There is no single universal source-selection formula
AI-assisted search products do not all use the same index, retrieval process, model or citation interface. Their behavior also changes over time. Any explanation that presents one fixed ranking formula for every system is too confident.
A more useful approach is to study the observable stages that a publisher can influence: discovery, eligibility, selection and presentation. This model does not claim access to a platform's internal weights. It gives teams a disciplined way to diagnose why a useful page may or may not appear.
A four-stage model
- Discovery. The system or one of its retrieval partners must know the URL or source exists. Sitemaps, internal links and external references can support discovery.
- Eligibility. The source needs accessible content, a stable canonical URL and enough visible information to interpret the page.
- Selection and synthesis. For a given query, the system chooses information it considers useful, relevant and supportable. Several sources may be combined.
- Presentation. The interface decides whether to display a link, citation marker, source card or no visible attribution at all.
A page can be retrieved without being cited, and it can influence an answer without producing an observable referral. These are different events.
What makes a source easier to use?
| Signal | Strong implementation | Common weakness |
|---|---|---|
| Access | Useful server-rendered HTML and a 200 response | Empty shell, blocked resource or unstable redirect chain |
| Identity | Clear organization, author, service and market relationships | Several names or offers with no explicit connection |
| Answer completeness | Direct statement supported by method, example and constraint | Marketing claim with no usable explanation |
| Evidence | Primary records or relevant independent corroboration | Self-referential claims repeated across owned pages |
| Freshness | Date and update context where the topic changes | Undated statistics or obsolete product details |
| Consistency | Visible facts, metadata and external profiles agree | Conflicting addresses, names, services or canonical URLs |
Write passages that remain accurate when extracted
A useful passage should survive outside the visual layout. Define the subject by name, include the essential condition and avoid relying on “this,” “it” or a nearby graphic for meaning. Tables work well for explicit comparisons when every row has a clear label.
Weak: “We cover everything nationwide.”
Stronger: “The service supports scheduled B2B deliveries across the continental United States; same-day availability depends on origin, destination and booking time.” The second statement is narrower, easier to verify and includes a decision-relevant limitation.
Why citations vary between tests
- The wording and specificity of the question change retrieval needs.
- Market, language, account state and product mode can change available sources.
- Indexes and documents change between capture dates.
- A system may synthesize a response without exposing every source used.
- Some answers require current information; others can rely on stable background knowledge.
For that reason, a single screenshot is evidence of one observation, not a durable visibility rate. A rate requires a fixed query set, repeated collection and a stated denominator.
A repeatable source test
- Freeze the exact question set before collection.
- Record platform, product mode, language, market, date and account context.
- Capture the full answer and every visible source.
- Classify brand mention, cited brand mention and third-party source separately.
- Open each cited URL and confirm it supports the associated statement.
- Repeat on an agreed cadence without silently changing the query set.
A page-level review checklist
- One preferred canonical URL returns a useful 200 response.
- The page has a descriptive title and one primary heading.
- The organization, service and audience are named in visible text.
- Important claims include evidence, date or scope.
- Structured data matches what a visitor can read.
- The page links to relevant primary material and related internal pages.
- Limitations are explicit where they change the decision.
Primary references and limits
Search-engine documentation explains discovery, crawling, indexing and canonical selection, but it does not disclose a universal citation algorithm for every AI product. Use official documentation for implementation facts and treat live answer tests as dated observations.
Where to go next
Use the Nexus AI Visibility Framework to place this topic inside a complete measurement system, or review the AI Visibility Audit scope.
13 Jul 2026 — Expanded with an independent structure, examples, implementation guidance and primary references.