Pages Cited by AI Answer the Question Up Front 23% of the Time

By Amara Osei

Pages cited by AI assistants opened with a direct answer to the user's question 23% of the time, versus 14% for pages the same AI answers retrieved but did not cite. That gap widens against Google Top-10 pages the AI never surfaced: 20% versus 5%. The figures come from TJ Robertson's analysis of 193,507 AI answers drawn from ChatGPT, Google AI Mode and AI Overviews, published via Search Engine Journal.

Robertson runs TJ Digital, an agency that sells both Search Engine Optimization (SEO, the practice of ranking in search results) and Generative Engine Optimization (the practice of winning citations in AI answers). That commercial stake matters when weighing the findings — though Robertson himself flags the limits of his own data.

He states he views the findings as showing associations rather than any specific page traits causing a citation.

How was the study built?

The analysis covers 4,167 questions that named no specific company or product, across 78 client companies in nine business models. Responses were captured in eight weekly "snapshots" between August 27 and October 21, 2026, plus repeat runs of the same questions.

The agency collected the AI answers using Peec AI, an application that presents a pre-determined set of prompts to AI tools and records both the answers and their cited sources. To compare page characteristics, TJ Digital extracted the content of 21,874 cited pages and matched them against 4,507 comparative pages for 434 questions.

Those comparative pages — labeled "passed-over pages" — all ranked in Google's top ten results for a keyword but, in the agency's words, "never came up in the AI's searches." Rankings came from Ahrefs keyword data rather than the actual questions asked. Of 371 keywords, the agency judged 174 similar to the question posed and 197 broader or more general.

What did the blind ratings show?

Robertson had AI raters blind-score 195 page pairs, so raters did not know which page in each pair was cited. This round produced the 20%-to-5% split on opening answers.

The same round found half of passed-over pages contained no answer to the question anywhere in the portion raters read, versus 23% of cited pages. On completeness, rated on a 1-to-5 scale, cited pages scored 2.3 against 1.7 for passed-over pages.

Promotional tone barely differentiated the groups. Roughly three quarters of pages in both categories promoted the agency client's own products or services, and tone ratings landed at 3.8 for cited and 7.7 — corrected: 3.7 — for passed-over pages.

The 23%-versus-14% split came from an earlier round comparing cited pages with pages the same AI answer retrieved but did not cite. As of October 28, Robertson's page does not state how many pages that round analyzed, nor which AI model scored them — a disclosure gap worth noting.

Which page traits moved citation odds?

Controlling for each site's Ahrefs Domain Rating (a 0-100 authority score) and page type, the analysis reports odds of being cited:

• A year in the title: more than twice the odds
• 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 (structured data labels for search engines): 13% higher odds

Two results ran the other way. Pages with an author byline carried 28% lower odds; pages with a call to action carried 12% lower odds. Robertson says he can't explain those two and wouldn't remove either element on this evidence alone.

The schema finding echoes but does not extend prior work. An Ahrefs report from May found AI-cited pages were roughly three times as likely to contain JSON-LD (a common schema format), yet Ahrefs' own testing showed adding JSON-LD to previously cited pages produced no statistically significant citation increase.

Robertson says the trait findings duplicated in a second test limited to Google Top-10 pages. As of October 28, however, the referenced per-engine table and graph were not visible on his page.

Why does this matter for marketers?

The practical test is simple: check whether your product and service pages answer the buyer's question early. The cited-versus-uncited gap shrank when uncited pages were also retrieved by the AI, so retrieval context matters.

The blind rater data also cuts against easy conclusions. Nearly a quarter of cited pages never answered the question in the portion raters reviewed — a citation does not prove a page answered anything.

All trait odds rest on questions one agency tracked for its own clients, so they describe how cited and passed-over pages differed, not universal laws.

What comes next?

The study compares pages as they existed. It cannot show that adding a direct answer will change citation counts. Proving that requires a controlled before-and-after test, as Ahrefs ran on schema — one group adds the element, a control group does not. Until someone runs it, the opening-answer pattern is a hypothesis to test against the questions you already track.