When an AI assistant answers a question and cites three sources, a pipeline ran. Understanding its stages explains why content that ranks well sometimes goes uncited, and why content that ranks modestly sometimes gets quoted.
The pipeline
Query understanding. The user's question is interpreted and often decomposed into several sub-queries. A single question like "is a heat pump worth it for a 1930s house" may become separate retrievals about heat pump efficiency, insulation in older construction, and regional cost data.
Retrieval. Each sub-query hits a search index — sometimes a conventional web index via a search API, sometimes a vector index of embedded documents, frequently both. This stage returns candidates, typically far more than will be cited.
Re-ranking. Candidates are scored for relevance to the specific sub-query, often by a dedicated model that reads the passage and the query together. This is more semantically precise and considerably more expensive than the retrieval step, which is why it runs on a shortlist.
Generation with grounding. The model composes an answer from the surviving passages. Citations attach to the passages actually used.
Classic SEO influences the retrieval stage most directly, because that stage often uses a conventional index. The re-ranking and generation stages operate on your passages, not your page, and that distinction drives most of what follows.
Why chunking changes what you should write
Retrieval systems do not consider whole documents. They split content into chunks — a few hundred words each — embed them, and match at the chunk level. Whatever ends up in the answer is a chunk, or a few.
This has direct writing consequences:
Sections must survive removal from context. A passage that begins "As we saw above, this means the second approach is usually better" is unusable in isolation. One that begins "Zirconia crowns cost more than porcelain-fused-to-metal because the milling process is slower" survives being lifted out. Write sections that would make sense if someone read only that section, because that is frequently what happens.
Headings are retrieval anchors. Descriptive headings that restate the question in the user's language give both the embedding and the re-ranker a strong signal. "Cost factors" is weaker than "What changes the price of a dental crown."
Front-load the answer. If the answer to the section's implicit question is in the first sentence or two, a chunk boundary is unlikely to separate the question from its answer.
What makes a passage citable
Across the systems whose behavior can be observed, a consistent pattern shows up in what gets quoted:
- Specific, checkable claims. Numbers, ranges, named mechanisms, and concrete conditions. Vague competence — "it depends on many factors" — is unquotable, because there is nothing to quote.
- Stated scope and conditions. "In the US, for single-family homes built before 1940" gives a model the qualifiers it needs to use the claim safely. Unqualified claims are risky to reproduce and often get skipped.
- Corroboration. Claims that agree with other retrieved sources are safer to assert. Content that is the lone outlier tends to be dropped, whether or not it is right.
- Recency signals. Visible dates, and content that acknowledges the current state of a changing topic.
- Absence of promotional framing. Passages that read as marketing copy are less likely to be surfaced as factual grounding. This is the point where SEO content written the usual way fails hardest.
What you can measure
Measurement here is genuinely immature, and it is better to say so than to invent precision.
What you can do: maintain a set of questions your customers actually ask, run them against the major AI interfaces on a schedule, and record whether you are cited and what was said about your category. This is manual, sampled, non-representative, and still more informative than nothing. Track citation frequency over time as a directional signal.
What you cannot yet do reliably: attribute revenue to AI citations, measure impression volume inside AI interfaces, or A/B test your way to citation. Referral traffic from AI interfaces is visible in analytics and growing, but it captures only the fraction of users who click through, which is by construction a minority.
The strategic bottom line
Almost everything that improves citability also improves conventional search performance: clear structure, specific claims, honest scoping, current information, and crawlable delivery. There is no meaningful trade-off to manage.
The one genuine change in emphasis is the shift from optimizing pages to optimizing passages. If your content is organized so that any given section can be lifted out, understood alone, and quoted accurately, you are aligned with how these systems consume the web — and you have also written something a human reader can skim, which was always the goal.