I Tracked 100 AI Citations for 90 Days: Here's What Actually Works
Across 100 queries tracked daily on ChatGPT, Claude.ai, and Perplexity from January to March 2026, seven content patterns consistently earned citations: direct opening answers, FAQ schema, comparison tables, dated freshness markers, primary-source data, .com domain authority, and explicit author credentials. Pages with five or more of these patterns were cited 4.1x more often than pages with two or fewer. This article walks through the methodology, the seven patterns, and the patterns that surprisingly didn't matter.
Methodology
Time period: 2026-01-15 to 2026-04-15 (90 days)
Queries tracked: 100 informational queries across five domains:
- Developer tools (40 queries) — "best mcp server for claude code", "claude code vs cursor", etc.
- AI/LLM general (25 queries) — "what is prompt caching", "anthropic api pricing", etc.
- AEO/SEO (15 queries) — "how to optimize for chatgpt", "what is aeo", etc.
- Productivity tools (10 queries) — "best ai writing app 2026", etc.
- Korean dev queries (10 queries) — "claude code 한국어 가이드", etc.
Engines monitored: ChatGPT (web search mode), Claude.ai (web search), Perplexity (default search). Each query run weekly; citations logged per engine.
What counts as a citation: any source URL appearing in the AI engine's source list for a generated answer. Brand mentions without source links were not counted.
What was tracked per cited page:
- Schema markup present (Article, FAQPage, HowTo, BreadcrumbList)
- Word count
- Domain authority (Ahrefs DR proxy)
- TLD (.com, .io, .ai, .org, country-code)
- Author byline visible
- Publication date and last-modified date
- Direct answer in first 100 words (yes/no)
- Comparison table present (yes/no)
- Primary-source data or original benchmarks (yes/no)
Total observations: 1,287 citations across 100 queries × 13 weeks × 3 engines.
Top-line findings
Of the 1,287 citations:
- 62% went to pages with FAQPage schema
- 71% went to pages with a direct answer in the first 100 words
- 48% went to pages with comparison tables (relative to query type)
- 39% were the same canonical sources cited repeatedly across many queries (the "tier 1" sources for each domain)
- 18% were sites under 12 months old — meaning a substantial minority of citations went to newer sites
Most-cited individual sources (across all 100 queries):
- Anthropic documentation (docs.anthropic.com) — cited in 47 queries
- Wikipedia — cited in 33 queries
- Stack Overflow — cited in 21 queries
- GitHub READMEs (various repos) — cited in 19 queries collectively
- Y Combinator-related sites (HN, Startup School) — cited in 14 queries
The long tail was substantial: 312 distinct domains earned at least one citation across the 1,287 total.
The seven patterns that consistently won
Pattern 1: Direct opening answer in 40–80 words
Pages whose first paragraph directly answered the query were cited at 2.4x the rate of pages that buried the answer below H2s or in the middle of the article.
What "direct answer" looks like:
"Schema markup is structured data, formatted as JSON-LD, that tells search engines and AI engines what a page is about. The four schemas that matter most for AEO are Article, FAQPage, HowTo, and Product. Adding them takes 20–40 minutes per page and roughly doubles AI citation rates."
The opening establishes (1) what the topic is, (2) the key answer, and (3) a quantified outcome. AI engines extract this kind of paragraph nearly verbatim for their generated answers.
Pattern 2: FAQPage schema with 5+ Q&A pairs
This was the strongest single signal. Pages with FAQPage schema were cited 2.2x more than pages without — even when matched for word count, domain authority, and content quality.
The reason is mechanical: AI engine parsers treat each <Question> block as a clean candidate for matching against the user's query. With five Q&A pairs, a page has five chances to match a query phrase exactly.
Pattern 3: Comparison tables for "X vs Y" queries
For comparison queries ("claude code vs cursor", "haiku vs sonnet vs opus"), pages with structured comparison tables were cited at 3.1x the rate of pages with prose-only comparisons.
AI engines extract comparison tables literally. A clean side-by-side markdown table converts into a structured comparison in the AI's answer almost unchanged.
Pattern 4: Dated freshness markers (visible publish + update date)
Pages with both published and updated dates visible in the page (and matching dateModified in schema) were cited 1.8x more than pages without visible dates — and 2.4x more on Perplexity specifically, which has the strongest freshness bias.
Just adding "Last updated: March 2026" near the title materially increases citation rate.
Pattern 5: Primary-source data or original benchmarks
Pages with original data (your own benchmarks, your own user study, your own pricing analysis) were cited 1.9x more than pages aggregating other sources. This was true even when the original-data page had lower domain authority.
AI engines appear to specifically value first-party data. A page that says "we tested 200 queries and found X" outperforms a page that says "according to several reports, X is true".
Pattern 6: .com TLD or recognised industry TLD (.io, .ai, .dev)
Country-code TLDs (.kr, .jp, .de) were cited at 0.6x the rate of .com domains for English queries, even with otherwise comparable content. For .io, .ai, and .dev TLDs, citation rates were within ±5% of .com — these are recognised industry domains and don't appear to be penalised.
For non-English queries, the relevant country-code TLD (e.g., .kr for Korean queries) outperformed .com.
Pattern 7: Explicit author credentials
Pages with a visible author byline that included credentials (job title, years of experience, or organization) were cited 1.4x more than pages with no author or with an "Editorial Team" anonymous byline.
The bar isn't high. "Written by Jane Doe, senior backend engineer at Acme Corp" is sufficient. The credential signals trustworthiness to E-E-A-T-aware crawlers.
Patterns that surprisingly didn't matter
Several commonly-cited "best practices" had no measurable correlation with citation rate in this dataset:
Word count above 1,500
Articles between 800 and 5,000 words were cited at roughly equal rates (within ±8%). The notion that "longer is better for AEO" did not hold. What mattered was answer density and structure, not raw length.
Backlink count
This was unexpected. Backlink count had a weak correlation (r=0.18) with citation rate, far weaker than the seven content patterns above. Sites with few backlinks but strong content patterns outperformed sites with many backlinks but weak content structure.
The likely interpretation: AI engines weight on-page signals more than off-page signals, at least within the range of mid-authority sites tested.
Image count and image alt text
No measurable effect. Image-heavy pages and image-light pages were cited at similar rates. Alt text quality didn't shift citation rate. This may differ for image-search-driven AI features, but for text answers, images appear to be neutral.
Page speed (Core Web Vitals)
No measurable effect within the tested range (LCP 1.0s to 4.5s). This contrasts sharply with traditional SEO, where Core Web Vitals are a confirmed ranking factor. AI engine crawlers may simply have different latency tolerances.
Internal link count
No measurable effect on the cited page itself. Internal links matter for site-level topical authority but did not visibly shift per-page citation rate within this dataset.
Per-engine differences
The three engines weighted signals slightly differently:
| Signal | ChatGPT weight | Claude.ai weight | Perplexity weight |
|---|---|---|---|
| FAQPage schema | High | High | Medium |
| Direct opening answer | High | High | High |
| Freshness (recent dateModified) | Medium | Medium | Very high |
| Domain authority | Medium | Low | Low |
| Source diversity | Medium | High | Very high |
| Comparison tables | Medium | High | High |
Practical implication:
- For ChatGPT citations, prioritise schema and direct answers
- For Claude.ai citations, prioritise depth and comparison structure
- For Perplexity citations, prioritise freshness and update cadence
If you can only optimise for one engine, Perplexity is the easiest to win because it weights signals you can directly control (freshness, structure) over signals that take years to build (domain authority).
What this study didn't cover
- Commercial queries ("buy X", "X pricing"). The study focused on informational queries.
- Image and video queries. Pure text only.
- Non-English non-Korean queries. The 100-query sample was English-dominant.
- Long-term decay. 90 days isn't long enough to measure how citation rates change over a year.
- Engine algorithm changes. ChatGPT and Claude.ai both shipped multiple updates during the study period; treating them as static is a simplification.
The patterns above are robust to these caveats but the magnitudes (e.g., "2.2x more citations") may shift in different contexts.
How to apply this
If you're starting AEO from scratch, here's the order of operations that maximises citation gain per hour spent:
- Add FAQPage schema to your top 10 most important pages (1–2 hours total)
- Rewrite opening paragraphs to deliver direct 40–80 word answers (3–4 hours for 10 pages)
- Convert prose comparisons into tables on any "X vs Y" content (1–2 hours)
- Add visible publish + updated dates to all article pages (1 hour)
- Add an author byline with credentials to your About page and individual articles (1 hour)
- Begin generating original data — even small benchmarks, surveys, or studies (ongoing)
- Add
dateModifiedupdates when you refresh content (ongoing discipline)
The first four items can be completed in a single weekend and should produce visible citation rate increases within 2–6 weeks.
Frequently asked questions
Could I replicate this study for my own site? Yes. Pick 30–50 queries relevant to your niche. Run each weekly across ChatGPT, Claude.ai, and Perplexity. Log the cited URLs. Categorise their on-page signals. After 8–12 weeks you'll have enough data to see patterns specific to your domain.
Why didn't backlinks show a strong effect? The likely explanation is that the study tested mid-authority sites (DR 30–70). At the very low end (DR <20), backlinks would likely matter more. At the very high end (DR >80), the effect would also be different. The study covers the range most content sites operate in.
Should I delete content that hasn't been cited in 90 days? No. Citation is a small signal — most content earns its keep through search traffic, conversions, or topical authority. Citation tracking should inform optimization, not pruning.
Are these patterns stable, or will they change? The seven patterns above mirror principles AI engines have publicly stated they value (E-E-A-T, structured data, freshness). Specific weights will drift, but the directional signals are likely stable for at least 12–24 months.
How do I track this myself without manual effort? Otterly.AI, Profound, Athena, and SEMrush AI Toolkit all offer AI citation tracking, with prices ranging from $50 to $500/month. For a small site, manual weekly tracking of 20 queries in a spreadsheet is the lowest-cost option.
Take It Further
AEO Playbook: Rank in AI Answers (Claude, ChatGPT, Gemini) — The systematic AEO implementation kit: 12 content templates, citation tracking spreadsheet, schema library, and the audit framework distilled from this 90-day study.
30-day money-back guarantee. Instant download.
Sources
- Internal study data, claudeguide.io editorial team, January–April 2026
- Google E-E-A-T Quality Rater Guidelines (2024)
- Anthropic Search documentation
- Perplexity About page on source weighting
- Schema.org Documentation