Does E-E-A-T still hurt AI citations when you control for content type?

A plausible mechanical explanation: visible expertise signaling on a page — author bylines with credentials, medical-reviewer attribution, "expert reviewed by" labels, professional-bio sidebars — appears most prominently on long-form, expert-reviewed content. Those same pages tend to be dense, with multiple sections, FAQs, and visual breakouts.

AI assistants answering "what is high blood pressure" often generate a confident synthesis from training without surfacing any citation at all. The well-credentialed long-form page never gets attributed. A sparser page with a clean definition near the top fits more naturally into the AI's answer structure — and gets pulled in instead.

That fits a hypothesis we've heard articulated elsewhere: AI can verify parseable answers but can't verify credentials, so it leans on what it can directly check.

The honest takeaway: the E-E-A-T pillar isn't a reliable citation predictor. Three independent studies (the original homepage research, the first run of this study, and the redesigned run) all point the same direction. Whatever those signals do for Google's framework, they aren't what AI citation behaviour rewards.

We selected 40 URLs in a balanced 2×2 design crossing two factors. Industry: healthcare (YMYL territory where Google's framework says credentials matter most) vs SaaS (technical/commercial, where author authority should matter less). Content type: educational pages whose purpose is to answer a topic question vs authority pages whose purpose is to convey institutional credibility. Ten pages per cell.

Each URL was scanned with the same GeoSource scoring engine that produced the original study's pillar scores. The only things varying across the three studies are the URL set and the query phrasing.

Queries (redesigned). Every query is concept-based and never names a specific brand. Educational pages were paired with topical concept queries ("what causes migraines", "what is CRM software"). Authority pages were paired with concept queries about credibility or category ("how should you choose a hospital for serious illness", "what are the largest payment processing companies"). Whether an authority page surfaces under those queries is the test. The first-run design used brand-named authority queries, which inflated the authority-cell citation rate by roughly 47 percentage points (see "Why this study has two versions" above).

Citations were checked across ChatGPT, Perplexity, and Claude — 120 checks total. A URL counts as cited if either the domain appears in the response text or the platform lists a URL from that domain as a source. We re-ran the analysis restricting to formal URL citations only (excluding incidental prose mentions) — the numbers were identical, so the citations are real, not just brand names floating through answers.

Limitations. Sample size. Ten URLs per cell is small — read the direction across cells rather than any single magnitude. Bimodality. Most URLs are either fully cited or not cited at all, which compresses what any aggregate measure can detect. Authority pages tested indirectly. Asking "how should you choose a hospital" measures whether Mayo Clinic's about page happens to be the natural answer — that depends as much on the query as the page. A future study could vary E-E-A-T signals directly on the same URL (A/B with and without author bylines) to isolate the page-content effect. What "E-E-A-T" means here. The score we're calling E-E-A-T is the GeoSource pillar's measure of visible credential signaling on the page — it's not Google's full E-E-A-T framework, which includes off-page signals our scanner can't see.