A prospective donor opens ChatGPT and types "what charities are working on childhood malnutrition in South Asia, and which ones are reputable?" A volunteer asks Perplexity "who near me helps resettle refugees?" A journalist on deadline asks Google's AI overview for "the most-cited research on safe drinking water access." In each case a machine writes the answer, names a few organizations, and links to a handful of sources. If your nonprofit isn't one of them, you were never in the room.
That moment matters more for nonprofits than for most businesses, because trust and discovery are the whole game. Here's the part most teams miss: you are probably better positioned to win it than the average company, and you're losing the advantage to fixable problems. AI answer engines reach for content that is factual, specific, well-sourced, and free of sales pressure. That describes a good impact report far better than it describes a product landing page. The catch is that your best material is usually buried in a PDF, undated, or written as a wall of prose a model can't cleanly quote.
This guide walks through how to analyze what AI engines currently see when they look at your organization, and the concrete fixes that move you from invisible to citable. Most of it is editing and clean HTML, not budget.
What GEO is, and why it's not separate from SEO
GEO stands for generative engine optimization: making your content easy for AI answer engines like ChatGPT, Claude, Perplexity, and Google's AI overviews to find, understand, and cite. SEO is making your content easy for a search engine to rank in a list of links. The two overlap far more than the new acronym suggests, and the practical differences between GEO and SEO are narrower than the hype implies. A page that's fast, crawlable, well-structured, and clearly written tends to do well in both. If you've neglected SEO basics, GEO is not a fresh start that lets you skip them. It's the same foundation with a second reader in mind.
So the honest first move for a nonprofit with a small team is to fix the basics that serve Google and AI engines at once, then add the GEO-specific touches. Chasing AI citations on a site that Googlebot can't crawl is painting a room with a hole in the floor.
Why nonprofits start ahead
Three things models look for when deciding what to cite line up with what mission-driven organizations already produce.
Models prefer factual, verifiable claims over promotional language. Your annual report is full of them: meals delivered, wells dug, cases handled, dollars spent per program. Models prefer named, authoritative sources on a topic. A nonprofit that has worked in one field for fifteen years is exactly that kind of source. And models weight content that other reputable sites reference. Nonprofits get cited by news outlets, academic work, and government pages in ways most commercial sites never do.
The problem is rarely a lack of authority. It's that the authority is locked up. Impact figures live inside a 40-page PDF that a model won't open or can't parse cleanly. The "About" page describes feelings instead of facts. The strongest research sits behind a "request the report" form. You have what the engine wants. You're just not handing it over in a form it can use.
How to analyze your current AI presence
Before you change anything, find out what the engines already say. This is the audit, and you can do most of it yourself in an afternoon.
1. Ask the engines about yourself
Open ChatGPT, Claude, Perplexity, and Google's AI overview, and ask the questions a donor or volunteer would actually type. "Is [your organization] a legitimate charity?" "What does [your organization] do?" "Best nonprofits for [your cause] in [your region]." Read the answers like an auditor, not a proud founder. Are the facts right? Is your impact data current or five years stale? Does the model cite your own site, or a third party describing you? When it's wrong, note where the wrong information seems to come from. Perplexity is useful here because it shows its sources, so you can see exactly which pages the model trusted.
2. Check what's even crawlable
If an engine can't fetch your pages, none of the rest matters. In Google Search Console, look at the Pages report for what's indexed versus excluded. Run a site:yourdomain.org search to see roughly what Google holds. Then check your robots.txt file at yourdomain.org/robots.txt for two separate things: whether normal crawlers are blocked, and whether AI crawlers are.
AI crawlers identify themselves with specific user-agents: GPTBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot, and Google-Extended (Google's control for whether your content trains and feeds Gemini, which is separate from Googlebot). Plenty of nonprofits block these without realizing it, sometimes because a security plugin or a well-meaning developer added a blanket block. Blocking them is a legitimate choice, but make it on purpose. If you block GPTBot, you are opting out of being read by the engine behind ChatGPT. That's a strategy decision, not a checkbox a plugin should make for you.
3. Check the technical floor
Run your top few pages through PageSpeed Insights. You're looking at Core Web Vitals, the loading and stability metrics Google treats as a ranking input. The "good" thresholds are an LCP (largest contentful paint, when the main content appears) at or under 2.5 seconds, an INP (interaction to next paint, how fast the page responds to a tap) at or under 200 milliseconds, and a CLS (cumulative layout shift, how much the page jumps around as it loads) at or under 0.1. Nonprofit sites built on heavy themes with a stack of donation and tracking scripts routinely miss the LCP target on mobile, and a slow page is both a worse experience for a donor on a phone and a weaker signal to the engines. Confirm HTTPS is in place and the site is genuinely usable on a phone, since most of your traffic is mobile.
4. Run an accessibility pass
Run a page or two through axe DevTools or WAVE. This matters twice over for a nonprofit. First, serving people with disabilities is usually core to the mission, and an inaccessible site quietly excludes them. Second, accessibility and machine-readability come from the same place: semantic HTML. The most common failures we see are low color contrast (WCAG 2.1 success criterion 1.4.3 wants at least 4.5:1 for normal text), missing visible focus states for keyboard users (criterion 2.4.7), images without alt text, and headings used for visual size rather than document structure. A model parsing your page leans on the same heading hierarchy and alt text a screen reader does. Fix it once, help both.
5. Find your trapped content
Make a list of your most valuable facts and figures, then check where each one lives. If your flagship impact number exists only inside a downloadable PDF or an image of an infographic, treat it as invisible. Get it into real HTML text on a real page.
The fixes that move the needle
Once you know what the engines see, here's where to spend effort, roughly in priority order.
Lead with the answer. Under a heading that mirrors a real question, put a direct one- or two-sentence answer first, then expand. "How is my donation used?" should be followed immediately by "About 84 cents of every dollar goes directly to programs" (your real figure), not three paragraphs of mission philosophy before the number arrives. This is what we call answerability: models skim for the passage that answers the query. Make it the first thing under the heading.
Write claims that survive being quoted. A model cites sentences it can lift off your page and still trust out of context. "In 2025 we provided 1.2 million meals across 14 districts" is quotable: it names the subject, the magnitude, and the scope. "We helped so many people this year" is useless the moment it's separated from your page. Read each key sentence as if it were pasted into a stranger's document. If it still means something specific, it's quotable. Use your own real numbers, never invented ones.
Get impact data out of PDFs and into pages. Your annual report can stay a PDF for the board. But the headline figures, program outcomes, and financial breakdown need to also exist as HTML text, ideally in a table, on an indexable page. A table of "program, people served, spend" is far easier for both a model and a screen reader to extract than the same data trapped in a designed infographic.
Make your identity unambiguous. Models build a picture of your organization as an entity. Help them by using your full legal name consistently, keeping one canonical "About" page that states what you do, where, and since when, and making sure your profiles on Candid (GuideStar), Charity Navigator, and Wikidata are accurate. These third-party sources feed model knowledge, and when they disagree with your own site, the model gets confused about who you are.
Add structured data. Schema.org markup tells engines what a page is in a format they parse directly. Use the Organization type (or its NGO subtype) on your About page with your name, logo, and social profiles, FAQPage markup on any FAQ, and Article markup on posts. You can validate it with Google's Rich Results Test. This is developer work, but small and well-documented.
Don't block the AI crawlers, unless you mean to. If step 2 turned up a block on GPTBot, ClaudeBot, PerplexityBot, or Google-Extended that nobody decided on deliberately, remove it. Allowing them is what makes citation possible.
Keep content dated and current. Models favor content that looks maintained. Show a real "last updated" date, refresh your impact figures when new ones land, and retire pages describing programs you've ended. A 2019 statistic presented as current does more harm than no statistic.
Consider an llms.txt file. This is a young, voluntary convention: a plain-text Markdown file at yourdomain.org/llms.txt that points AI engines at the handful of pages worth reading, each with a one-line description. No engine requires it and it won't get you cited on its own. It takes twenty minutes and signals you've thought about how machines read you. Low priority, low cost, worth doing once the rest is in place.
Being found is only half the journey. Once a donor lands on your answer, make giving frictionless: donate-ready QR codes built for small NGOs turn a phone-in-hand moment into a completed gift instead of a hunt for the donate button.
Honest limitations and common mistakes
GEO is not measurable the way SEO is. There's no AI-citation equivalent of Search Console showing exactly when ChatGPT mentioned you. You can watch referral traffic in GA4 from sources like chatgpt.com and perplexity.ai, keep an eye on branded search in Search Console, and re-run the step-1 questions every month or two to see whether the answers improve. Treat it as a trend you nudge, not a dial you set.
A few mistakes to avoid. Don't write for the machine at the cost of the donor. A page stuffed with question-headings and keyword-loaded sentences reads as hollow to the human who actually decides whether to give. Don't expose data you shouldn't. "More crawlable" is not the goal when it means publishing beneficiary names, locations, or anything that puts vulnerable people at risk. Privacy outranks visibility every time. And don't expect overnight results. Models update their picture of the web on their own schedule, and citation is earned over months.
There's also no guarantee. Nobody can promise a citation in ChatGPT any more than anyone could promise a number-one Google ranking. What you can do is make your organization the easiest, most trustworthy source for a model to reach for, then let the work compound.
Key takeaways
- AI answer engines now sit between many donors, volunteers, and journalists and your nonprofit. If the model doesn't cite you, you're invisible at the moment of intent.
- Nonprofits start with a GEO advantage: factual, authoritative, non-promotional content is exactly what models cite. The usual problem is that this content is trapped in PDFs, undated, or written as unquotable prose.
- Audit first: ask the engines about yourself, check crawlability and AI-crawler access in
robots.txt, test Core Web Vitals in PageSpeed Insights, run an accessibility pass, and find your trapped data. - Then fix: lead with the answer, write quotable claims with real numbers, move impact data into HTML, make your identity consistent across Candid and Charity Navigator, add
Organization/FAQPagestructured data, and stop blocking AI crawlers by accident. - GEO and SEO share one foundation. Fix the basics that serve both before chasing AI-specific tactics, and never trade donor clarity or beneficiary privacy for machine-readability.
FAQ
Should nonprofits block AI crawlers like GPTBot?
Only if you have a specific reason to. Blocking GPTBot, ClaudeBot, or PerplexityBot in robots.txt opts you out of being read and cited by those engines. For most nonprofits whose goal is wider reach, allowing them is the right call. The exception is content involving vulnerable people or sensitive data, which shouldn't be widely published in the first place.
Is GEO actually different from SEO for a nonprofit?
They share most of the same foundation: crawlability, speed, clean structure, clear writing. GEO adds an emphasis on self-contained, quotable claims, accurate third-party entity data, and structured markup so a model can extract a clean answer. If you've done SEO well, you're most of the way to GEO already.
Will any of this hurt our Google ranking?
No. Everything here, faster pages, semantic HTML, structured data, accurate content, and accessibility, is also what Google rewards. There's no tension between optimizing for AI engines and for traditional search.
Do we need an llms.txt file?
It's optional and low-priority. It's a voluntary convention that no engine requires, and it won't get you cited on its own. It's cheap to add once the higher-impact fixes (crawlability, structure, quotable content, structured data) are done, so treat it as a finishing touch rather than a starting point.
