"Research shows" is a false authority signal; "(SparkToro, 2024)" with a link is a real one. Back every significant claim with a verifiable, dated statistic from a named source — citation density measurably lifts AI inclusion.
By Christopher Foster-McBride, Founder, AI Knowledge Signal · Reviewed 2026-10-13
Why it matters
AI systems reward verifiability. In the GEO study (Aggarwal et al., KDD 2024), rewriting a source to add quotations, statistics or citations to other sources were the most effective of nine tested methods, raising that source’s visibility in generated answers by up to ~40%; keyword stuffing did not help. The study measured how much of an answer drew on an already-retrieved source — not whether a page gets retrieved or clicked. Vague attribution — “research shows”, “experts say”, “studies suggest” — gives a reader nothing to check and a model nothing to attribute.
The discipline is simple: every significant claim gets a verifiable, dated statistic from a named source, placed near the claim and classified by source tier.
What good looks like
Claims are evidenced, named, dated, and tiered:
Every significant claim is backed by a verifiable, sourced, dated statistic.
The organisation and year are named — "(SparkToro, 2024)" — and the source is hyperlinked where possible.
Primary sources (research papers, official reports, original datasets) are preferred over blog coverage.
Sources sit next to the relevant claim, not only in a list at the end.
Step by step
Attach a sourced statistic to each claim
For every significant claim, find a specific, dated statistic that supports it. “Search is changing” becomes “58.5% of US Google searches ended without a click to any website (SparkToro & Datos, 2024).”
Name the organisation and year, and link it
Always state the source organisation and year — "(McKinsey, 2024)" — and hyperlink the source name. Never use "research shows" or "experts say" without specifics; those are false authority signals. When you repeat someone else’s figure, repeat its basis too (sample, date range, what was measured).
Prefer primary sources and classify by tier
Link the research paper, official report, or dataset rather than a blog about it. Tag each source Tier 1 (primary), Tier 2 (credible synthesis), or Tier 3 (tertiary) so the strength of each claim is transparent.
Place the source next to the claim
Put the citation immediately beside the claim it supports, not only in a references list at the bottom. Co-location lets a model bind the evidence to the claim when it extracts the passage.
Combine several evidence types
Use quotations, statistics and source citations together where the content supports them — they were the three strongest of the nine methods in the GEO study. The study tested combinations, but don’t expect the individual gains to add up.
Worked examples
1 Before / after: vague vs. cited
Before (weak)
After (cited)
"Search is changing."
"58.5% of US Google searches ended without a click to any website (SparkToro & Datos, 2024)." — linked
"Research shows it works."
"Adding quotations, statistics or citations raised a source’s visibility in generated answers by up to ~40% (Aggarwal et al., KDD 2024)." — linked
"Most companies use AI now."
"72% of organisations surveyed had adopted AI in at least one business function (McKinsey, The state of AI, 2024)." — linked
2 What the GEO study tested
Nine rewrite methods were applied to sources an engine had already retrieved, and the change in each source’s share of the generated answer was measured. Re-read Table 1 of the paper before quoting a per-method figure.
Prefer data from the last ~24 months unless it’s a foundational source; flag older figures (Phase 6.2).
Limits of this advice
The GEO results come from one study with a simulated engine and Perplexity. A 2025 benchmark (C-SEO Bench, Puerto et al., NeurIPS 2025) found most such rewrite methods largely ineffective; treat the ~40% figure as one study’s upper bound. Cite sources because readers can check them, not for a guaranteed lift.
A citation to a weak or misquoted source is worse than none. Re-read every source (Phase 6.2).
Checklist
Every significant claim has a verifiable, dated statistic
Source organisation and year named, and hyperlinked
Primary sources preferred over blog coverage
Each source classified Tier 1 / 2 / 3
Sources placed next to their claims, not only at the end
Quotations, statistics and source citations used together where the content supports them
