Most of the pages that ChatGPT and Google’s AI cite do not open by answering the question they were cited for. Those that do appear in AI answers noticeably more often than the pages passed over beside them, according to a study of 193,507 answers published on 25 October by TJ Digital, a Nevada-based agency that sells SEO and generative engine optimisation (GEO) services.

The finding lands as B2B content teams face growing pressure to show they appear in AI-generated answers as well as in Google’s rankings. TJ Digital’s dataset is among the largest public attempts so far to describe what cited pages have in common. It is also observational, built on questions the agency tracks for its own clients, and a University of Toronto paper released two weeks earlier found that which pages an engine retrieves explains far more of the citation picture than anything written on them. For marketing teams, that makes answer-first structure a cheap test worth running rather than a lever that guarantees citations.

What did TJ Digital’s study of 193,507 AI answers find?

TJ Digital, run by founder TJ Robertson, collected answers from ChatGPT, Google AI Mode and Google AI Overviews in eight weekly snapshots between 27 August and 21 October 2026. The questions were 4,167 unbranded buyer queries tracked in the Peec AI monitoring tool for 78 client businesses across nine business models. The answers carried 943,673 citations to 209,859 URLs.

To judge page quality, AI raters scored pairs of pages blind, without knowing which had been cited. In one round, the cited page opened with a direct answer 23% of the time, against 14% for a page the same AI answer had retrieved but not cited. In a second round of 195 pairs, set against Google top-10 pages the AI never retrieved, the figures were 20% and 5%. Half of those passed-over pages failed to answer the question anywhere in the section the raters read, but so did 23% of the cited pages.

Other traits tracked citation too. After adjusting for Ahrefs Domain Rating and page type, across 21,874 cited and 4,507 passed-over pages, a page published or updated in the past 12 months had 88% higher odds of being cited. Tables carried 69% higher odds, first-hand experience 45% and visible prices 28%. A year in the page title more than doubled the odds. Two results ran the other way: pages with an author byline had 28% lower odds, and pages with a call to action 12% lower. Robertson wrote that he could not explain either result and would not remove bylines or calls to action on this evidence alone.

For B2B teams, the page-type figures carry weight. Product pages took 28% of all citations, homepages 12% and category pages 10%, while comparison pages took 2.5% and “alternatives to” pages 0.2%. Company websites, including competitors’ sites, accounted for 67% of citations overall, rising to 87% for B2B industrial and distribution businesses.

How strong is the evidence that answer-first pages earn more citations?

TJ Digital’s report states plainly that its results are associations and cannot show that any trait causes a citation. Several gaps limit how far the answer-first figures travel.

The report does not define what counted as a “direct answer” or how much of each page the raters read. Search Engine Journal, covering the study on 28 October, noted that the report had not disclosed which AI model did the rating or how many pages sat behind the 23% versus 14% round. The Google comparison also used Ahrefs keyword data as a stand-in for the questions asked: of 371 keywords, 197 were broader than the question they represented, and all came from one crawl location on one date.

Sampling shapes the result too. The questions are those TJ Digital tracks for paying clients, so the findings describe that client base rather than AI search at large, and the agency sells the GEO services its recommendations point towards. The data may well hold, but the nine-point gap between cited and retrieved pages is a pattern to test, not a measured effect of rewriting an introduction.

The study’s own volatility figures add a further caution. When the same engine answered the same question a week later, only 24% to 28% of previously cited pages were cited again, falling to 11% to 16% after seven weeks. A single before-and-after check on a rewritten page would struggle to separate a real gain from that churn.

What does independent research say decides AI citations?

Mahe Chen and Nick Koudas of the University of Toronto put numbers on the retrieval question in a preprint submitted on 10 October. Analysing 158,099 answers from ChatGPT, Gemini and Perplexity collected over 90 days, they found that page features explained about 2.8% of the variation in ChatGPT’s citation counts within a single website. Retrieval, meaning whether ChatGPT’s own search returned the page at all, explained about 32%.

The same paper found that common GEO tactics, including FAQ blocks, schema markup and summary boxes, appeared to help when comparing different websites but showed no benefit when comparing pages within one site. The authors attribute the cross-site effect to those features clustering on sites that are already popular. One content signal did hold within sites: titles and paragraphs that closely matched the question asked. That effect was stronger still in a second dataset drawn from a niche B2B software market, proposal software for architecture and engineering firms.

That overlaps with part of TJ Digital’s result, which also found that pages whose openings matched the question did better. It is consistent with earlier work The Helm has covered, including an AirOps analysis that ranked retrieval position as the strongest predictor of ChatGPT citation and a critical survey of 45 GEO studies that found no technique reliably lifted visibility across engines.

TJ Digital’s own recommendations follow the narrower reading: answer the question in a page’s first lines, keep important pages current, show prices, and track a fixed set of questions across several engines for weeks rather than judging from one answer. The report compares pages as they already existed and includes no controlled test of whether adding a direct answer to an existing page changes how often AI engines cite it.