Earned Media analysed 760 AI search responses across Australia and the US, more than 5,000 citation events in total, to test what actually drives visibility in ChatGPT, Gemini, Claude and Perplexity. Across both markets, unique referring domains correlated significantly higher with AI citations than Google rankings or Domain Rating.
The study, at a glance
The Earned Media AU/US GEO & Backlink Study
- Study scope: 760 AI responses across 190 real-world queries in 9 Australian and 10 US verticals, tested across ChatGPT, Gemini, Claude and Perplexity.
- Key finding 1 (referring domains): unique referring domains show a real, statistically reliable positive relationship with AI citations in both Australia (r = 0.27, p = 0.001) and the US (r = 0.43, p < 0.0001).
- Key finding 2 (Google rankings): ranking on page one of Google has no statistically meaningful relationship with AI citations in Australia (r = 0.09, p = 0.31) and only a marginal one in the US (r = 0.15, p = 0.076).
- Key finding 3 (quartile multiplier): domains in the top 25% for referring domains earned 1.9x more AI citations in Australia and 2.3x more in the US than the bottom quartile.
We ran 90 real-world queries across 9 verticals in Australia (home services, healthcare, real estate, ecommerce, personal finance, legal, education, travel, fitness) and 100 queries across 10 verticals in the US, each fired at four AI engines with web search switched on. That's 360 AI responses in Australia and 400 in the US, producing 3,069 and roughly 2,500 to 3,200 citation events respectively. We then matched every cited domain against Ahrefs (Domain Rating, referring domains, total backlinks) and Semrush (share of keywords ranking on page one of Google).
Do referring domains predict AI citations better than Google rankings?
Yes, consistently, in both markets.
Across both countries, one thing stood out above everything else: domains with links from a wider range of other websites got cited by AI noticeably more often. Not domain authority scores. Not Google rankings. The sheer number of different sites linking in.
In Australia, that pattern held across 137 domains we measured, and it's a real, reliable relationship, not one factor among many, with less than a 0.1% chance it's a coincidence. In the US, the same pattern showed up even more strongly across 141 domains, with less than a 0.01% chance.
Compare that to Domain Rating (a single aggregate authority score), which moved the needle less in both markets, and to Google page one rankings, which barely moved it at all.
Do Google page one rankings get you cited by AI in 2026?
Barely. In Australia, not even that.
Whether a domain had a big share of keywords ranking on page one of Google had almost no bearing on how often AI cited it in Australia. Statistically, that's indistinguishable from random chance. In the US, there was a faint hint of a link, but it fell short of the bar researchers normally use to call a result reliable.
The takeaway: a domain can dominate page one of Google for its core keywords while still being invisible to ChatGPT and Gemini. What moved the needle instead was the width of its backlink profile.
How much do referring domains increase AI citations?
We split domains into quartiles (best 25% vs. worst 25%) to see how big the practical gap actually is:
- Domains in the top 25% for referring domains earned 1.9x more AI citations than the bottom 25% in Australia, and 2.3x more in the US.
- Domains in the top 25% for Google page one ranking share earned only 1.5x more (AU) and 1.6x more (US) citations than the bottom 25%.
Link building is having a much greater effect than rankings, in both countries, on entirely different query sets.
It's not just the big brands winning
Even with heavyweights like Reddit, Forbes, Canstar and Finder in the mix, the top 10% most-cited domains captured only around 28% of total citations in each market. Most of the citation activity is spread across a long tail of smaller sites. There's real room for challenger brands to break in, if they build the right link profile.
Who's actually winning right now
Australia
| Domain | Citations | Domain Rating | Referring Domains |
|---|---|---|---|
| canstar.com.au | 23 | 77 | 10,352 |
| finder.com.au | 23 | 78 | 14,202 |
| choice.com.au | 18 | 80 | 16,305 |
| moneysmart.gov.au | 18 | 85 | 19,297 |
| healthdirect.gov.au | 16 | 88 | 41,474 |
United States
| Domain | Citations | Domain Rating | Referring Domains |
|---|---|---|---|
| reddit.com | 35 | 95 | 1,410,080 |
| forbes.com | 35 | 94 | 818,860 |
| nerdwallet.com | 21 | 90 | 102,738 |
| healthline.com | 15 | 92 | 320,927 |
| medicalnewstoday.com | 11 | 91 | 185,231 |
Every one of these sits at the intersection of strong rankings and an exceptionally wide backlink profile, but it's the second variable that separates the frequently cited from the merely well-ranked once you look below the top names.
What this means
We'd stop short of saying links cause AI citations. This is correlational data, and AI engines are also weighing content structure, recency and topical fit. But the pattern is consistent, reliable, and it turned up independently in two countries, two query sets and two Ahrefs/Semrush pulls. That's a rare thing in SEO data right now.
The practical takeaway for anyone chasing visibility in AI Overviews, ChatGPT and Perplexity: don't just optimise for rank one. Build a backlink profile that's wide as well as strong. A domain with a modest Domain Rating but links from a genuinely broad set of referring domains is, on this data, a better AI citation bet than a domain that only ranks well.
The numbers behind this (for the data-literate)
| Relationship | Australia | United States |
|---|---|---|
| Citations vs. referring domains | r = 0.27, p = 0.001, n = 137 | r = 0.43, p < 0.0001, n = 141 |
| Citations vs. Domain Rating | r = 0.24, p = 0.005, n = 137 | r = 0.28, p = 0.001, n = 141 |
| Citations vs. % keywords on page one | r = 0.09, p = 0.31 (not significant) | r = 0.15, p = 0.076 (marginal) |
With the table above: r measures how tightly two things move together, from 0 (no relationship) to 1 (perfect lockstep). p is the probability the result is random noise; under 0.05 is the usual bar for "real". n is the number of domains the figure is based on. According to the Earned Media GEO study's regression model, each 10x jump in referring domains predicted roughly 1.4 more AI citations per domain in Australia and 2.2 more in the US.
Methodology
AU domain-level analysis covers 137 .com.au/.gov.au/.edu.au/.net.au/.org.au domains cited 3 or more times. US domain-level analysis covers 141 domains cited 3 or more times from the original 400-call, 4-engine study. Page-level URL analysis in the US reflects a separate re-collection run across ChatGPT, Claude and Perplexity only, after Gemini's API suffered a sustained outage during that phase of data collection.
Want a read on where your own domain stands? Request a free Truth Gap audit. We run your buyers' deal-breaker questions across the major AI models and show you exactly what's missing, and what's holding your citation rate back.