# What Actually Predicts AI Search Citations: Independent Studies Compared

Multiple independent studies have now measured what predicts AI search citations. They disagree on domain authority, agree on content alignment, and reveal that citation is a source-selection decision — not a ranking signal.

Canonical URL: https://machinerelations.ai/research/what-predicts-ai-search-citations-independent-studies-2026
Published: 2026-08-08
Research type: Reference

## Source Body

Content-prompt alignment is the strongest page-level predictor of AI search citations, with an effect size roughly three times larger than any other on-page signal. But the studies measuring this disagree sharply on whether domain authority matters, whether structured data helps, and which content formats get cited most. The disagreements are not noise — they reflect real differences in methodology, engine coverage, and what each study actually measured.

This analysis compares the major independent citation studies published through August 2026, maps where they converge and diverge, and contextualizes both against Machine Relations Index citation rate data measured across six AI engines and 6,020 domains.

## What the Studies Measured

Five independent research efforts have now published quantitative citation analysis with enough rigor to compare:

- **DiscoveredLabs** analyzed [2 million AI citation observations across 10,000 pages](https://discoveredlabs.com/research/what-drives-ai-citations) over six months, covering ChatGPT, Claude, Google AI, and Gemini. They computed 60+ structural, alignment, and infrastructure features per page and applied nine robustness checks including stability-selection Lasso and Double Machine Learning.

- **FactoryJet** measured [58 pages from 232 Google AI Overview citations](https://factoryjet.com/ai-citation-study) across 12 commercial e-commerce queries. They deliberately excluded publishers and marketplaces to focus on competitive pages, and measured rendered HTML rather than source code.

- **GetAISearchScore** analyzed [658 citations across 485 domains and 30 queries](https://getaisearchscore.com/blog/ai-citation-analysis-658-sources-30-queries) on Perplexity's sonar-reasoning-pro model. They ran three replicates per query at temperature=0 and scored 441 domains on structural factors.

- **Zyppy/Signal** published a [meta-analysis of 54 peer-reviewed studies, experiments, and patents](https://signal.zyppy.com/p/ai-citation-ranking-factors) across ChatGPT, Gemini, and Perplexity. They scored each factor on repeatability, evidence strength, and official documentation support.

- **SIGI** published an [observational analysis of 22 on-page metrics](https://generativeintelligence.institute/publications/sigi-2026-037) and their structural correlation with AI citation frequency.

Each study asked the same question with different scope, different engines, and different analytical frameworks. The result is a literature that converges on some factors and diverges instructively on others.

## Where the Studies Agree

Three findings hold across every major study regardless of methodology:

**Content-prompt alignment is the highest-leverage page-level signal.** DiscoveredLabs measured this at a standardized regression coefficient of +0.37 — selected in 100% of 200 bootstrap samples and significant across three of four engines tested. Zyppy's meta-analysis rated query-answer match at 9.2/10. GetAISearchScore found a [62× citation rate difference](https://getaisearchscore.com/blog/ai-citation-analysis-658-sources-30-queries) between same-topic and cross-topic domains. FactoryJet's list density finding (median 110 list items per cited page) reflects the same principle: pages that chunk information into retrievable units aligned to buyer questions earn citations.

**Traditional on-page SEO factors have weak or no direct effect on citation.** Schema markup showed no significant effect in DiscoveredLabs' controlled model. GetAISearchScore found structural scores correlated with citations at r = 0.009 (functionally zero). FactoryJet found that FAQPage schema appeared on only 41% of cited pages despite widespread recommendation. Core Web Vitals showed no significant citation effect after domain controls were applied. [Haide Digital's LLM ranking factor analysis](https://haide.digital/resources/llm-ranking-factors) reached the same conclusion: traditional on-page SEO metrics do not transfer to AI citation selection. [Averi's citation benchmark analysis](https://resources.averi.ai/benchmarks/ai-search-citation-benchmarks) across AI search engines reinforced this pattern — the factors that predict AI citation are structurally different from those that predict search ranking.

**llms.txt files have no measurable citation impact.** Google's official AI optimization guide [states](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) that "Google Search doesn't use them." Ahrefs found 97% of llms.txt files received zero requests across 137,000 domains. Zyppy rated llms.txt at 2.0/10. No study found evidence that any AI engine consults these files during citation selection.

## Where the Studies Disagree: The Domain Authority Question

The sharpest disagreement across studies is whether domain authority predicts AI citation:

| Study | Scale | Engines | Domain Authority Finding |
|---|---|---|---|
| DiscoveredLabs | 2M citations, 10K pages | ChatGPT, Claude, Google AI, Gemini | Mean absolute SHAP value of 0.38 — 6× more influential than strongest page-level feature |
| FactoryJet | 58 pages, 232 citations | Google AI Overviews only | "Authority buys organic position. It did not buy citation here." |
| GetAISearchScore | 658 citations, 485 domains | Perplexity only | Only significant predictor, but explained just 2.2% of variance |
| Zyppy meta-analysis | 54 studies synthesized | Multiple engines | Rated 5.0/10 — weak average correlation across studies |

This disagreement is not random. It reflects three methodological differences that the original studies did not reconcile:

**1. Engine coverage matters.** FactoryJet measured only Google AI Overviews. GetAISearchScore measured only Perplexity. DiscoveredLabs measured four engines over six months. MRI citation rate data, which tracks six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, and Perplexity) across 6,020 domains, shows that citation patterns diverge substantially by engine. The same domain can appear at different citation rates on different engines because each engine's retrieval architecture makes independent source-selection decisions.

**2. Within-tier vs. cross-tier comparison.** FactoryJet excluded publishers and platforms, comparing only competitive service pages. Within a narrow authority tier, domain authority cannot differentiate — every competitor has a similar backlink profile. DiscoveredLabs compared across the full domain spectrum, where the difference between a 10-DA site and a 90-DA site is structural. Both findings can be true simultaneously: authority provides a floor that gates entry, but does not determine citation rank among peers who clear the floor.

**3. Citation rate vs. citation occurrence.** DiscoveredLabs measured how often a domain gets cited (rate). FactoryJet measured which pages appeared in a citation list (occurrence). These measure different things. A high-DA domain may earn a higher citation rate across thousands of queries while still failing to dominate any single query's citation list.

## What MRI Data Reveals That Page-Level Studies Miss

Page-level studies measure what makes one page more citable than another page on the same domain. The Machine Relations Index measures something different: which sources AI engines select when answering real buyer questions, and how consistently they return to those sources over time.

MRI citation rate data across 6,020 measured domains reveals a pattern that page-level analysis cannot capture: **source type predicts citation behavior more reliably than any page-level or domain-level metric.**

Market databases (G2, Crunchbase), analyst firms (Gartner, Forrester, Deloitte), and market research publishers (Grand View Research, MarketsAndMarkets) maintain high citation rates not because of superior content-prompt alignment or structured data — but because AI engines classify them as authoritative source types for specific question categories. A market database gets cited on competitive landscape questions. An analyst firm gets cited on strategic planning questions. A press distribution network gets cited on event and announcement questions.

This is visible in the MRI's [source-role classification](https://machinerelations.ai/research/ai-answer-engine-source-type-citation-patterns-2026): domains serving the same source role cluster at similar citation rates regardless of their page-level optimization. The variation between source roles exceeds the variation within them.

The implication for the citation factor literature: studies that control for page-level features but not for source-role classification will attribute to domain authority what actually belongs to source-type selection. And studies that measure only one engine or one vertical will miss the cross-engine divergence that makes citation a multi-surface problem rather than a single optimization target.

## Content Freshness Preferences Differ by Engine

DiscoveredLabs produced the most granular engine-specific freshness data:

| Engine | Median Cited Content Age | Share Under 6 Months |
|---|---|---|
| Claude | 5.1 months | 60% |
| Google AI | 6.0 months | — |
| Gemini | 7.8 months | — |
| ChatGPT | 8.0 months | 40% |

Claude's strong recency preference and ChatGPT's tolerance for older content create a practical tension: optimizing freshness for one engine's citation threshold may not affect another's. DiscoveredLabs recommended calibrating to the most demanding engine to maintain competitiveness across all of them.

Zyppy rated freshness at [7.0/10](https://signal.zyppy.com/p/ai-citation-ranking-factors), noting that its importance varies by query type — current events queries demand fresh content, while definitional queries tolerate older sources.

## Content Format and Structure: What Holds Up

Across studies, a few structural patterns survived controlled analysis:

**FAQ sections.** FactoryJet found visible FAQ sections on 72% of [cited pages](https://factoryjet.com/ai-citation-study). DiscoveredLabs measured FAQ sections at β = +0.07 (the strongest structural signal on third-party pages). Zyppy rated AI-ready structure at 8.6/10. The pattern is consistent: extractable question-answer pairs aligned to sub-queries help.

**Content length.** FactoryJet found a median of 2,813 rendered words among cited pages, with the largest cluster between 2,000 and 3,500 words. DiscoveredLabs found a positive effect (β = +0.13) for page length. Only 3 of 58 cited pages exceeded 6,000 words, contradicting common advice to write very long pages. [DigitalApplied's 2026 data study](https://digitalapplied.com/blog/ai-search-citation-ranking-factors-2026-data-study) confirmed this middle-length finding — cited pages tend to be comprehensive but not exhaustive.

**Answer position.** DiscoveredLabs found that the paragraph most often cited sits at median depth 0.36 (top third of the page). Zyppy rated "answer near the top" at 8.8/10, noting Google's Gemini uses a strict retrieval cap per URL. The pattern: put the answer first.

**Factual specificity.** Zyppy rated "factually specific" at 8.3/10 and "explicit phrasing" at 8.1/10. Hedged or vague statements reduce citation probability. AI engines prefer definitive claims backed by referenced sources.

## Citation Instability Is Measurable

GetAISearchScore documented a finding that complicates all other results: citation lists are probabilistic, not deterministic.

Even at temperature=0, only [47.5% of Perplexity citations](https://getaisearchscore.com/blog/ai-citation-analysis-658-sources-30-queries) appeared consistently across three replicates. 29.3% appeared in only one of three runs. The mean Jaccard similarity between citation lists was 64.4%, ranging from 18% to 100%.

Any single-observation study of citation factors measures a noisy signal. [Indexably's AI citation research](https://indexably.io/blog/ai-citation-research) also documented this instability, noting that citation lists shift between runs even with identical queries. Studies with larger sample sizes and longer observation windows — DiscoveredLabs' six months, MRI's continuous daily measurement — are structurally more reliable than snapshot analyses. Not because of better methodology per se, but because they average out citation instability that single-run studies cannot detect.

[AI+Automation's research on the SEO floor](https://aiplusautomation.com/research/the-seo-floor) adds a practical dimension: there appears to be a baseline level of SEO competence required for pages to enter the citation candidate pool at all, below which no amount of alignment or structure helps. This aligns with the DiscoveredLabs finding that domain authority provides a structural floor rather than a scoring gradient.

## How Machine Relations Contextualizes These Findings

The independent citation studies collectively confirm a shift that Machine Relations has been measuring since the Machine Relations Index launched: citation selection by AI engines operates on different inputs than traditional search ranking.

Three properties distinguish citation from ranking:

1. **Source-type selection** determines which domains enter the citation candidate set. Page-level optimization operates only within that set. No amount of content-prompt alignment gets a personal blog cited as a market database.

2. **Multi-engine divergence** means that citation is not one optimization problem but six (or more). Optimizing for one engine's retrieval architecture may reduce performance on another.

3. **Citation instability** means that any single measurement is a sample, not a census. Continuous measurement over time — not a one-time audit — is the only way to detect whether a change in citation behavior reflects real improvement or normal variance.

These properties are why Machine Relations treats citation as a source-relationship discipline rather than a content-optimization exercise. The page-level studies are right that alignment matters. The domain-level studies are right that authority provides a floor. Both miss that the decisive variable is whether AI engines classify your domain as the right source type for the question being asked.

## FAQ

### Does domain authority matter for AI citations?
It depends on the comparison. Across the full domain spectrum, authority provides a structural floor — low-authority domains rarely get cited. Within a tier of similar-authority domains, authority does not differentiate. The independent studies disagree because they measured different comparisons, not because one was wrong. [Attrifast's analysis comparing AI citations to backlinks](https://attrifast.com/blog/ai-citations-vs-backlinks) found that link-based authority metrics correlate weakly with citation frequency, supporting the view that citation and ranking operate on different signals.

### What is the single most important factor for getting cited by AI search engines?
Content-prompt alignment — how closely your content matches the specific question the AI engine is answering. [DiscoveredLabs](https://discoveredlabs.com/research/what-drives-ai-citations) measured this at a standardized effect of +0.37, roughly three times larger than the next strongest page-level signal. This held across four engines and survived nine robustness checks.

### Does structured data or schema markup help with AI citations?
Independent studies consistently found no significant direct effect. DiscoveredLabs reported no significant effect after domain controls. GetAISearchScore found structural scores correlated at r = 0.009. [Google's official guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) states: "Structured data isn't required for generative AI search."

### Why do different AI engines cite different sources for the same question?
Each engine uses a different retrieval architecture, different grounding corpora, and different freshness thresholds. Claude cites median content age of 5.1 months; ChatGPT tolerates 8.0 months. Google AI Overviews draws from its own search index; Perplexity runs its own web crawl. Citation is not one system — it is six independent source-selection processes running in parallel.

*Last updated: August 8, 2026. Analysis based on Machine Relations Index v2 citation rate data and independent studies published through August 2026.*

## Attribution

This research is published by Machine Relations Research, the research program of machinerelations.ai — the public research and standards initiative that publishes the glossary, research, evidence, and measurements for the Machine Relations discipline. Provenance and editorial standards: https://machinerelations.ai/about

## Machine-readable related links

### Related concepts

- [Machine Relations Index (MRI)](https://machinerelations.ai/glossary/machine-relations-index)
- [Machine Relations (MR)](https://machinerelations.ai/glossary/machine-relations)
- [Content Freshness](https://machinerelations.ai/glossary/content-freshness)
- [AI Citations](https://machinerelations.ai/glossary/ai-citations)

### Supporting research

- [AI citation measurement methodologies compared — why different indices rank the same publishers differently](https://machinerelations.ai/research/ai-citation-measurement-methodologies-compared-2026)
- [The Domain Authority Inversion: Why AI Engines Cite Mid-Tier Vertical Publications Over Tier-1 Press](https://machinerelations.ai/research/domain-authority-inversion-ai-citations-vertical-publications-2026)
- [AI Search Citation Volatility: What Five Independent Studies Reveal About Weekly Stability](https://machinerelations.ai/research/ai-search-citation-volatility-weekly-stability-2026)
- [Cross-Domain Brand Authority vs Backlinks: What Actually Drives AI Citation Selection](https://machinerelations.ai/research/cross-domain-brand-authority-vs-backlinks-ai-citations-2026)

### Framework context

- [Machine Relations Index](https://machinerelations.ai/index)
- [Machine Relations Stack](https://machinerelations.ai/stack)
- [Evidence Base](https://machinerelations.ai/evidence)