TJ Digital: 23% of AI-cited pages open with a direct answer, versus 14% of uncited ones

The agency TJ Digital analyzed 193 507 ChatGPT and Google answers. 23% of cited pages opened with a direct answer, versus 14% of pages that were found but not cited. The gap is small, and analyst TJ Robertson calls it a correlation, not a cause. The agency that collected the data sells SEO and GEO services, so rewriting a whole site is premature. First, check the openings of your important pages against their target questions. • ChatGPT cites about 27% of the pages it finds; AI Overviews and AI Mode cite 85–92% • Indig: 44.2% of ChatGPT citations fall in the first 30% of the text, only 1.47 times above the baseline • A date within the last 12 months gives +88% odds of a citation, tables +69%

AI-processed from Habr AI; edited by Hamidun News
TJ Digital: 23% of AI-cited pages open with a direct answer, versus 14% of uncited ones
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1ChatGPT cites about 27% of the pages it finds; AI Overviews and AI Mode cite 85–92%
Indig44.2% of ChatGPT citations fall in the first 30% of the text, only 1.47 times above the baseline
Key metric #3A date within the last 12 months gives +88% odds of a citation, tables +69%

The advice to "give the answer in the first paragraph" has become a mantra of GEO optimization, the fight for a place in AI answers. TJ Robertson of the agency TJ Digital analyzed 193 507 answers from ChatGPT and Google's AI answers (AI Mode and AI Overviews). 23% of the pages the AI found and cited opened with a direct answer, as did 14% of the pages it found but did not cite. The gap is small, and Robertson himself calls it a correlation, not a cause. The practical takeaway for site owners: don't rewrite everything. Check the openings of your important pages, each against its target question.

What the measurements showed and where the gaps are

AI models graded the pages twice, blind. Compared with pages from Google's top 10 that the AI never opened at all, the gap is larger: 20% versus 5% across 195 page pairs. Even in the best group, though, three pages out of four open without a direct answer, so most cited pages do fine without one.

The numbers need caveats. The agency that collected the data sells SEO and GEO services itself. For the 23% vs 14% comparison, neither the number of pages nor the grader model had been disclosed as of October 28. Almost a quarter of the cited pages did not answer the buyer's question in the part the graders looked at, so a link does not guarantee an answer.

Kevin Indig's measurement sounds louder: of 18 012 ChatGPT citations, 44.2% fell in the first 30% of the content. But if citations were spread evenly, those 30% would get 30% anyway. A recalculation by info.link gives a lift of only 1.47 times, and the peak sits at 10–20% of the page. The rule "answer in the first 100–200 words" is called exaggerated there. Within a paragraph the picture is the opposite: 53% of citations come from the middle and only 24.5% from the first sentences, so cramming everything into the first sentence is pointless.

Where a page gets lost before the first paragraph

How the service picks a fragment is unknown. OpenAI's developer documentation limits search context to 128 thousand tokens, and a page is not necessarily read in full, but that description cannot be carried over to the ChatGPT chat itself. What is known is that ChatGPT's main loss happens after search. According to TJ Digital, it cites about 27% of the pages it finds, while AI Overviews and AI Mode cite 85–92%.

Domain and position matter too. Pages from Google's top 3 are cited in AI Overviews 13% of the time, and pages from positions 4–10 8% of the time. Meanwhile, 10% of domains received 81% of all citations.

In a preprint by Mahe Chen and Nick Koudas covering 158 099 answers, every cited page had been found by ChatGPT's own search through sub-queries, which the model rewrites from scratch for each answer. Getting into this internal search is more strongly associated with citation than any of the page properties studied. Search Engine Journal adds features adjusted for domain and page type: a publication or update date within the last 12 months (+88% odds) and tables (+69%), as well as personal experience and prices.

How to check a page's opening in three steps

1. Pick a page and one question people search for to find it, phrased the way a buyer would phrase it. The test: it fits in one sentence.

2. Copy the first paragraph into an empty file and show it to a person who has not seen the page, asking them your question. If they answer correctly, the paragraph passes.

A company story or reflections on the importance of the topic do not count as an answer. An answer is a number, a decision, a definition or steps. 3.

Move the answer to the top and leave the rest below, making sure no fact is lost. For example, instead of "We have been working since 2012…" write: "Laptop repair takes one to three days, diagnostics are free, and a screen replacement costs from 4 500 ₽."

After the edit, check indexing. Google says only indexed pages make it into AI Overviews and AI Mode links, and the features may split a question into several searches. Check the URL in Google Search Console and look for the page using the target question and two or three rephrasings. According to Yandex, Alice AI relies on highly ranked sources and builds each answer anew, so repeat the check on different days and look at the page in Yandex Webmaster.

The findings cannot be carried over to all services. Answer position was measured on ChatGPT and Google, and in a small sample by the agency AiBoost also on Perplexity and Gemini. There are no such measurements for Alice AI and GigaChat. In AiBoost's sample of 100 British pages, 47.3% of cited fragments lay in the first 30% of the text, but the effect is weakest for service and product pages and strongest for articles. The sample is small, and the source is an agency blog.

Bottom line: an answer at the start of the page is necessary hygiene, but not a guarantee. Without demand, indexing, rankings and brand recognition, it will not pull a page into citations.

Story

Key Questions Answered

What the measurements showed and where the gaps are?

AI models graded the pages twice, blind. Compared with pages from Google's top 10 that the AI never opened at all, the gap is larger: 20% versus 5% across 195 page pairs. Even in the best group, though, three pages out of four open without a direct answer, so most cited pages do fine without one. The numbers need caveats. The agency that collected the data sells SEO and GEO services itself. For the 23% vs 14% comparison, neither the number of pages nor the grader model had been disclosed as of October 28. Almost a quarter of the cited pages did not answer the buyer's questi

Where a page gets lost before the first paragraph?

How the service picks a fragment is unknown. OpenAI's developer documentation limits search context to 128 thousand tokens, and a page is not necessarily read in full, but that description cannot be carried over to the ChatGPT chat itself. What is known is that ChatGPT's main loss happens after search. According to TJ Digital, it cites about 27% of the pages it finds, while AI Overviews and AI Mode cite 85–92%. Domain and position matter too. Pages from Google's top 3 are cited in AI Overviews 13% of the time, and pages from positions 4–10 8% of the time. Meanwhile, 10% of

How to check a page's opening in three steps?

1. Pick a page and one question people search for to find it, phrased the way a buyer would phrase it. The test: it fits in one sentence. 2. Copy the first paragraph into an empty file and show it to a person who has not seen the page, asking them your question. If they answer correctly, the paragraph passes. A company story or reflections on the importance of the topic do not count as an answer. An answer is a number, a decision, a definition or steps. 3. Move the answer to the top and leave the rest below, making sure no fact is lost. For example, instead of "We have been

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…