Auditing Your Nonprofit's AI Footprint: What ChatGPT, Gemini, and Perplexity Say About You
Your organization already has a reputation inside AI assistants, whether you have looked at it or not. Donors, funders, journalists, and job candidates are asking these systems about you, and the answers they get are assembled from sources you may never have reviewed. An AI footprint audit is the disciplined way to find out what is being said, where it comes from, and what you can actually do about it.
Most nonprofit leaders have, at some point, typed their organization's name into ChatGPT out of curiosity. It is a strange moment. The assistant produces a confident paragraph about who you are and what you do, and it is usually about eighty percent right. The remaining twenty percent is where the trouble lives: an outdated program that ended two years ago, a budget figure from an old filing, a founder's name spelled wrong, or a description that makes your work sound like something adjacent to what you actually do.
That one-off curiosity check is not an audit. An audit is systematic. It asks a consistent set of questions, across multiple assistants, more than once, and records what comes back in a form you can compare over time. The difference matters, because a single answer tells you almost nothing. These systems vary their responses between runs, weight different sources, and change behavior as models are updated. Only a repeatable process turns anecdote into information you can act on.
The stakes have risen quickly. AI assistants are now a routine first stop for people evaluating charities, and the answers they produce increasingly shape decisions before any human conversation happens. A major donor's family office, a corporate partnerships manager, a foundation program officer doing a quick sanity check, a reporter deciding whom to call: all of them may form an impression of your organization from a synthesized paragraph. Our analysis of how AI assistants decide which charities to recommend covers the underlying mechanics, and the audit described here is how you find out where you personally stand.
This article walks through the whole process. You will learn what an AI footprint actually consists of, how to design a prompt set that reflects the questions people really ask, how to run the audit across the major assistants without special tools, how to score what you find, what the common failure patterns mean, and how to fix the underlying sources rather than chasing individual bad answers. The work takes a few focused hours the first time and far less on each repeat.
What an AI Footprint Actually Is
Your AI footprint is the aggregate of everything an AI assistant can say about your organization when asked. It is not stored anywhere as a profile you can log into and edit. It is reconstructed on demand from two ingredients: what the model absorbed during training, and what it retrieves live from the web at the moment of the question. Both ingredients matter, and they behave very differently.
Training knowledge is baked in and slow to change. If your organization was described a certain way across the web during the period a model was trained, that framing persists until a newer model supersedes it. This is why an assistant sometimes describes a program you sunset years ago with complete confidence. You cannot edit that memory directly. You can only change the public record enough that future training and live retrieval reflect something better.
Retrieval is faster and more addressable. Assistants that search the web while answering will pull from whatever they can find and cite right now, which means your current website, your third-party profiles, recent news coverage, and directory listings all have real influence. This is the part of your footprint you can move within weeks rather than years, and it is where an audit produces the most immediately actionable findings.
One more distinction is worth drawing clearly. Your search ranking and your AI footprint are related but not the same thing. Ranking well for your own name in traditional search does not guarantee that an assistant will mention you when someone asks a category question like which organizations serve homeless youth in a given city. Assistants synthesize across sources rather than ranking pages, so a nonprofit with modest search traffic but strong, consistent third-party data can outperform a better-optimized peer. We explore that divergence in depth in our comparison of generative engine optimization and traditional SEO.
Presence
Does the assistant know you exist, and does it surface you when someone asks about your cause, your geography, or your service category without naming you? Presence is the foundation. An organization that only appears when named directly is invisible in exactly the moments that matter for new supporter acquisition.
Accuracy
Are the facts right? This covers your legal name, your location, your leadership, your program list, your service area, your founding year, and your financial scale. Factual errors are the most damaging findings because they are the easiest for a skeptical evaluator to notice and the hardest to explain away later.
Framing
Even when every fact is correct, the emphasis can be wrong. An assistant might describe a workforce development organization primarily as a food pantry because the pantry gets more press. Framing problems are subtle, they rarely feel like errors, and they quietly steer donors toward or away from the work you most want funded.
Sources
Which pages does the assistant cite or lean on? This is the most operationally useful part of the audit, because sources are things you can influence. If every answer about you traces back to a single outdated directory listing, you have found both your problem and your first fix.
Designing a Prompt Set That Reflects Reality
The single biggest mistake in a first audit is testing only your own name. Asking an assistant to tell you about your organization is the easiest question it will ever face, and the answer is usually flattering. The questions that determine whether you gain a supporter are the ones where you are not mentioned at all, where the person is describing a need and asking who can meet it.
Build a prompt set of roughly fifteen to twenty five questions across several categories. Write them the way an actual person would type them, in plain language, sometimes messily. Include your geography, because most giving decisions are local or regional and assistants weight place heavily. Include the sceptical questions too, because evaluators ask those, and you want to know what comes back.
Categories to Cover
A balanced prompt set tests discovery, evaluation, and comparison, not just recall
- Direct identity: what does your organization do, where is it based, who leads it, how big is it. These establish baseline accuracy.
- Unnamed category discovery: which organizations help with your issue in your city or region. This is the acquisition question, and the hardest one to pass.
- Donor evaluation: is this organization financially responsible, how much goes to programs, is it well rated, is it trustworthy. Expect ratings data to dominate these answers.
- Comparison: how does your organization compare to two named peers, and which should someone support for a specific goal. Comparison prompts reveal your positioning starkly.
- Service seeker: where can someone get the help you provide. If you deliver direct services, being absent here is a mission failure, not just a marketing one.
- Adversarial: has this organization faced criticism, controversy, or complaints. You want to know what surfaces before a funder finds it.
- Practical logistics: how to volunteer, how to donate, what the hours are, how to reach a person. Wrong operational details cost you real participants.
Write the prompt set once and then freeze it. The value of the audit compounds only if you ask the same questions each time, so resist the temptation to reword prompts on later runs because you thought of a better phrasing. Keep a separate list for new prompt ideas and add them as additions rather than replacements, so your trend data stays intact.
One practical note on scale. A small organization can run a meaningful audit with fifteen prompts across three assistants, which is about an hour of work. A larger organization with multiple program lines and service areas should build a prompt set per program, because an assistant may represent one part of your work well and another part not at all. Averaging across the whole organization hides exactly the gap you need to see.
Running the Audit Without Special Tools
A commercial AI visibility platform will automate this work, and there is a growing market of them. They are useful at scale and they cost money. For most nonprofits, a spreadsheet and a disciplined afternoon produce ninety percent of the insight at zero cost, and doing it manually the first time teaches your team far more than reading a dashboard would. Start manual. Buy tooling later if the volume justifies it.
Step One: Neutralize Your Own Context
Assistants personalize. If you have spent months chatting with an assistant about your own organization, it may remember you and give a flattering, context-laden answer that no stranger would ever receive. Before you begin, turn off memory and personalization features, use a logged-out or temporary chat session where the product allows it, and avoid signalling who you are in the prompt.
This step is not optional. An audit run inside a personalized session measures your own history, not your public footprint, and it will make you feel much better than the facts warrant.
Step Two: Run Each Prompt Multiple Times
These systems are probabilistic. The same question can produce a different list of organizations on consecutive runs. Run each prompt three times per assistant in fresh sessions and record all three. What you are looking for is not a single answer but a pattern: appearing in three of three runs is a strong position, one of three is fragile, zero of three is a gap.
This is also the step people skip, and skipping it is how organizations conclude they are doing fine based on one lucky response.
Step Three: Cover the Major Assistants
Test ChatGPT, Google's Gemini and AI Overviews, Perplexity, and Claude at minimum. They behave differently. Perplexity is citation-heavy and leans on live retrieval, which makes it excellent for identifying which specific pages shape your footprint. Google's surfaces carry enormous reach because they appear in ordinary search. ChatGPT has the largest general audience.
Do not assume a good result on one transfers to another. Divergence between assistants is one of the most common and most useful findings, because it usually points to a specific source that one system trusts and another does not.
Step Four: Record Structured Observations
For each run, capture the assistant, the date, the prompt, whether you were mentioned, your position in any list, which other organizations appeared, any factual errors, and every source cited. Paste the full response text into a notes column. Screenshots are helpful for anything you may need to show a board.
The citation column is the one that earns its keep. After a few dozen runs, a pattern emerges showing the handful of domains that are effectively writing your public description, and that list becomes your remediation plan.
Set aside about three hours for a first full pass with a twenty prompt set and three runs across four assistants. That sounds like a lot until you compare it to the cost of a single misinformed major donor conversation. Subsequent quarterly runs take less than half the time, because the spreadsheet already exists and you are only filling in a new column.
Turning Raw Answers Into a Score You Can Track
A pile of pasted responses is evidence, not information. To make the audit useful over time, and to make it presentable to a board or a leadership team, convert what you observed into a handful of simple measures. Precision is not the goal here. Consistency is. A crude score applied the same way every quarter tells you far more than an elaborate one applied once.
Four Measures Worth Keeping
Simple, repeatable, and understandable by people who were not in the room
- Mention rate: the share of all runs where your organization appeared at all. Track it overall and separately for the unnamed category prompts, which are the ones that matter for growth.
- Accuracy rate: the share of mentions that contained no factual error. One wrong figure counts as an error even if everything else was right, because that is how a reader will treat it.
- Framing quality: a simple three point judgment on whether the description led with your core work, mentioned it secondarily, or missed it. Have two people score independently and discuss disagreements.
- Source concentration: the number of distinct domains cited across all runs, and the share held by the top three. High concentration means a single source is dictating your reputation, which is both a risk and an opportunity.
Add one qualitative column that no score can replace: the single sentence from the audit that would most concern a donor. Boards remember that sentence long after they have forgotten the percentages, and it is usually the thing that unlocks the budget or attention needed to fix the underlying problem.
Resist the urge to benchmark your scores against other organizations. There is no reliable public dataset of nonprofit AI visibility, and any external comparison you construct from a handful of peer audits will be too noisy to trust. Benchmark against your own previous quarter instead. Movement in your own numbers is the only comparison you can defend.
Common Findings and What They Actually Mean
Audits tend to surface the same handful of problems across very different organizations. Recognizing the pattern quickly saves you from treating a structural issue as a one-off mistake, and it points you toward the right kind of fix.
Invisible in Category Questions
You appear reliably when named and never when not. This is the most common finding and the most consequential. It means the assistant has a record of you but does not associate you strongly with your cause and place. The fix is usually about consistent, machine-readable descriptions of what you do and where, repeated across the third-party sources these systems trust, rather than anything on your own website.
Confused With a Similarly Named Organization
Nonprofit naming conventions produce a lot of collisions, and assistants merge entities that share a name or an acronym. If answers about you contain facts belonging to another organization, your problem is entity disambiguation. Consistent use of your full legal name alongside your common name, plus a clear and distinct presence in structured directories, is what separates you again.
Stale Financials and Sunset Programs
Assistants lean on filings and directory profiles, which lag reality by a year or more. Expect your budget figures to be dated. The concerning version is when a program you discontinued still dominates your description, or a program you launched last year is entirely absent, because that gap tells funders you are something other than what you now are.
Crowded Out by Larger Peers
Category answers name the same few national organizations regardless of the local framing of the question. This reflects source volume rather than merit. Local and specific presence is the counterweight: regional news coverage, local directories, community foundation listings, and clear geographic language in your own materials all help an assistant understand that the question about your city has a local answer.
Invented Details
Occasionally an assistant will produce a plausible fabrication: a program that never existed, an award you never won, a partnership you never had. Fabrications thrive where public information is thin, because the model fills gaps with what usually goes there. The remedy is not correction but supply. Publishing clear, complete, easily retrievable facts reduces the space in which invention happens.
Notice that almost none of these findings are fixed by editing your homepage copy. That is the central lesson of a first audit. Your footprint is built mostly from what other sources say about you, which is why the remediation work looks more like reputation infrastructure than marketing.
Fixing the Sources Rather Than the Answers
There is no complaints department for an AI assistant. You cannot file a correction and watch the answer change tomorrow. What you can do is change the inputs, consistently and patiently, until the synthesized output follows. This is slower than a website edit and considerably more durable.
Start with a single canonical fact sheet held internally: legal name, common name, founding year, mission in one sentence, current program list with one line each, service area, leadership, current budget scale, and contact details. Everything else flows from this document. The most common cause of a confused AI footprint is an organization that describes itself six different ways across six different places, and no assistant can resolve that inconsistency into a clean answer.
Third-Party Profiles
Nonprofit directory and ratings profiles are disproportionately influential because they are structured, comprehensive, and treated as authoritative. Bringing every profile into alignment with your canonical fact sheet, and completing the optional fields most organizations leave blank, is usually the single highest-return action an audit produces.
Your Own Site's Machine Readability
A clear about page, a plainly written program list, an accessible leadership page, and correct structured data give retrieval systems something unambiguous to work with. The technical side of this is more approachable than it sounds, and our checklist for llms.txt, schema markup, and AI crawlers walks through it step by step.
Trust and Transparency Signals
Financial documents, governance details, and impact reporting published in accessible formats give assistants something concrete to cite when someone asks whether you are trustworthy. This overlaps heavily with what a careful human evaluator wants, which is discussed further in our piece on optimizing for machine-readable trust signals.
Independent Coverage
Sources that are not you carry more weight than sources that are. Local journalism, community foundation writeups, partner organization pages, university and government program listings, and credible sector publications all reinforce the same facts from outside. A modest amount of this coverage does more for category visibility than a great deal of self-published content.
Sequence the work by leverage rather than by ease. Fixing an inconsistent legal name across every profile is unglamorous and takes an afternoon, and it will do more for your footprint than a month of content production. Save the ambitious projects for after the foundational cleanup is finished, because ambitious content built on inconsistent facts just adds another conflicting source.
Expect a lag. Changes to directory profiles and website structure can show up in retrieval-based answers within weeks, while training-baked descriptions may take a model generation to shift. Plan your remediation on a six to twelve month horizon and judge progress by your audit trend rather than by whether a specific answer changed this week.
Making the Audit a Routine Instead of a Project
A one-time audit produces a snapshot and a to-do list. A recurring audit produces something considerably more valuable: an early warning system. Because these systems change underneath you without notice, a description that was accurate in the spring can drift by the autumn, and you will only know if someone is looking.
Quarterly is the right default cadence for most organizations. Anything more frequent produces noise that is hard to distinguish from ordinary model variation. Anything less frequent means a serious problem can sit unnoticed through an entire fundraising season. Add an unscheduled run whenever something significant changes: a rebrand, a merger, a leadership transition, a new flagship program, or any moment of negative press.
Assign ownership to a specific person rather than a committee. This works well as a defined responsibility for a communications or development staff member, with a standing calendar block and a short written summary that goes to leadership. Organizations that have built internal capacity through an AI champions program often find this is a natural assignment for someone already curious about how these systems behave.
Keep the reporting short. One page with the four measures, the trend against last quarter, the most concerning sentence observed, and the three actions being taken is enough for a board. Anything longer will be skimmed, and the point of the exercise is decisions rather than documentation. If you want a broader frame for how this fits into organizational AI readiness, our fifteen minute AI audit offers a companion lens focused inward rather than outward.
Mistakes That Undermine an Otherwise Good Audit
- Auditing while logged in: personalized sessions flatter you and produce results no stranger would ever see.
- Testing only your name: the easiest question, the least informative answer, and the reason many organizations think they are fine.
- Running each prompt once: a single response is a sample of one from a system that varies deliberately.
- Changing the prompts between runs: destroys comparability and turns a trend into a series of unrelated snapshots.
- Chasing individual answers: arguing with one bad response instead of fixing the source that produced it wastes effort and changes nothing.
- Stopping after the findings: an audit with no owner for the remediation list is an expensive way to feel informed.
Conclusion: Look Before Someone Else Does
Every nonprofit already has an AI footprint. The only variable is whether anyone inside the organization has examined it. That asymmetry is uncomfortable, because the people most likely to encounter a flawed description of your work are precisely the people whose opinion you can least afford to have shaped by an unreviewed paragraph.
The audit itself is not difficult. It requires a thoughtful prompt set, a few disciplined hours, a spreadsheet, and the willingness to look at what comes back without flinching. What makes it valuable is repetition and follow-through. The first run tells you where you stand. The fourth run tells you whether the work you did in between actually moved anything, which is the question that separates a genuine practice from an interesting afternoon.
The remediation work is also more familiar than it first appears. Consistent facts, complete third-party profiles, clear public documentation, and independent coverage are the same things that build trust with human evaluators. Optimizing for machine legibility and optimizing for donor confidence turn out to be substantially the same project, which is fortunate for organizations with limited capacity for parallel initiatives.
Start this quarter with fifteen prompts and three assistants. Whatever you find will be more useful than the assumption you are currently operating on, and the organizations that begin looking now will be several audits ahead by the time this becomes standard practice across the sector.
Find Out What AI Says About Your Nonprofit
We help nonprofits audit how AI assistants represent them, identify the sources driving those descriptions, and build the practice of checking regularly. If you have never looked, we can help you look properly.