Preferences

Privacy is important to us, so you have the option of disabling certain types of storage that may not be necessary for the basic functioning of the website. Blocking categories may impact your experience on the website. More information

Accept all cookies

Wikipedia in AI Answers: a Data Study of Its Decline over 13 Months

Wikipedia went from the most cited source of the panel to third place in 13 months. Data study on 301025 AI answers, model by model.

Published on:
30/8/2026
-
Updated on:
30/8/2026
Line chart of the Wikipedia citation rate on ChatGPT, Gemini, Perplexity, Claude and Grok between August 2025 and August 2026, measured by Sorank
Line chart of the Wikipedia citation rate on ChatGPT, Gemini, Perplexity, Claude and Grok between August 2025 and August 2026, measured by Sorank
Thibault Besson-Magdelain fondateur de Sorank

About Author

Thibault Besson-Magdelain

Founder of Sorank, 5+ years of experience in SEO, GEO enthusiast.

The short answer

Wikipedia went from 5.28% of measured answers in August 2025 to 1.39% in August 2026, after a peak of 11.31% in October 2025. It is still the leading source on ChatGPT, which cites it in 3.43% of its answers, and almost absent from Claude at 0.03%. Over the same period the leading review platform reached 2.7%.

The single largest movement in 13 months of data is not a platform rising, it is the encyclopedia falling.

Key figures

  • 301025 AI answers analysed between August 2025 and August 2026
  • 1820738 citations extracted from those answers
  • 131349 distinct domains cited at least once
  • 5 models measured separately: ChatGPT, Gemini, Perplexity, Claude and Grok
  • 79% of the questions asked in French, the rest in English

Wikipedia model by model

  • ChatGPT: 3.43%
  • Gemini: 1.38%
  • Grok: 1.08%
  • Perplexity: 0.19%
  • Claude: 0.03%

The sectors where Wikipedia still weighs

  • Automotive: 4.39%
  • Fashion: 3.58%
  • Sports: 2.67%
  • Entertainment: 2.23%
  • Consulting: 2.1%
  • Retail: 2.04%

How to read these numbers

A citation rate of 2% means the source was cited twice for every 100 measured answers. Rates are computed per model against the number of answers of that model, so a model with a smaller sample never distorts a column. A source cited several times inside one answer is counted once per company block, which keeps a single verbose answer from dominating a month.

What changed over 13 months

Averaged over the first three months against the last three, Wikipedia moved from 9.56% to 1.62%, while the leading review platform moved from 0.36% to 2.7%. The two lines crossed in April 2026.

Month by month

  • August 2025: 5.28%
  • September 2025: 7.94%
  • October 2025: 11.31%
  • November 2025: 11.16%
  • December 2025: 7.9%
  • January 2026: 8.69%
  • February 2026: 10.53%
  • March 2026: 9.09%
  • April 2026: 0.17%
  • May 2026: 0.98%
  • June 2026: 2.25%
  • July 2026: 1.57%
  • August 2026: 1.39%

What this means in practice

Wikipedia started the period as the most cited source in the panel and ended it well behind review and video platforms. The decline is visible on every model except the most conservative one, which still leans on it heavily. Two readings are possible: the models diversified their retrieval, or the questions in the panel moved towards commercial intent. Both are probably true, and both point the same way for a business, which is that an encyclopedic entry is no longer the shortcut it was.

How to use this study

  1. Keep your entity data consistent, since encyclopedic content still feeds entity recognition.
  2. Stop treating an encyclopedia entry as a citation strategy on its own.
  3. Move the effort towards the sources that rose over the same period, starting with reviews.

How this study was measured

  1. We run a panel of buyer questions every month, the questions people type before choosing a supplier.
  2. Every question is sent to five AI: ChatGPT, Gemini, Perplexity, Claude and Grok.
  3. Every answer is parsed, every cited source is extracted and normalised to a domain and a platform.
  4. The citation rate is the number of citations of a source per 100 measured answers, computed per model so an uneven number of answers per model never distorts a column.
  5. The month is closed, the dataset is rebuilt, and the public files are republished on data.sorank.com.

Limits of this study

  1. This panel is not a neutral sample of the web. The questions come from the sectors our users operate in.
  2. The panel leans French: 237930 answers come from questions asked in French and 63095 from questions asked in English.
  3. Model coverage is uneven: 87008 answers from ChatGPT, 87004 from Perplexity, 87000 from Gemini, 20011 from Claude and 20002 from Grok. The two smallest samples are read as direction rather than as precise values.
  4. A citation is not a click. This study measures what the AIs cite, not what a reader does next.

The dataset behind this study

Every number on this page comes from the public Sorank dataset. The underlying figures, the monthly detail and the filters are available here:

Open the source data on data.sorank.com

Data updated on 2026-08-26.

How to cite this study

This study is free to quote, in an article, a deck or a report. Please cite the source and the measurement period, because the numbers move every month.

Sorank (2026). Wikipedia in AI Answers: a Data Study of Its Decline over 13 Months. Sorank GEO studies, measured on 301025 AI answers collected between August 2025 and August 2026. Available at https://www.sorank.com/data/wikipedia-ai-citation-decline-study

Related studies

The same panel, read from another angle:

Measure your own citations

This study describes a market. Sorank measures your site: which questions mention you, which mention someone else, and what to publish to change that, tracked weekly across the same five AI.

Start with 3 days free. A payment card is required to open the trial.

Frequently questions asked

Which AI cites Wikipedia the most?

Wikipedia is cited by all five models measured. ChatGPT cites it most, in 3.43% of its answers, and Claude least, in 0.03%. The full model by model breakdown is on this page.

How was this study measured?

On 301025 answers from five AI, collected between August 2025 and August 2026, with every cited source extracted and normalised. The citation rate is the number of citations per 100 measured answers. The public dataset is on data.sorank.com.

Does this mean Wikipedia is losing importance everywhere?

No. On the most conservative model it remains the leading source. The decline is real at panel level and uneven by model, which is exactly why an average is a poor basis for a decision.

Our Blog for Ambitious Company