What actually predicts AI citations? Six findings.
We ran four studies of how AI platforms cite and recommend content. A handful of patterns showed up in every one of them. Those are the patterns worth acting on. This page is the consolidated read, with links to each underlying study.
Research at a glance
The six findings
Each of these showed up across at least two of our independent studies. We're presenting them qualitatively here; specific numbers live on the individual study pages.
Brand recognition swamps page quality
Across every study, well-known brands got cited even with weak on-page signals; lesser-known brands struggled to get cited even with strong ones. In our ecommerce work, brands with very low GEO scores routinely outperformed brands with very high ones. Page quality moves the needle, but brand familiarity is the bigger lever — and one that page-level optimization can't fix.
Query phrasing matters more than page content
When we changed the wording of a query — switching "is Mayo Clinic trustworthy" to "what makes a hospital trustworthy" — the same page's citation rate collapsed without any change to the page itself. Whether your content gets cited is mostly decided by how the query gets asked, not by anything you optimize on the page.
E-E-A-T signaling doesn't predict AI citation
Three independent studies — the original homepage study, the E-E-A-T 2×2 follow-up, and the ecommerce survival study — all showed the same thing. Visible E-E-A-T signals (author bylines, "expert reviewed by" labels, credential sidebars) don't make AI more likely to cite you. In the ecommerce study they actively hurt. Whatever those signals do for Google's framework, they aren't what AI assistants are reading.
Citation rate is the wrong top-line metric
Being mentioned isn't the same as being recommended. Brands with identical citation rates can be treated completely differently — one as the AI's top pick, another buried in a list of competitors. The yes/no signal hides that. Our follow-up strength scoring separates active recommendations from neutral mentions.
Content type sets the citation ceiling
Informational pages (definitions, how-tos, condition guides) get cited at dramatically higher rates than product or landing pages. Some queries don't produce sourced citations at all — AI just answers from training. Users can't change their product category, but they can choose what kinds of content to publish, and that's one of the highest-leverage levers we found.
Platforms agree more than they disagree
When we asked the same query on ChatGPT, Perplexity, and Claude, they almost always agreed on which sources to cite. A page that gets cited by one platform usually gets cited by all; a page that's ignored by one is usually ignored by all. Per-platform optimization rarely pays off — per-query optimization does.
Which studies support which findings
A finding is only listed above if at least two independent studies pointed at it. This matrix shows where each finding showed up — a green check means the study tested it directly, a blue check means the study showed the same direction as a side observation, and a dash means the study didn't test that question.
| Finding | Citation study | E-E-A-T 2×2 | Ecommerce | Multi-turn |
|---|---|---|---|---|
| Brand recognition swamps page quality | ||||
| Query phrasing matters more than page content | ||||
| E-E-A-T / Authority never predicts citation | ||||
| Citation rate is the wrong top-line metric | ||||
| Content type sets the citation ceiling | ||||
| Citations are mostly all-or-nothing |
The studies behind these findings
Each finding above is sourced from at least one of these. Each study page includes its full methodology, limitations, and qualitative outcomes.
The original GEO citation study
61 sites across 17 industries, testing which content signals correlate with AI citations and building the first version of the GEO scoring model. Where we first noticed the brand-recognition confound and the surprising negative E-E-A-T result.
Read the full study →E-E-A-T & content type follow-up
A controlled 2×2 follow-up to test whether the negative E-E-A-T result was a content-type confound. 40 URLs across healthcare vs SaaS, educational vs authority pages. We ran the study, found a query-design bug in the first attempt, redesigned, and reran — and the negative E-E-A-T direction survived the fix.
Read the full study →Ecommerce recommendation-survival study
40 brands × 12 product categories × 4 shopping-journey stages. Tested how brands survive a multi-stage shopping conversation, then ran a follow-up strength classifier on every AI response to score recommendation strength beyond binary citation. The results shipped as the Recommendation Readiness Score.
Read the full study →What this means if you're trying to get cited
The biggest leverage is upstream of your page. How your audience phrases their queries, what category your page competes in, and how well-known your brand already is — these matter more than the on-page signals our scoring engine measures. If you optimize the page in isolation, you're working on the small lever.
Pick the right queries to compete for. Brand-named queries ("who founded X", "is X trustworthy") almost always pull citations back to that brand. Concept queries ("what causes migraines", "what is CRM software") rarely do unless your page is genuinely the best answer. Both have a place in a content strategy, but they behave very differently.
Don't over-index on E-E-A-T signaling. Author bylines and credential cues didn't predict AI citation in any of our studies. They may help with Google SERPs, but they aren't what AI platforms are reading. Focus content effort on the pillars that did show up positively.
"Cited" isn't the goal — "actively recommended" is. Being named at position #8 in a list of competitors isn't the same outcome as being the top pick. Track the difference. Our Recommendation Readiness Score and follow-up strength analyzer are built around exactly this distinction.