How an AI Answer Gets Built: From Prompt to Citation

When someone asks ChatGPT, Perplexity, Gemini, Claude or Google's AI Overviews a question about your category, the reply arrives as one confident paragraph. It reads as though the engine simply knew the answer and typed it out.

It did not. That paragraph was assembled in stages: from the question as the engine interpreted it, from what the model already knew, from what it searched for in the moment, and from which of those sources it decided to trust enough to name.

That matters commercially, because a brand can drop out at any one of those stages, and each exit has a different cause and a different remedy. "We are not showing up in AI answers" is not one problem. It is at least four, and treating it as one is why so much AI visibility work produces activity without movement.

The four stages of an AI answer

Answer engines generally build a response in four moves.

1. Plan. The engine takes the question and works out what to search for, usually rewriting it and splitting it into several searches.
2. Retrieve. It pulls candidate documents through search, blending literal keyword matching with meaning-based matching.
3. Rerank. It re-scores candidate passages against the question to decide which are worth using.
4. Generate. It writes the answer from the passages that survived, and cites some of them.

Each stage narrows the field. Planning decides what gets looked for at all. Retrieval casts the net. Reranking is the fine filter. Generation is the final cut, where a relevant, retrieved source can still go unnamed.

Stage one: the prompt is a doorway, not a target

The engine rarely searches for the question word for word. It interprets and expands it, fanning one question into several searches, each aimed at a different facet.

"Best project management tool for agencies" is not one search. Underneath it there is usually a features search, a pricing search, and a comparison search naming specific competitors. The exact words a buyer uses shape which of those run, and those searches decide which sources are pulled in at all. Ask the same underlying question two ways and you can get two different sets of sources, and therefore two different answers.

The consequence for a brand is simple. You are not competing to appear for one question. You are competing across the spread of searches that question fans out into. A brand can be a perfect fit for the headline question and still miss every specific search the engine actually runs beneath it.

The high-value moment is usually a comparison or decision phrasing, because that is where a buyer is choosing. Two brands can target the same broad topic while the engine, underneath, runs comparison searches that only one of them is prepared for.

This is where topical authority stops being an SEO abstraction. A single optimised page cannot cover a fan-out. A connected set of pages can.

Stage two: training memory and live retrieval

Once the engine knows what to search for, it draws on two different kinds of knowledge.

The first is the model's training knowledge: what it absorbed when it was built. It is fast and always available, but frozen, general, and blind to anything that changed after the cutoff.

The second is live retrieval. The engine runs the searches from stage one, pulls current documents, and grounds its answer in them. The technique is usually called retrieval-augmented generation. Grounding is the important word: a grounded answer rests on sources the engine can point to, and pointing is what produces a citation.

Retrieval typically blends two methods, keyword matching and vector matching on meaning, then merges the result lists. Pages that score well on both tend to be favoured.

For a brand, this creates two routes to relevance. Be established enough that the model knows you from training, which helps in answers with little live retrieval. Be structured and current enough to be retrieved live, which is what gets you into grounded, cited answers. The strongest position is both, which is why entity consistency and fresh, well-structured content are not alternatives.

Stage three: what is searched is not what is cited

This is the stage most people miss, and it is the one that explains most disappointments.

Retrieval casts wide. It brings back far more candidates than will ever appear in the answer. The engine then reranks: a model reads each candidate passage alongside the question and re-scores how well that passage actually answers it. Most retrieved sources do not survive.

Two details matter. The reranker scores passages, not whole pages, so a strong page can contain a weak passage and lose at this stage. And the scoring is against the specific question the engine wrote, not against your topic in general.

After reranking, generation applies its own logic: source quality, a preference for variety so the answer does not lean on one domain, and the plain fact that an answer only has room for a few citations.

So a source can be searched but not retrieved, retrieved but not reranked, or reranked but not cited. "We were in the sources but not in the answer" is a real, common, and entirely explainable outcome.

"Google AI Overviews are pulling 30 to 50 per cent of organic traffic. If you're not optimised for them, you're losing half your potential visibility."

John Russell, Director, Digital Womble

Being searched means the engine looked in your direction. Being cited means it chose you as a source worth standing behind. Closing the distance between those two states is most of the work.

Stage four: the answer is written, not copied

The final answer is composed, not lifted. The engine takes the surviving passages, adds what it already knows, and writes new sentences.

That is why a brand can be described in words that appear on no single page, and why an answer can be wrong in a way no audit of your own site explains. If the surviving sources are thin, dated, or disagree with each other, the engine will still produce one smooth, confident sentence from them.

Citations are attached to some of that. They are both the engine's evidence and your only reliable record of which sources it trusted.

Where a brand drops out, and what actually fixes it

| Symptom | Stage | Cause | Fix | | --- | --- | --- | --- | | Never mentioned at all | Plan | No content for the searches the engine runs | Cover the fan-out, including comparison phrasings | | Mentioned only for the brand name | Retrieve | Weak or missing page for specific facets | Dedicated pages, explicit titles, crawlable HTML | | Retrieved but never quoted | Rerank | Passages are vague or buried | Direct, self-contained answer passages | | Quoted in competitors' comparisons | Cite | Thin corroboration, no attributable expertise | Named sources, quotes, third-party corroboration | | Quoted but described wrongly | Generate | Stale or conflicting facts about you | Canonical facts, consistent entity data |

Why this changes the fix

Once you know which stage is failing, the work stops being "do more AI SEO" and becomes a specific instruction. A reranking problem is a writing problem. A retrieval problem is an architecture problem. A citation problem is usually an evidence problem.

That also means results are not guaranteed in advance. Engines re-run their own judgements constantly, and nobody can promise a citation. What can be engineered is the position: being in the candidate set for the searches that matter, having passages that survive a rerank, and being corroborated by sources the engine already trusts.

"AI systems reward multi-source expertise and direct quotes. Single-page optimisation doesn't compete. You need topical authority with real, attributed quotes."

John Russell, Director, Digital Womble

How we work on it

We run this as a service: a prompt set mapped to the pages that carry revenue, an evidence-led audit of which sources engines actually cite in your category, a ranked fix list separating architecture from writing from corroboration, and content built to be the passage an engine quotes rather than the page it merely reads.

What that costs, and what is included, is on the content strategy pricing page. If you want to see the mechanics on your own site first, start with the free SEO and AI search growth toolkit.

Related reading: how to rank in Google AI Overviews, why Google's models favour content that cites real sources, and what E-E-A-T eligibility actually requires.