AI doesn’t pick the best nonprofit. It picks the best passage.
How AI search decides who to recommend, why the answers keep changing, and what a small organization can control.
A donor asks an AI tool where to give. It names a few organizations in a paragraph. It’s tempting to assume the tool weighed them and found them worthy.
It didn’t. It ran a process, and each organization either survived every stage or dropped out at one. The process is inconsistent in at least four ways, and each has a different fix.
How the process works
The internals of ChatGPT, Perplexity, and Gemini aren’t public, and they change often. But the broad pattern is documented, and Google has described its own version.
- The question gets split. AI search engines break each prompt into multiple sub-queries before choosing sources. Google’s name for this is query fan-out: breaking a question into subtopics and issuing many queries at once. airopssurferseo
- Pages get retrieved. Keyword and vector search run together, and the results are merged so documents that rank well in both rise. peec
- Passages get re-scored. A reranking model reads each candidate passage against the query, and a strong page can contain a weak passage. peec
- An answer gets written from the surviving passages, and some get cited.
One consequence: you’re rarely competing for one keyword. You’re competing to be the best source across a neighborhood of related sub-questions the engine generates itself. geotoolbox
Four kinds of inconsistency
1. The tools disagree with each other. Each engine generates its own sub-queries and has its own retrieval, so the same question can trigger different searches. That’s my inference from the mechanism, and I haven’t tested it directly. One study of five million fan-outs found AI platforms rewrite and expand queries before searching. neilpatel
2. Wording decides who qualifies. On October 4, 2026, I asked Gemini “food help in the Mission.” It returned five restaurants, each with a rating and an open-now status. I changed it to “food pantry in the Mission,” and it returned Mission Food Hub, Mission Action, and the Salvation Army. One word separated a person who needed groceries from a brunch recommendation.
A different question showed the same effect in a different way. I asked for neighborhood meal programs with low administrative overhead. The answer named three large organizations backed by expense ratios and charity scores, then explained that small grassroots groups were left out because they don’t publish standard cost ratios or structured financial data. [Confirm the tool and exact prompt.] The tool wasn’t saying those groups ran worse programs. It had no number to put next to them.
3. You disagree with yourself. I looked up Mission Food Hub by name in three tools. They agreed on the basics: founded May 2020, 701 Alabama St., culturally appropriate groceries, run under Carnaval San Francisco/CANA. They disagreed on what a hungry person needs. Perplexity laid out why: the website listed Monday, Wednesday, and Friday distributions at 10 a.m.; Instagram said Fridays at 11 a.m. at a different address; and local coverage said Fridays from about 11 to 2:30. Across the runs, two phone numbers also appeared. The tool wasn’t hallucinating. It was reporting that the organization’s own channels didn’t match.
4. The tool remembers. During these sessions, one assistant tailored its answer to my past work, referring to my interest in community greening when I’d asked about something else. Another listed my earlier searches as a source. These are signs of how the products are designed, and I can’t say how much they changed which organizations were named, since I didn’t compare the same prompt with and without history.
The consequence is uncomfortable for a nonprofit. Two donors asking the same question can get different answers based on what each has asked before, where they are, and which settings they’ve turned on. There’s no neutral version to check and no way to see what a given donor sees. A clean audit tells you a baseline, not what any particular person gets.
So don’t aim to win a ranking. Aim to be accurate and consistent in the places every version of the answer draws from: your website, your listings, your press, your funders’ directories. If those agree, most variations come out right.
Three ways to fail, three fixes
The most useful idea I found is a diagnostic one from Peec’s research, which I’ve adapted: fix the stage that failed first. Rewriting a paragraph for the reranker can’t fix a page that was never retrieved. peec
| Stage | What failure looks like | Likely cause | What a nonprofit can do |
|---|---|---|---|
| Not retrieved | You don’t appear for anything | Nothing crawlable, no pages matching the sub-questions, or a blocked crawler | Have real web pages in text, not a PDF or social-only presence. Cover the specific questions donors and clients ask. Check that your site isn’t blocking search bots |
| Retrieved but losing | You surface for some questions and never get named | The relevant passage is buried, vague, or mismatched | Lead each page with a plain answer: what you do, who you serve, where, when |
| Good passage, no citation | Your page is strong, but others get picked | Thin corroboration or inconsistent facts across sources | Make facts match across your site, listings, and press. Get described by outside sources |
Small organizations are most likely to fail at the first stage, and it has nothing to do with the quality of their work. One easy-to-miss trap: a blanket AI block in robots.txt can catch search bots along with training crawlers and drop a site from the citation pool without any ranking change. contextbolt
What helps, in order
-
- Be retrievable. A real, crawlable page for each program, in text.
- Answer the specific question. Because of fan-out, ten pages each answering one question beat one page covering ten. Write for “Is this free?” “Who qualifies?” “What do I bring?” in the words your community uses, like “food pantry” and “free groceries,” not program jargon.
- Lead with the answer. Passages are scored on their own, so a founder story above the answer makes a weak candidate.
- Publish what the question needs. Address hours and eligibility for “where” questions. Budget and program expense ratio for “can I trust you with my money” questions.
- Be corroborated and consistent. Complete your Candid profile, 211 listing, and any food bank or funder directory, and make them match your site. When something changes, update every channel the same day.
- Be present where people talk. Peec found ChatGPT fan-outs mentioning Reddit rose from about 0.15% to 3.68% between January and May 2026, which the authors read as a pull toward experiential content. What residents and volunteers say about you may matter as much as what you say about yourself. neilpatel
What doesn’t help is chasing individual keywords. Surfer’s analysis of 1,600 fan-out runs found most fan-outs sit close to, but not identical to, the original query. Cover the neighborhood of questions rather than a single phrase. surferseo
Size isn’t destiny, but infrastructure is
The assumption is that AI favors big, famous organizations. The evidence is mixed. One March 2026 analysis found only 37.9% of AI Overview citations also ranked in the top 10 for the same query. A SIGIR 2026 study, as Surfer summarizes, found that generative search engines are significantly less likely than traditional search engines to retrieve sources from popular websites. A small, specific, well-written page has a real chance. contextboltsurferseo
The unfairness is lower down. The early stages reward infrastructure: a crawlable site, someone to maintain it, and up-to-date listings. Those are what the smallest, most community-rooted groups tend to lack. A tool can’t recommend an organization it never retrieved, and “never retrieved” looks the same whether the cause is a great pantry with a Facebook page or an organization that doesn’t exist.
That points to who can fix it. Funders can pay for a basic web presence and listing upkeep inside a grant. Community foundations and food banks can publish open, structured partner directories. Local press can run “who’s doing the work” roundups. None of it requires small nonprofits to become technical.
One thing my own checks suggested, and I’d hold it loosely: when I asked about a specific local organization, all three tools described it accurately, including what it doesn’t do. Asking about a known name seemed to work better than asking “who should I give to.” That fits how retrieval works, but it’s one case.
What I’m confident about, and what I’m not
Firmer ground: AI search splits prompts into sub-queries (Google says so itself), and retrieval, reranking, and generation are separate stages that can fail separately.
Less certain:
- Most of the quantitative research comes from companies that sell AI-visibility tools, and some figures reach me secondhand.
- Each product’s internals aren’t public and change often.
- The mapping from failed stage to nonprofit action is my reasoning, not a tested intervention.
- I haven’t found research specific to nonprofits, so I’m assuming findings about brands carry over.
- My own runs were informal, from one day to the next, one city to the next, some with personalization in play. They illustrate the mechanism. They don’t prove it.
Follow us on LinkedIn – Build, Grow, Convert.
About Mike Doherty
Mike Doherty serves as Chief Experience Officer at Greening Projects, a nonprofit organization dedicated to transforming underutilized urban spaces into vibrant green areas that benefit communities and the environment. With a passion for urban revitalization and community-centered approaches, Mike oversees the end-to-end experience of residents, volunteers, municipal partners, and donors involved in the organization’s green space conversion projects. His role encompasses strategic vision, community engagement, and ensuring that every interaction reflects Greening Projects’ commitment to creating accessible, sustainable urban oases. Under his leadership, the experienced team focuses on making green space development collaborative, impactful, and meaningful for all stakeholders while fostering stronger, healthier neighborhoods through environmental transformation.
