Back

How AI Overviews Generate Answers: The Sub-Query Layer That Decides Your Citation

Tengence Team ·
ai-search GEO/SEO
How AI Overviews Generate Answers: The Sub-Query Layer That Decides Your Citation

Summary: An AI overview is not "a summary of the top-ranked article". It usually decomposes a question into several sub-questions, retrieves and extracts each one, assembles them, and attaches citations to a few chosen passages. Once you see that layer, you understand why some long articles are never cited while a 200-word Q&A page appears daily — the unit selected is a paragraph, not a document.

Key Takeaways

1. From one question to one answer block

A user types a sentence; internally, three steps follow.

Step one: understand and decompose. The question is split into independently retrievable sub-questions. Ask "how should a cross-border DTC site do GEO" and it may split into: what GEO is, how DTC differs from marketplace sellers in GEO, the concrete action list, and the time to results. Each sub-question gets its own retrieval.

Step two: retrieve and extract. For each sub-question the retrieval system returns candidate pages, then extracts passages that answer that sub-question. This follows retrieval-augmented generation: find evidence first, then compose, rather than letting a model answer from memory. RAG was described systematically in a 2020 paper whose core idea is coupling a retriever with a generator so that output is constrained by evidence.

Step three: assemble and attach citations. Extracted passages are composed into a coherent answer, and badges are attached to the passages used. Badge count is limited, typically two to four — a structural scarcity, not something optimization can break.

These three steps explain most observations. Why is your 3,000-word article never cited? Because every section is setup, and no section answers one sub-question alone. Why does an 800-word Q&A page appear daily? Because each paragraph is a conclusion plus a condition, exactly the shape extraction wants.

Want to know how much of the ChatGPT and Perplexity answers your brand owns?

Explore TENGENCE GEO

2. The sub-query layer: one page shoots at several targets

Sub-query expansion means one page may match several sub-questions — or none.

Matching several is the good case: the same article cited across different prompts is the most stable form of visibility. The precondition is clear internal segmentation, each section answering one question. If the whole article answers "one big question", it can only match the outermost sub-query and hit rate stays low.

Our practice: split the topic into a main question plus three to five sub-questions, one subheading per sub-question, and the first sentence under each subheading states the conclusion. One article then shoots at four to six targets instead of one.

One step that gets skipped in assembly is consolidation. When several sub-answers point to the same conclusion, the system merges them into one sentence with a single badge; when sources conflict, it usually picks the more conservative, more conditional phrasing. For writers this means: instead of stating a conclusion categorically, state the conditions — the conditional version survives conflict better.

Layer What it maps to Writing requirement Test
Main question Page title and summary One sentence on what this page solves Title can be quoted alone
Sub-question One section per H2 Conclusion first, conditions after Still true when lifted out
Supporting evidence Data and sources per section Numbers with basis and year Third party can verify
Boundary note Conditions at the end of each section State what applies and what does not Does not contradict the conclusion

The most overlooked row is the boundary note. A conclusion without boundaries is risky during assembly: a sentence like "works for mid-sized companies" pulled in as a universal claim would mislead, so conditional phrasing is preferred. The more specific you write, the higher the adoption rate.

3. Why the citation went to someone else

Across projects we have reviewed, five reasons dominate:

  1. The paragraph cannot stand alone. Heavy use of "this solution", "the configuration above", "as noted earlier" leaves the slice semantically dangling.
  2. No verifiable evidence. All judgment, no number, year or source.
  3. Entity is unclear. No product name, organisation name or standard name appears, so the system cannot tell whom the page is about.
  4. Freshness looks doubtful. Version numbers, prices or policies are stale, or the page carries no explicit update date.
  5. It cannot be retrieved. noindex, login gating, or content rendered client-side means nothing is available at crawl time.

The fifth is the most wasteful and the most common. In one audit we found a client's entire help centre loading asynchronously: the crawled HTML held headings but no answers, so every FAQ was written for nothing. The fix is trivial — render content in the initial response and let collapsing be a visual state only.

The third ties into entity building. A page that only ever says "our product" or "this solution", never the full product name and owning organisation, is hard for a machine to attach to an entity. This is the same logic as knowledge panels and Wikidata work: first let the machine know who you are, then discuss whether it wants to cite you.

People who do not understand the mechanism assume "rank higher and you get cited". The actual chain is longer: be retrieved, be judged an extractable passage, be selected during assembly, and then fit within a badge quota. Break any link and your name never appears — which is why "we rank third but were never cited" is a common observation rather than a contradiction.

There is a further layer: after the answer block is shown, whether the user keeps asking, clicks a badge or bounces back to search becomes input into whether the answer was sufficient. Passages that are immediately understandable and let the user leave quickly tend to be reused; passages that pull users into a long article may gain clicks short term and lose position long term. That is not fair, but it is the mechanism.

4. Four rewriting steps

Our revision process, in order.

Step one: add a conclusion sentence to every section. Thirty to sixty words, conclusion first, reason second. Delete every "as … continues to develop".

Step two: attach evidence next to the conclusion. Number, year, basis and source immediately after the conclusion — not relegated to a data-sources list at the bottom. The bottom list serves readers; in-passage evidence serves extraction.

Step three: remove references. Replace every "this", "the above" and "the former" with a specific noun. The test: hand a random paragraph to a colleague; can they follow it?

Step four: state boundaries. Append an applicability clause to each conclusion — scale, industry, region, version. The article reads less absolute and gets cited more.

Run this as single-variable changes: first pass only the summary and key takeaways, second pass adds sources and author information, with two weeks of observation between passes and the same prompt set each time. Changing everything at once feels fast but destroys attribution — the real reason many teams rewrite long articles and cannot say which step worked.

5. Boundaries: the ceiling the mechanism sets

Three things to accept.

Display is decided by the system. Which queries get an overview is not yours to control; your job is to be cited when one appears.

Citation slots are scarce. Two to four badges means most qualified content still misses. That is structure, not quality.

Comparison queries favour third parties. Ask "which is better" and the system prefers sources that ran comparative tests. Self-described advantages lose by default; public, comparable third-party data wins.

The controllable part is clear: whether a paragraph stands alone, whether evidence sits next to the conclusion, whether the entity is explicit, whether the page can be crawled. None requires guessing at an algorithm; all are basic writing and engineering discipline.

One engineering issue sits directly on retrieval: paginated and parameterised URLs. Many sites expose the same content at several URLs (parameters, pagination, tracking codes), and which version gets crawled is uncertain. If versions differ — one missing a data block — the citation can land on the incomplete one. Canonicalise to a single version and make sure every version renders identical body content.

FAQ

Q: Will an AI overview copy my exact text?
It reuses sentence shapes and facts, usually reorganised. That is why the conclusion sentence matters — that is often the line that gets reused.

Q: Should I prepare a separate condensed version for AI?
No. Google's guidance is that AI features need no special optimization. Restructure existing pages into sections, conclusion sentences and evidence instead.

Q: Do long articles still make sense?
Yes, but the shape must change. Their value is covering multiple sub-questions, not length itself; one section per sub-question beats a structureless 3,000-word piece.

Q: How many citations does one answer block carry?
Typically two to four. Treat that as a quota, not a target you can expand by writing more.

Q: Do structured data help an AI overview cite me?
They help machines resolve topic and entity, which indirectly improves selection, but they do not guarantee citation.

Related Reading

Data Sources

  1. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Lewis et al., 2020 — the retriever-plus-generator architecture behind evidence-constrained answers.
  2. AI features and your website — Google Search Central — official guidance on how AI surfaces treat site content.
  3. Creating helpful, reliable, people-first content — Google Search Central — the evidence and people-first standard extraction rewards.
← Previous: AI Search Result Page Optimization: Treat It as Four Modules, Not One Ranking Slot Next: GEO Diagnostic Tools Compared: The Five Capabilities That Actually Matter →
Consult
QR code

WeCom