asked is rarely the query that ran

Answer engines rewrite. A buyer typing “what should I use to invoice clients if I’m a one-person plumbing business” does not produce one search for that sentence. It produces several shorter ones: invoicing software for tradespeople, best invoicing app sole trader, and so on. And the union of those results forms the candidate set.

Two consequences follow, and both are useful. First, exact-phrase optimization is pointless, because your phrase is not what ran. Second, breadth of coverage beats precision: pages that address a topic from several angles get pulled into more of those sub-searches.

This is the stage where SEO is still doing the work

Candidates get ordered before the model reads them, and that ordering leans on the same signals search has always used: topical relevance, link and brand authority, freshness, and whether the page appears to answer the question rather than merely mention it.

This is precisely why GEO does not replace SEO. Everything that makes you rank makes you a candidate; being a candidate is the price of admission to the stages where the newer work happens.

The model needs something it can lift

Now the model reads the top candidates and writes an answer. It is looking for passages that stand on their own and say something definite. A paragraph that only makes sense after the three above it is hard to use; a paragraph that asserts a specific, checkable thing is easy.

This is the stage where good marketing pages lose to mediocre documentation. The marketing page is written to build to a conclusion. The documentation states the conclusion and moves on, and that is the shape that gets synthesised.

Being used and being credited are not the same

The final stage attaches sources to statements. It is possible, and common, for your page to inform an answer without being linked, because another page said the same thing more quotably, or because the answer engine simply cites sparsely.

Watch for the pattern where you are mentioned accurately and cited elsewhere. The model clearly knows your specifics and is crediting an aggregator that repeated them. The fix is usually to be the clearest statement of your own facts, not to publish more of them.

Four symptoms, four different weeks of work

Run your question, then work down this table until a row matches. The first match is your stage.

One caution: check this on more than one answer engine before deciding. A retrieval failure on one engine and a synthesis failure on another is entirely possible, and it means two different jobs rather than one.