What is 'Reference Authority' in AI search?

Reference Authority is the measurable presence of a brand in the evidence AI engines draw from when they answer buyer-intent questions. It has two surfaces: Citation Authority — structured URLs an engine cites (visible in Perplexity and Google AI Overviews) — and Model Authority — how consistently a brand appears in the training-time knowledge of engines that answer from memory (ChatGPT, Claude, Gemini, Grok).

What do 'structured citations' mean across the five AI engines?

A structured citation is a machine-readable URL that an AI engine exposes alongside its answer, letting us see exactly which source it drew from. In this pilot, only Perplexity returned a formal citations array with 2,806 URLs. ChatGPT, Claude, Gemini, and Grok answered from training memory, producing prose with bolded brand names but no link-level evidence. So 'structured citations' only exist at scale for Perplexity in this sample; the other four engines express authority through brand recall, not URLs.

Why does only Perplexity return citations in this pilot?

Perplexity is a retrieval-augmented engine: it fetches live sources for each query and exposes them as a citations array. ChatGPT, Claude, Gemini, and Grok answered these prompts primarily from their training corpora during the pilot window, returning prose with bolded brand mentions but no structured URLs. This is a real behavioral split, not an instrumentation gap.

How accurate are the findings in this pilot?

Directionally accurate, not statistically final. The pilot is 1,200 observations (80 prompts × 3 runs × 5 engines), not a full survey. Brand extraction uses bolded text + Title Case + a 500+ term stoplist, which catches many brands but is not perfect named-entity recognition. Engine behavior is a snapshot from July 2026 — any engine could switch to live retrieval and change the picture. Use the numbers as evidence of an architectural split, not as precise market shares.

Is a pilot of 1,200 observations enough to publish conclusions?

No — this is a directional pilot, not a peer-reviewed study. Findings are consistent with the underlying architectural differences between retrieval and training-data engines, but the pilot uses 80 unique prompts × 3 runs × 5 engines. A full study will expand to ~400 prompts and run quarterly. Every number here should be read as 'in this sample.'

Why is YouTube the #1 cited source in Perplexity?

Perplexity's ranker gives disproportionate weight to fresh, transcript-rich, high-engagement pages — and YouTube fits all three criteria. In this pilot YouTube appeared in 70% of queries and accounted for 11.8% of all cited URLs, more than Wikipedia, Reddit, LinkedIn, and Forbes combined.

What does this mean for PR and thought leadership?

PR and sustained thought leadership build both surfaces. Tier-1 news placements and expert bylines earn Citation Authority in Perplexity and AI Overviews today, and — because they are indexed, quoted, and syndicated — they become part of the training corpus that shapes Model Authority in ChatGPT, Claude, Gemini, and Grok over time. The mistake is measuring only citations and concluding PR 'doesn't work' for the other engines.

Pilot findings. 1,200 observations across 5 engines. Directional — see Limitations before citing.

New study · Reference Authority · Pilot · July 2026

The Reference Authority Report

Q: Can I republish these findings?

Yes. Everything on this page and the underlying dataset are licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Republish with attribution to Smart Money Media and a link back to this page. Email contact@smartmoneymedia.org for chart exports, the raw CSVs, or interview requests.

A pilot study of 1,200 answers across ChatGPT, Claude, Gemini, Grok, and Perplexity — measuring which sources AI engines actually cite, and which brands they remember when they don't cite anyone at all.

Reference Authority has two surfaces: Citation Authority (what engines link to) and Model Authority (what engines have memorized). This pilot is the first measurement of both, side by side.

Headline finding

1 of 5

engines returned structured citations

Only Perplexity exposes a machine-readable citations array (240/240 responses, 2,806 URLs). ChatGPT, Claude, Gemini, and Grok answered in prose with bolded brand names — no links. For those four engines, "reference authority" is not a backlink game; it is brand recall from training data.

Download pilot dataset (CSV)

How the five engines behaved

Engine	Obs	Avg length	URLs in text	Bolded brands
ChatGPT	240	2,192	2.9%	85.0%
Claude	240	1,758	1.2%	90.4%
Gemini	240	6,231	4.2%	93.3%
Grok	240	2,602	5.0%	94.6%
Perplexity	240	3,519	structured	94.6%

Citation Authority — what Perplexity actually cites

Top 5 domains capture ~21% of all Perplexity citations (n = 2,806 URLs, 240 queries).

Rank	Domain	Citations	Share	Query coverage
1	youtube.com	330	11.8%	70%
2	en.wikipedia.org	95	3.4%	29%
3	linkedin.com	63	2.3%	20%
4	reddit.com	62	2.2%	28%
5	forbes.com	37	1.3%	16%

Source-type mix

Brand / vendor / other1,976 · 70.4%
Social / UGC (YouTube, LinkedIn, Reddit, Medium, X)486 · 17.3%
Gov / edu141 · 5.0%
Encyclopedia (Wikipedia, Britannica)108 · 3.9%
Major news95 · 3.4%

YouTube alone accounts for 11.8% of Perplexity citations — more than Wikipedia, Reddit, LinkedIn, and Forbes combined. Major news outlets account for just 3.4%.

Model Authority — what the other four engines remember

Brands mentioned across ChatGPT, Claude, Gemini, and Grok (n = 960 answers). Rate = share of answers containing the brand.

Rank	Brand	Avg mention rate
1	Microsoft	10.5%
2	Shopify	9.5%
3	Google	8.6%
4	Google Workspace	8.1%
5	Stripe	7.4%
6	Slack	7.4%
7	QuickBooks	7.2%
8	Notion	5.1%
9	Zoom	4.7%
10	Salesforce	4.58% (tie)
10	PayPal	4.58% (tie)

The consensus set is dominated by canonical vendors with a decade of press, product coverage, and cross-linked reference material. That's the tell: Model Authority is earned by being written about across the open web for years — the same input that PR provides.

Methodology

Sample

80 buyer-intent prompts across 13 categories · 5 engines · 3 runs per cell = 1,200 observations.

Measurement

Citations parsed from Perplexity's structured array. Brand mentions extracted via bolded-text and Title Case matching, filtered against a 500+ term stoplist.

Reproducibility

Raw row-level dataset, prompt list, and extraction scripts released under CC BY 4.0 alongside the final report.

Limitations

Pilot sample. 80 unique prompts is directional. The full study will expand to ~400 prompts.
Brand extraction is heuristic. Bolded text + Title Case + stoplist catches most brands but is not NER-clean. A named-entity pass will replace it in the next wave.
Engine behavior is a snapshot. Any of the four training-data engines could enable live retrieval by default and change the picture. Findings are true for July 2026.
Perplexity ranker opacity. Why YouTube ranks first is inferred, not confirmed by Perplexity.
English-only, US-market prompts. Not generalizable to other locales.

Frequently asked

Want PR that builds reference authority?

The same third-party coverage that earns Perplexity citations today is what teaches ChatGPT, Claude, Gemini, and Grok to name you tomorrow. Smart Money Media engineers both surfaces — earned press for citations, and long-tail reference material for model recall.

Book a call

Republishing & press

Licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Republish with attribution to Smart Money Media and a link back to this page. Email contact@smartmoneymedia.org for chart exports or interview requests.

Download raw dataset (CSV)See Wave 1 — The AI Citation Gap