Study of 193,507 AI Answers Finds Cited Pages Open With Direct Answers

Pages cited by ChatGPT, Google AI Mode and AI Overviews opened with a direct answer to the question in 23% of cases, versus 14% for pages the same answers retrieved but did not cite. The figure comes from an analysis of 193,507 AI answers conducted by TJ Robertson's agency TJ Digital. Robertson sells both SEO and Generative Engine Optimization services, and he frames the findings as associations, not causes.

What did the study measure?

The dataset covers 4,167 questions that named no specific company or product, drawn from 78 client companies across nine business models. Responses were captured in eight weekly snapshots between August 27 and October 21, 2026, plus repeat runs. The agency used Peec AI, a tool that feeds preset prompts to AI systems and records the answers and their sources.

To compare page traits, the agency extracted 21,874 cited pages and matched them against 4,507 comparative pages across 434 questions. It called these comparative pages "passed-over pages": pages that ranked in Google's top ten for a keyword, per Ahrefs data, but that the AI never surfaced.

Of the 371 keywords involved, 174 matched the question closely; 197 were broader. That asymmetry matters. Keyword-to-question fit was uneven, and the study has not quantified how much that skewed results.

How strong was the direct-answer effect?

Robertson ran two rounds of blind page ratings. AI raters scored 195 page pairs without knowing which page was cited.

In one round, 20% of cited pages opened with a direct answer, against 5% for passed-over pages. Half of passed-over pages contained no answer anywhere in the portion raters read; 23% of cited pages also missed the answer. Cited pages scored 2.3 on a 1-to-5 completeness scale; passed-over pages scored 1.7.

The gap narrowed when uncited pages were still retrieved by the AI. That is the 23% versus 14% split from the earlier round. As of October 28, Robertson's page had not stated how many pages that round analyzed, nor which AI model evaluated them.

Promotional tone showed little difference. Roughly three quarters of pages in both categories promoted the researcher's own products or services, and tone ratings were 3.8 versus 3.7.

Takeaway: citation correlates with answering early, but citations do not prove a page answered the question.

Which page traits moved citation odds?

The analysis controlled for Ahrefs Domain Rating and page type, then reported trait odds:

• A year in the title: more than twice the odds of citation
• Publish or update date within 12 months: 88% higher odds
• Tables: 69% higher odds
• First-hand experience: 45% higher odds
• Prices: 28% higher odds
• Schema markup: 13% higher odds
• Author byline: 28% lower odds
• Call to action: 12% lower odds

Robertson said he cannot explain the byline and CTA penalties and would not remove either element on this evidence alone.

The schema finding echoes and complicates earlier work. An Ahrefs report from May found AI-cited pages were roughly three times as likely to contain JSON-LD, yet Ahrefs' controlled test showed adding JSON-LD to previously cited pages produced no statistically significant citation gain. Robertson also claims the trait patterns replicated in a second test limited to Google Top-10 pages, but the per-engine table and graph referenced on his page were not visible as of October 28.

Takeaway: correlational odds from one agency's client questions are testable patterns, not playbook rules.

Can you act on this?

The study compared pages as they existed. It cannot show that adding an opening answer will change citation counts. A before-and-after experiment with a test group and a control group, like the one Ahrefs ran on schema, would settle the question.

Until such a test runs, the opening-answer pattern remains a hypothesis to check against the questions you already track.