Pages cited by AI Overviews, ChatGPT, and Google’s AI Mode are more likely to put the answer in their first lines than pages those systems skipped. TJ Robertson reached that finding by analyzing 193,507 AI answers, and Search Engine Journal covered the results on October 28. The gap shows up in the data, but the method carries limits that content teams should weigh before rewriting anything.

Two comparisons produced two headline figures. Among pages the same AI answers retrieved, 23% of cited pages started with a direct answer, against 14% of retrieved pages that went uncited. Against top-ranking Google pages that the AI never retrieved, the split was 20% for cited pages and 5% for the others. The smaller first gap suggests that once a page is retrieved, its opening matters less than it does for getting retrieved at all.

How the data was built

Robertson’s company, TJ Digital, sells both SEO and generative engine optimization services, and the responses came from questions it tracks for clients. They covered 4,167 questions that named no company or product, across 78 client companies and nine business models. Eight weekly snapshots between August 27 and October 21, 2026, were supplemented by repeat runs. Peec AI, a tool that submits fixed prompts to AI engines and records answers and sources, handled collection.

For page-level work, the agency extracted 21,874 cited pages and 4,507 comparison pages tied to 434 questions. Each comparison page ranked in Google’s top ten for a question but never surfaced in the AI’s searches. Rankings came from Ahrefs keyword data rather than the exact questions asked. Of the 371 keywords involved, 174 were classed as close to the question and 197 as broader.

What the blind ratings showed

AI raters scored 195 pairs of pages, blind to which page of each pair earned the citation. The 20% versus 5% figure comes from this round. Among the comparison pages, half held no answer anywhere in the portion raters read, against 23% of cited pages. On a 1-to-5 completeness scale, cited pages averaged 2.3 and comparison pages 1.7.

Promotional tone did not separate the groups. About three fourths of pages in each were businesses promoting their own offerings, and tone scores came out at 3.8 and 3.7.

Where the evidence is thin

The 23% versus 14% figure came from an earlier rating round. As of October 28, Robertson’s page did not state how many pages that round covered or which AI model did the rating. The raters were AI, not people, so scores inherit whatever that model favors. Robertson himself calls the findings associations, not proof that a trait causes a citation.

The analysis also reports odds of citation after controlling for site authority via Ahrefs Domain Rating, plus page type. A year in the title carried more than double the odds. A date of publication or update in the last year raised the odds by 88%, tables by 69%, first-hand experience 45%, and prices 28%. Author bylines were tied to 28% lower odds and calls to action to 12% lower. Robertson offers no explanation for those two and would not strip either element on that basis.

Schema markup showed 13% higher odds. Ahrefs tested that in May by adding JSON-LD to pages already cited in AI Overviews and found no statistically significant citation gain. A table and chart broken out by engine were not visible on Robertson’s page on October 28, so engine-by-engine differences cannot be checked.

What content teams can do with it

The study compares two groups of pages as they stood. It cannot show what happens when a page changes. About a quarter of cited pages (23%) still lacked an answer in the part raters saw, so a citation does not certify a good answer.

Only a controlled test settles whether answer-first rewrites earn citations: rewrite one group of pages, leave a matched group alone, and measure both before and after. Until then, treat the pattern as a hypothesis. Teams that track buyer questions should check whether key pages answer in the first paragraph and log citation rates against that, since a lead rewrite is cheap and reversible.

Based on reporting by Matt G. Southern at Search Engine Journal, published October 28, 2026, on TJ Robertson’s analysis of AI citations.