44% of ChatGPT citations come from the first 30% of a page

What 50,000 ChatGPT citations actually show about page structure

Otterly.ai published the largest dataset on ChatGPT citation behavior to date this month. Their team tracked 50,000 ChatGPT references across English-language business and technology queries between January and March 2026, then reverse-engineered which parts of the cited pages the model actually quoted. The headline finding: 44.2% of citations come from the first 30% of source content. Just under half of every ChatGPT citation is being pulled from the opening third of the page, regardless of how long the page is.

This number explains a pattern Vyzz has been seeing in audits all quarter. Small businesses with healthy domain authority and reasonable on-page content are still being ignored when an AI assistant answers a query in their category. The pages have the right facts. The facts are buried.

If you run a service business and your strongest claim sits in paragraph six, the model walks past it. The fix usually comes down to moving what you already wrote into the first 100 words.

What "front-loading" means in practice for the three SMB page types

Service pages, FAQ pages, and case studies fail this test in different ways. The reorder move is different for each.

Service pages

Most SMB service pages open with a warm framing paragraph. "We've been serving the community for fifteen years." "Our team is committed to a personalized experience." Then the actual differentiators (specific procedures, accepted insurance, price ranges, service area) arrive in paragraph four. A model scoring the first 300 characters for relevance to "best dentist accepting Delta Dental in [area]" finds nothing extractable in the opening and either cites a competitor or skips the citation.

The reorder: the first sentence of a service page should name the service, the entity, and the specific. "[Practice] offers same-day crowns in the $950 to $1,400 range and accepts Delta Dental PPO and Aetna." Then the warmth, the team, the years of service. The reader who wants the warmth will scroll. The model that needs the answer has it.

FAQ pages

The FAQ format works against itself when the question is followed by setup. A medspa page that asks "Does it hurt?" and then opens the answer with three sentences of context before getting to "Most patients describe it as a brief stinging sensation, mild and short" loses the citation to a competitor whose answer arrives in sentence one.

The reorder: lead with the answer. One sentence. Then add the nuance.

Case studies

Case studies almost always bury the result in the back half. The setup, the methodology, the journey, then the number. A model looking for "how much can a regional services brand expect from this kind of work" needs the result up front to cite it.

The reorder: open with the result and the period. "A regional services brand we audited grew booked-call volume 28% across two quarters by republishing service-page openings with prices in plain text." Then the rest. Same words, different order.

How to audit your own pages for citation gravity

The audit you can run yourself in a single afternoon: open the top 10 pages on your site by traffic. For each one, copy the first 300 characters into a notepad. Read those 300 characters as if you are a buyer who has never heard of the business.

Ask three questions about those 300 characters. Is the specific service, product, or topic of the page named in the opening? Is there a number, a price, a date, a procedure name, or any other extractable fact present? Will a buyer get the answer they came for before they have to care about the brand voice?

If the answer is no for any of those three, that page is bleeding citations. The rewrite is reordering, not adding.

A second test that costs nothing: type your business name into ChatGPT and read the response. Then type the most common service query for your category and area and read whether your business appears. Most SMBs do this once and never again. The pages cited in those answers are the pages where the opening 300 characters did the work. Look at them. They are doing nothing magical.

Why most SMB pages still bury the lead

The reason this pattern is so widespread comes down to a writing convention that worked for a decade. SEO copy in 2014 to 2022 rewarded long-form pages with deep keyword coverage, a setup paragraph that warmed the reader, and an answer delivered after enough on-page time to satisfy dwell-time signals. Google rewarded the shape, copywriters were trained in it, and most SMB sites today reflect that lineage. The page is structured the way the reader of 2018 was assumed to read.

The retrieval model in 2026 does not read that way. It reads the opening, scores the passage, and decides whether to cite. A page that worked perfectly for human readers and Google's old quality signals can score zero with the model. The site has not changed. The reader has.

This is also why so many SMBs assume the answer is more content. The instinct is to add another blog post, another FAQ entry, another case study. The pages already cover the topic. The problem sits in the order: the strongest sentence is on the page, just not where the model looks first. Adding more pages with the same shape produces more pages the model walks past.

What the rest of the data says

Beyond the front-loading finding, Otterly.ai's report (Otterly.ai 2026 Citations Report) measured two other signals worth knowing. Domain authority still amplifies: domains with 32,000+ referring domains earn roughly 3.5x more citations than domains in the 5,000 range. Brand search volume shows the strongest single-signal correlation with citation likelihood (r=0.334).

The structural read on those numbers: authority and demand still matter, and they always will. But the citation-location effect operates inside whatever authority you already have. A mid-authority SMB that front-loads its top 10 pages will outperform a higher-authority site that does not. Front-loading is the one move available to a small business that does not require waiting on backlinks or brand awareness to compound.

For a broader view of how AI Overviews are reshaping search behavior, Search Engine Land's March 2026 update put AIO coverage at 48% of all Google searches, up from 15% a year earlier.

What to do this week

Pick the three pages on your site that get the most traffic. Open each one. Read the first 300 characters out loud. If those 300 characters do not contain the strongest claim, the most specific number, or the named service that the page is about, reorder them today. Move the sentences you already have into the opening, and let the warmth scroll lower.

The reader who scrolls patiently will still find the warmth in paragraph four. The model that pulls the citation will find the answer in sentence one. Same content, different order, measurably more visibility.