Citation Hygiene: Source Selection Rules for AI Answers

TL;DR

A defensible citation rubric scores each source 0-10 across authority, primacy, freshness, transparency, and method. Apply the rubric before publishing, set freshness budgets per claim category, and resolve conflicts deterministically. AI engines reward this with higher citation confidence.

What is citation hygiene?

Citation hygiene is the editorial practice of:

1. Selecting the right source for each claim.
2. Scoring that source against a stable rubric.
3. Disclosing the source clearly enough that a reader (or LLM) can verify it.
4. Resolving conflicts when sources disagree.

When done consistently, citation hygiene improves trust, reduces correction load, and — because LLMs increasingly evaluate retrieved sources — raises the odds of being cited back.

Why hygiene matters for AI search

AI engines do not just retrieve content; they evaluate it. Modern retrieval pipelines look at:

• Whether claims have linked evidence.
• Whether the linked evidence resolves to authoritative domains.
• Whether the page's freshness matches the claim's volatility.
• Whether disagreement among sources is acknowledged.

Pages with poor hygiene get downweighted; pages with strong hygiene get cited. Hygiene is a citation-readiness signal.

Source taxonomy

Primary sources

• Original research, datasets, government records, official documentation, statutes, court rulings, vendor changelogs, first-party announcements.
• Highest weight in the rubric.

Secondary sources

• Reputable journalism, peer-reviewed analyses, industry analyst reports, well-cited textbooks.
• Useful when primaries are unavailable or for context.

Tertiary sources

• Encyclopedias, syndicated summaries, listicles.
• Acceptable for definitions; not for time-sensitive claims.

Disallowed

• Auto-generated content with no human review.
• AI-generated summaries without primary verification.
• Anonymous forum posts as load-bearing evidence.
• Press releases used to establish independent fact (allowed only as primary for the announcement itself).

The 0-10 source rubric

Score each candidate source on five axes (each 0-2). Sum is the source score.

A score of 8-10 is strong; 5-7 is supportive; below 5 should not carry a load-bearing claim.

Freshness budget by claim category

Each claim has a half-life. Apply a budget:

If the source is older than the budget, downgrade Freshness to 0 regardless of authority.

Conflict resolution protocol

When two sources scoring ≥7 disagree:

1. Primacy wins. Prefer the primary source over the secondary, even if the secondary is more recent.
2. Recency breaks ties at the same primacy level.
3. Methodology breaks remaining ties. Reproducible methods beat anecdotes.
4. If still unresolved, present both positions, name the disagreement explicitly, and link both sources. Never silently pick one.
5. Document the decision in research notes so reviewers can audit.

Disclosure standards

• Inline link or footnote at the point of claim.
• Anchor text reflects the source name, not generic "click here".
• Date of access for time-sensitive primary sources.
• A visible "Sources" or "References" section at page end.
• Avoid nofollow on legitimately cited primary sources — it suppresses retrieval signal without preserving any privilege.

Workflow

1. Draft claim.
2. Identify candidate sources.
3. Score each via the rubric.
4. Pick highest-scoring within freshness budget.
5. Cite with proper anchor and date.
6. If conflict, apply resolution protocol.
7. Tag the page in your CMS with the highest-and-lowest scores; pages with any source < 5 trigger editor review before publish.

Common pitfalls

• Citing your own marketing as if primary. Vendor pages are primary for what the vendor said, not for independent fact.
• Stale evergreen pages. "How AI works" articles citing 2019 papers in 2026 will fail freshness.
• One mega-link. A single Wikipedia link cannot ground three different claims.
• Hidden disagreement. Quietly picking one side of a real dispute looks like sloppiness to readers and AI evaluators alike.
• Anchor-text generic. "Read more" or bare URLs hide source identity; LLMs prefer named anchors.

FAQ

Q: How many sources should a typical article cite?

A strong long-form article ties every load-bearing claim to a source, typically 5-15 distinct citations. Padding citations to hit a number hurts hygiene; under-citing leaves claims unsupported.

Q: Are vendor blog posts acceptable sources?

As primary sources for the vendor's own announcements, decisions, and product behavior — yes. As neutral evidence for industry-wide claims — no.

Q: What about AI-generated summaries as sources?

Never as load-bearing citations. They obscure the underlying primary source and may hallucinate. Use the primary source the AI summary itself was based on.

Q: How do I handle a source that was reliable but went dark?

If the original URL is gone, link the canonical archive (Internet Archive, Perma.cc, official mirror) and note the access date. Do not silently drop the citation.

Q: Does adding many citations hurt readability?

No, when they live as inline footnotes or compact superscripts. The few extra characters are far cheaper than the credibility hit of unsupported claims, and AI engines explicitly look for them.