Why AI may favour large charities: the visibility gap and Citewake's open question H-008

Some commentators think AI assistants favour large, well-documented charities and leave small ones out. The evidence so far is a handful of early reports and small studies, none of them a systematic test across the four main assistants. Citewake treats it as an open research question, registers it as H-008, and will test it from February 2027.

This guide sets out what has been said, what has been measured, and exactly how Citewake will test it. Citewake has no charity results yet.

What is the visibility gap?

The phrase describes a difference in how easily donors can find different organisations. In an October 2026 article for Stanford Social Innovation Review, Andrew Dunckelman and Dale Nirvani Pfeifer ("The New Nonprofit Visibility Gap") argue that search once gave donors a menu of options to explore, while AI tends to give one confident answer that favours large, well-documented organisations (SSIR). Their main example is a donor asking where to give to save children's lives, and the chat tools recommending GiveWell's top charities. They stress that the advice is worthwhile and that the concern is what is left out.

A peer-reviewed study in The Foundation Review (Alegria and Barnett, Giving Compass, 2026) addresses a related problem. Donors say they want to give locally but often give to large, nationally known nonprofits. In a 2024 evaluation, 291 US donors were randomly assigned to use either the Giving Compass platform or general internet tools to choose a nonprofit for a $50 donation. Users of the platform were significantly more likely to give locally and more often chose nonprofits with smaller net assets (The Foundation Review). That study tested a purpose-built tool, not ChatGPT, Claude, Gemini or Perplexity.

What has anyone measured about AI recommendations?

Whole Whale, a nonprofit consultancy, put 255 questions across 17 cause areas to seven AI models, and collected answers both from the models' memory and with web search on. It reports that in breast cancer the single top organisation received 13% of all mentions and the top ten received 60%, while in poverty the leader received 4%, so recommendations there were more spread out (Whole Whale, 2026).

Two points about that study matter. First, concentration differed by cause, so a single "AI favours big charities" statement hides variation. Second, the authors used AI to count and de-duplicate organisation names and caution against reading the numbers literally. They ran the questions through programming interfaces rather than the consumer apps. It is a useful proof of concept and not a final answer.

Why might AI favour large charities?

These are possible reasons, not findings.

• More public documentation. Large charities tend to have more web pages, news coverage and register entries to draw on. SSIR's authors say recommendations favour well-documented organisations.
• Evaluator coverage. Assistants may lean on evaluators and registers. Whole Whale reports Charity Navigator and GuideStar appearing across most models. A charity no evaluator has rated has less to be recommended from.
• Name recognition. A model that learned about a charity from training data may repeat the best-known names.

Each can be tested. None has been established for these four assistants.

What exactly is H-008?

H-008 is entry eight in Citewake's public research register, on the research page. It reads: when people ask AI where to give, recommendations concentrate on a small number of large, well-known charities, and smaller and local charities rarely appear, even when the question names a city or region. It was registered on 29 October 2026, before any charity data was collected.

It has two registered tests:

• H-008a, concentration. Supported if the five most-recommended organisations receive more than 50% of all recommendations in at least four of five charity categories, in each market panel.
• H-008b, small charities. Supported if organisations with annual income under US$1 million (or the local equivalent), as shown by their country's official register, receive under 10% of recommendations in each category and panel, including on questions that name a city or region. An organisation whose income cannot be found in an official register is reported separately and never counted as small.

The five categories are animal charities, environmental charities, children's charities, humanitarian aid organisations and health charities. They sit in Citewake's categories. Results will be reported for the United States and for a composite of European countries, never pooled.

Why register it in advance?

Because it stops the analysis from being shaped by the results. The thresholds, the categories and the question type were fixed before any answer was collected. If the data do not meet the thresholds, the hypothesis is reported as not supported. For why this matters in general, see reproducibility in AI search research and how to build a benchmark set of questions.

When will there be an answer?

The first charity edition is due from February 2027, then quarterly, and results will be published as Special Edition: Charities on the editions pages. The method is on the methodology page. Until then, the sensible position is to stay open: the gap may be real, large, small or different by cause. If you run a small charity, the practical steps do not depend on the answer. See how to check whether an AI recommends your charity.