Intro
Most “AI-friendly content” advice stops at being selected. That is half the job.
Recent research analysing 21,143 citations across answer engines reframes Generative Engine Optimisation as evidence-container design: a page must first be eligible for source selection — through authority, recognizability, language, and domain context — and then useful for absorption into the generated answer through semantic alignment, structural legibility, and evidence density (Zhang Kai, He Xinyue & Yao Jingang, 2026).
This page operationalises that finding into six concrete design rules a content team can apply to any page in a Capston Core silo.
Audit your evidence containers
Why evidence-container design
Search engines reward pages that rank. AI engines reward pages that get used.
Two distinct behaviours sit behind every citation an engine emits. First, the engine selects a small set of candidate sources from its retrieval layer. Second, it absorbs fragments of those sources into the generated answer. A page can win the first round and lose the second — selected as a source but never quoted, paraphrased, or attributed in the visible answer. A page can also lose both rounds even when it ranks well organically, because ranking signals and citation signals only partially overlap.
Evidence-container design treats the page as a container engineered for both rounds. The selection layer asks: is this source eligible? The absorption layer asks: is this source useful?
Both layers must be designed deliberately. Neither is automatic.
The selection layer: be eligible
The research isolates four eligibility drivers that determine whether a page enters the candidate set at all.
- Source authority — domain trust signals, brand recognizability, and verifiable provenance. Premium brands carry recognizability that anonymous publishers do not, but only when the domain is identifiable from the URL and the page itself.
- Language match — the page’s language must match the prompt language, including locale nuance. A French page rarely surfaces in an English answer, and vice versa, unless the prompt is explicitly multilingual.
- Domain context — engines weigh sources by topical fit. A page about hospitality scoring is more likely to be selected for a hospitality prompt than a page on the same domain about general AI strategy.
- Recency signals — visible publication and update dates, especially when the content is time-sensitive. Engines downweight stale-looking sources on fast-moving topics.
Eligibility is a threshold, not a ranking. Pass it and the page enters the pool. Fail it and the rest does not matter.
This is the layer the Capston Core methodology addresses first, before any content rewrite.
The absorption layer: be useful
Once selected, the page competes on usefulness. The research identifies extractable evidence genres that engines reuse disproportionately: definitions, numerical facts, comparisons, procedural steps. Pages that surface these genres in modular, labelled blocks are absorbed more often than equivalent pages where the same facts are buried in prose.
Three usefulness drivers matter most:
- Semantic alignment — the page’s wording matches the wording the user is likely to use in the prompt. Synonyms help; aligned phrasing helps more.
- Structural legibility — headings, lists, tables, and short paragraphs make extraction cheap. Walls of text cost the engine compute and lose to better-structured alternatives.
- Evidence density — the number of citable units per thousand words. High-influence pages are longer than average, but only because they pack more extractable evidence, not more filler.
Designing for absorption is not about writing for machines. It is about respecting how engines parse, score, and quote.
Six design rules
A practical checklist a content team can apply page by page.
- Lead with a definition. Open with a one- to two-sentence definition of the page’s primary entity, in the same wording the audience uses. This is the single most-quoted unit on most pages.
- Surface numerical facts in their own line. Counts, percentages, dates, thresholds — give each its own sentence or row. Numbers buried inside long paragraphs are extracted less reliably than numbers that stand alone.
- Use comparisons explicitly. When two options, methods, or competitors are discussed, label the comparison and structure it (table, parallel bullets, before/after). Engines reuse comparison structures verbatim.
- Write procedural steps as numbered lists. Any “how to” content should appear as discrete, numbered steps with a verb-led first word. Steps embedded in prose are absorbed inconsistently.
- Keep modules short and labelled. Each H2 or H3 should cover one extractable idea. A module that mixes a definition, a procedure, and an opinion is harder to absorb than three separate modules.
- Anchor every claim to a named source. Where a fact comes from research, a study, an audit, or an internal dataset, name it inline. This raises absorption probability and protects against hallucination attribution.
- Make the domain identity visible on the page. Brand name, author, and entity definition should appear in the first 200 words. Engines that paraphrase still need to attribute.
- Date the page in plain text. A visible “Updated [month, year]” line in the body, not only in metadata, helps recency-sensitive selection.
Six is the floor. Eight is the working maximum before a checklist stops being followed.
How this fits into Capston Core
Evidence-container design sits between Capston’s measurement layer and its content layer. The AI answer evidence layer tells the team which pages are selected, which are absorbed, and which are ignored. The Capston Core methodology sequences the work across five stages. The Capston QA standards keep the six rules applied consistently across clients, languages, and partners.
A page that passes the six rules is not guaranteed to be cited. A page that fails them is reliably ignored.
→ Back to Capston Core
FAQ
Is evidence-container design different from SEO?
Yes. SEO optimises for ranking in a results page. Evidence-container design optimises for selection and absorption inside a generated answer. The two overlap on basics like authority and freshness, but diverge on structure, density, and modularity.
Do these rules apply to all page types?
The rules apply to any page intended to be cited: service pages, methodology pages, audit deliverables, knowledge-base entries. They apply less to transactional pages where citation is not the goal.
How long should an evidence-container page be?
Long enough to hold the evidence, short enough to stay scannable. Research shows high-influence pages run longer than average, but only because they carry more extractable units, not more filler.
What is the single biggest mistake content teams make?
Burying the definition. Most pages open with brand language or a hook, then define the topic three paragraphs in. By then the engine has scored the page on a weaker opening.
Reference
Zhang, Kai; He, Xinyue; Yao, Jingang (2026). From Citation Selection to Citation Absorption: An Empirical Study of Generative Engine Optimisation. arXiv:2604.25707v2. Analysis of 21,143 citations across answer engines.
Final CTA block
Audit how your pages perform as evidence containers.
Audit your evidence containers
Read the methodology