AI Search Citation Statistics (2026): 31 Sourced Figures

This page reports what the pages cited by AI answer engines look like. It covers how many sources an answer cites, which page types earn citations, how long the cited pages are, and how recently they were updated. Every figure comes from one year of citation records in an AI search visibility tracking dataset. Each number links to the method section below, which states the data window and its limits.

The figures describe the cited pages as a group. They do not show that any single feature causes a citation. A common feature of cited pages can be just as common on pages that are never cited. The data has no record of uncited pages, so it cannot make that comparison.

StatWharf publishes a related study of the searches ChatGPT runs before it answers. It is at How ChatGPT searches for software recommendations. Results of StatWharf’s own test runs are at ChatGPT citation experiments. The way StatWharf builds and sources its own comparison pages is set out on the methodology page.

Method

The data comes from an AI search visibility tracking database. The tracker runs prompts through AI answer engines on a schedule. It stores each answer, each source the answer cites, and a crawl record for each cited page. StatWharf queried this database read-only on October 11, 2026.

The window covers every citation record created in the 365 days before October 11, 2026. The first record in the window is dated October 13, 2025. The last is dated October 9, 2026.

StatWharf used four kinds of record. The first is the citation record, one row each time an answer cites a source. The second is the answer record, which links each answer to the request that produced it and to the engine label of that request. The third is the page record, which holds the crawled attributes of each cited page. These include page type, word count, count of statistics, tables, structured data, an llms.txt file, writing formality, rendering type, publication date and last updated date. The fourth is the map between answers and the web searches the engine ran while it answered.

All figures are counts and shares computed directly from these records. No figure is modeled, sampled, simulated or estimated. The denominator of each share is the set of pages with that attribute recorded. The text states each denominator. No client, organization, brand, product, prompt, keyword or cited website is named or can be read from the figures.

Limits of the data

The tracker runs prompts that its users choose. The prompts lean toward commercial software and service questions. This is not a random sample of all ChatGPT use or of all AI search.

The engine mix is uneven. The data holds answers under six engine labels. ChatGPT appears under two collection labels, and together they hold most answers. An OpenAI label holds the next largest group. Google AI Mode and Google AI Overview hold a few dozen answers each. One label names no engine. The counts are in the engine table below. Because ChatGPT dominates the data, the page figures mostly describe ChatGPT citations.

Page attributes come from a crawl of each page. They describe the page as crawled. The engine possibly read a different version of the page. Some citation records have no stored page record, and some page records lack a given attribute. Publication and update dates are missing or invalid for many cited pages. The freshness figures use only pages with a valid date that falls on or before the first citation.

Scope of the data

An answer in this data usually cites many sources, not one. The median is in the table above. This matters for any page that wants to be cited. The goal is to be one of the sources in the answer, not the only one.

Answers by engine

The two Google engines hold too few answers to support separate figures. This page therefore reports no Google-only result.

Fan-out searches behind the answers

Before it answers, ChatGPT runs its own web searches. These are called fan-out queries. The tracker stores them where the engine exposes them.

The StatWharf fan-out study found the same median in a separate panel of software prompts. In that study, ChatGPT kept the category phrase in most searches and added vendor names and the current year.

How concentrated citations are

Most cited pages appear once in the year and never again. A small group of pages is cited again and again, and that group takes a large part of all citations. For a publisher, a single citation is common. Repeat citation is the rarer result.

Which page types are cited

The tracker classifies each cited page by type. The table gives each type’s share of citations whose page has a stored type. Only the five largest types are shown.

Vendor pages and blog posts together take more citations than any other group. Listicles hold a large share as well. Those are the pages that compare several options for one need, which is the format StatWharf publishes.

What cited pages contain

These shares use the cited pages that have a full crawl record. That set holds 904,196 pages. Formality uses a slightly smaller set of 900,369 pages.

Most cited pages carry no structured data. Most sites behind them publish no llms.txt file. Neither feature is required for a citation in this data. Nearly every cited page serves its text as static HTML. A page that needs a browser to run scripts before its text appears is rare among cited pages. Formal writing is the norm.

Cited listicles

Listicles differ from the average cited page. The figures below cover the cited pages classed as listicles.

A cited listicle carries a table more often than the average cited page. It is also longer. Comparison pages that set vendors side by side in a table match the common shape of a cited listicle. The StatWharf methodology page describes how its comparison tables are built and sourced.

How fresh cited pages are

Freshness runs from the last updated date to the first day of citation in the window. Where no update date is stored, the publication date is used. These shares cover 634,925 pages with a valid date.

Recent pages make up the largest group. Old pages are still cited, though. An undated or stale page is not shut out, but a recently updated page is the more common case.

What the figures support

The data supports a few plain statements about pages cited by AI answer engines, chiefly ChatGPT, for the commercial prompts that the tracker’s users track.

• An answer cites many sources. A page competes to be one of them.
• Product pages, blog posts and listicles take the most citations.
• Cited pages are mostly static HTML and formal in tone.
• Structured data and llms.txt files are absent from most cited pages.
• Cited listicles often use a table and run to several thousand words.
• Many cited pages were updated in the months before they were cited.

The data does not show that adding a table, structured data or a recent date will cause a citation. It shows only what cited pages have in common. Uncited pages are not in the data, so no comparison with them is possible.

Counting rules

A figure on this page is one row in a table. Each row has one linked value. Shares are rounded to one decimal place. Medians are exact values from the records. Denominators named in the text are linked to this method in the same way as the table values. The count in the page title covers table rows only.

Questions about the method can go to editorial@statwharf.com.

Factual corrections are free: editorial@statwharf.com.