The research journey for a major gift starts long before a donor contacts your organisation. A funder's due diligence begins well before the formal application stage. A journalist investigating your sector types your name into an AI tool between reading two other sources.

What those tools return shapes the first impression. And for a lot of charities, that impression is vague, partly wrong, or built from information that has long since gone out of date.

This is not primarily an SEO problem. It is a problem with the state of your published information, and left unaddressed it becomes a governance problem.

A warning about everything written on this subject

Most writing about what AI tools get wrong is out of date by the time you read it. Model capability moves quickly. A confident claim that "AI cannot do X" is the single most likely thing in any article on this topic to be false by the time it reaches you, including this one.

So I have tried to keep the argument at the level that does not expire. Three questions survive every model release:

  • Is what has been published in the charity's name checkable against a source?
  • Does the tool have access to the organisation's actual facts, or is it working around a gap?
  • When something is wrong, who is answerable to the Board and the regulator?

Everything below sits on those. Where I make a claim about how a specific system behaves today, it is sourced and dated, and you should assume it has a shelf life.

Why this is now a stakeholder research question

The sector is already inside this. The Charity Digital Skills Report 2026 found 79% of UK charities using AI tools in some way, and 51% with no AI policy at all. The people researching you are using the same tools, under the same absence of policy.

On the search side, Google's own documentation describes its generative features as using retrieval-augmented generation, which it also calls grounding, drawing on its core Search ranking systems to pull relevant pages from the Search index. Its guidance for site owners, last updated 10 July 2026, states that to be eligible to appear as a supporting link a page must be indexed and eligible for a snippet.

That is the practically useful fact, and it is more durable than any statement about what a model knows. These systems increasingly answer from retrieved pages rather than from memory. Which means the question is no longer "what did the model learn about us", it is "what of ours is findable, and is it right".

What tends to be wrong

I run a version of this check as part of the Blueprint Audit: ask several tools about a client's organisation by name, record what comes back, and compare it against what is true. The patterns are consistent enough to name.

Geographic scope. An organisation operating locally in London is described as a "national charity". An international NGO with programmes across several regions is placed "primarily in East Africa". This is usually traceable to the organisation itself, which describes its reach differently in its annual report, its funding applications and its website footer.

Mission conflation. Two charities working on youth employment in the same city, two foundations funding arts education, and the description that comes back fits neither precisely but is plausible for both. If a funder is assessing whether you meet their criteria, a conflated mission is enough to screen you out of a process you were eligible for.

Leadership and governance that has moved on. If your chief executive left some time ago, the question is not whether a tool "remembers" the old name. It is whether your own site still shows the old name, or shows no name at all, or shows a Board list with no date on it. All three produce a wrong answer, and in all three the error originates with you. This is one of the most common findings in a website audit and it has nothing to do with AI: see Website Governance During Leadership Transitions.

Scale. "A small charity" against "a significant regional organisation" changes how a reader contextualises everything else. Income figures are particularly unstable, because the most prominent number about you across the indexed web is often a grant amount from a press release rather than a total income figure from your accounts.

Programme descriptions from a previous strategy. Organisations that have changed what they do, which is most established charities, find themselves described in terms of the model they moved away from. A housing charity that has shifted from direct provision to systems advocacy is still summarised as a provider. For a funder with specific priorities, that mismatch decides fit before any conversation happens.

The one behaviour worth planning around, because it is unlikely to change soon, is confidence. The Tow Center for Digital Journalism at Columbia tested eight AI search tools across 1,600 queries and published the results in March 2025 as AI Search Has a Citation Problem. The tools returned incorrect answers to more than 60% of the queries, and the researchers' sharper finding was that the tools were bad at declining: they offered speculative answers rather than saying they did not know, and more than half of responses from two of the tools cited fabricated or broken URLs. That study is about news attribution rather than charity profiles, and the numbers will have moved since. The structural point is the one to carry: a gap in what is available about you is not returned to the reader as a gap.

Why this belongs to governance and not to communications

The instinct is to file it under reputation management, something the comms team watches. That understates the exposure, because of who is doing the reading.

Grant officers and foundation programme staff shortlist from partial information before they engage. A summary that puts you in the wrong geography or presents an outdated governance structure removes you from a conversation without anyone telling you.

Institutional donors and major gift prospects research informally before requesting a meeting. If what they find contradicts what your development team has told them, the inconsistency is itself the signal they act on.

Journalists and policy researchers use these tools to get oriented quickly. An inaccurate description can shape an angle, or a decision about whether to include your voice.

Regulators and statutory funders will notice where a public description of your governance does not match your Charity Commission register entry. However that arose, it reads as a failure of institutional communication.

None of these require a dramatic error. A geographic scope slightly off, an out-of-date income range, a programme name that no longer exists: enough to introduce doubt at a moment when you are not in the room to correct it.

The governance point is this. Your organisation is being described to consequential audiences by systems you do not control, and the quality of that description is set almost entirely by the quality and consistency of what you have published. That is not a comms failure. It is an infrastructure failure with governance consequences, and it is the Board's to ask about.

What the tools are actually working from

It is tempting to answer this with a list of what AI systems supposedly reward. Most such lists are invented. Google's published guidance is unusually blunt about it, and worth reading against anything an agency tells you:

You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search.

And, on schema:

Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add.

That guidance also warns against writing content specifically for machines, against chunking pages for them, and against recycling what is already on the internet. The implication is not that structure is pointless. It is that structure is not a shortcut, and the thing being assessed is whether your page says something true, specific and its own.

So the honest version of the infrastructure problem is narrower than the usual advice, and more actionable:

  • Facts stated only in a PDF are facts you have chosen not to state on a page. Whether or not a given system parses your annual report well this year, a page that states your income, your geography and your governance in plain HTML is unambiguous, citable and easy to keep current. The PDF can stay. The fact should also live somewhere with a URL.
  • The same fact expressed three different ways is three facts. Website, Charity Commission register entry, LinkedIn, sector directories. Where these disagree, something has to reconcile them, and you do not get to choose what.
  • Undated information cannot be judged current. A Board list with no "as at" date is not obviously stale, which is exactly the problem.
  • Schema is disambiguation, not visibility. Organisation structured data, per Google's documentation, "can help Google better understand your organization's administrative details and disambiguate your organization in search results". For a charity with a name close to another charity's, that is a real benefit and it is the benefit to claim. The technical side is in Schema Markup for Nonprofit Websites.

The gap this all points at is entity definition: a specific, checkable statement of what your organisation is, who it serves, where it operates, at what scale, under what governance, with what track record. That information usually exists somewhere in your organisation. It is rarely in one place, in one wording, on a page with a date on it.

Being cited is the goal, not being hidden

There is a reflex among charities to treat this as something to opt out of. That is the wrong instinct, and the two decisions are separable.

Reserving your content against being used for model training is a legitimate position, and it is a different setting from allowing an answer engine to retrieve your page in order to cite it. Blocking retrieval does not remove you from these answers. It removes your website as the source and leaves whatever else exists about you, which is the position you were trying to escape.

The goal is to be the thing that gets cited, with a link back, accurately. That is what Answer Engine Optimisation for Nonprofits sets out in practice.

What a trustee should require

This is where the reader's real question usually sits, because the pressure is not only "what do these tools say about us", it is "we are being told to use AI and nobody has said what is safe".

One requirement covers most of it. Nothing is published in the charity's name until a named person has checked it against a source and accepts responsibility for it. Not "reviewed by the team". A name, against a version.

That is not a new obligation invented for AI. The Charity Commission's The essential trustee (CC3) already requires trustees to act with reasonable care and skill and to ensure the charity is accountable. A drafting tool changes how quickly text arrives. It does not move where the duty sits.

Two constraints sit underneath it:

Beneficiary and supporter data does not go into a third-party tool casually. Pasting case notes, service user details or a supporter list into an external system is a processing operation, and UK GDPR applies to it in the ordinary way. The ICO's guidance on AI and data protection is the reference point, and it makes clear that lawful basis, data minimisation, purpose limitation and accountability are not suspended because a system is probabilistic. The practical rule for staff is short: no personal data goes into a tool that has not been assessed, and the assessment is a job for someone, not a shrug.

Every figure and quotation about your own work must be traceable to a document you hold. This is the constraint that actually catches the failure mode. Fluent, plausible, unsourced numbers about your own impact are worse than no numbers, because they are the ones a funder will check.

Notice what is not on this list. There is no requirement to detect whether text was machine-written, no percentage of AI involvement to declare, no ban. Those are all attempts to police the tool. The requirements above police the output, which is the part that has the charity's name on it.

What to actually do

Four things, in order, and none of them is large:

  1. Sample what is currently said about you. Ask several tools, record the answers with dates, and note which of your pages or which third-party listing each answer appears to lean on. Sample rather than test, and look for the pattern.
  2. Fix the source, not the summary. Every recurring error points at a specific page that is missing, stale, undated, or contradicted by a directory listing you forgot you had.
  3. Put a date and an owner on the pages that decay. Governance, leadership, programmes, impact, income. This is worth doing whatever happens to AI, and the Board can ask for it directly. What Your Board Should Actually Ask About Your Website has the rest of that list.
  4. Write the one-line publishing rule down and tell people it exists. Half the sector has no AI policy. The named-person rule is a policy you can adopt this week and defend to a regulator, which is more than most longer documents manage.

Where website governance and institutional risk meet

Your website has always been your primary institutional communication channel. What has changed is that its content now feeds systems that describe your organisation to consequential audiences without your involvement and without your knowledge.

That shifts the calculus. Content that is vague, inconsistent, undated or buried in a download now carries a risk it did not carry before, not because a reader would be misled by it directly, but because it is the raw material for a description your stakeholders meet without context.

Getting this right is not a large investment. It is a systematic one: sample what is being said, review consistency, publish the facts where they can be read, and hold a cadence for keeping the decaying pages current. Those are the kinds of tasks a Blueprint Audit surfaces and sequences, alongside the structural work that makes a charity website fit for institutional purpose.

Further reading:

Talk to me about a Blueprint Audit