Milan Bogojevic Blog

Which AI Visibility Work Is Worth Doing?

12 min read

Illustration of a funnel: a scattered cloud of coloured shapes pours in at the top, and only three shapes, a circle, a triangle and a square, come out at the bottom above the words Signal over noise.
Illustration of a funnel: a scattered cloud of coloured shapes pours in at the top, and only three shapes, a circle, a triangle and a square, come out at the bottom above the words Signal over noise.

Which AI Visibility Work Is Worth Doing? An Evidence-First Triage for Nonprofits

A communications lead at a small charity showed me her to-do list last month. It had fourteen items on it, collected from an agency audit, two webinars and a LinkedIn carousel: add FAQ schema, write an llms.txt file, break articles into short chunks, get mentioned on Reddit, refresh every page date, add author boxes, and so on. She has one colleague and about four hours a week for this.

Her question was not "what is AI visibility?" She already knew. Her question was "which of these do I do first, and how will I know if it worked?"

That is the question most AI search content skips. This piece is my answer for small teams. It assumes you have already done the groundwork in my AI visibility guide and audit, and it picks up where that leaves off: once the foundations are in place, how do you decide what to work on next, and how do you stop doing things that don't help?

The real problem is sorting, not knowing

There is no shortage of AI visibility tactics. The shortage is evidence. Only a handful have been tested properly, and I know of no test that followed a tactic long enough to show the effect holds while the AI systems keep changing underneath it.

The industry knows this. In Aleyda Solis's 2026 survey of search professionals, the share with a dedicated AI search strategy rose from about a third to well over half in a year. Yet the most common problems they reported were measuring visibility reliably and understanding what drives it, and around a third could not say what AI search contributes to revenue.

So strategies are spreading faster than proof. For a nonprofit, that matters more than for a software company with a marketing department, because every hour spent on a tactic that does nothing is an hour not spent on a programme page, a funder report or a volunteer email.

The useful shift is to stop treating your list as a list. Treat it as a set of bets, sort the bets by how much evidence stands behind them, and test them one at a time. Limor Barenholtz at Similarweb recently published a detailed version of this approach for commercial brands, which she calls an evidence-sorted roadmap. What follows is my adaptation of it for teams of one to three people.

What the strongest evidence says right now

The most useful study I have read this year is a preprint from Mahe Chen and Nick Koudas at the University of Toronto, published in October 2026. It looked at more than 150,000 AI answers and nearly half a million citations across two commercial markets over 90 days each.

What makes it valuable is the comparison it draws. Most "what gets cited" studies compare different websites, which mostly tells you that popular websites get cited. This one also compared pages within the same website, which strips out the popularity effect. Three findings matter for you.

Matching the question still works. Within the same site, pages whose title and best paragraph matched the questions people ask were cited more often. This is the one content effect that survived the stricter comparison.

Structure mostly didn't. FAQ blocks, structured data, summary boxes and page speed looked effective when comparing different sites, but not when comparing pages on the same site. In other words, the sites that use them tend to be cited more, but adding them to your own pages is not shown to change much.

Being found beats being well built. When the researchers replayed questions through ChatGPT's search, whether a page was retrieved at all explained around 32% of the difference in which pages got cited. The page properties they measured explained 2.8%. They also found that ChatGPT leans towards brands it already knows, sometimes naming them with no supporting page at all.

Read that last point as a nonprofit. It means your organisation being known, consistently described across registers, partners and press, is doing a lot of the work. That is the case I made in why your website alone won't get you cited, and this data strengthens it.

The caveats are real. It is a preprint. It covers English-language commercial questions, not charitable services. Nobody has run this study on "where can I donate winter coats in Leeds" or on queries in Croatian or Serbian, where far fewer sources exist and the balance may be different. So I treat it as the best current evidence, not as proof, and I build every plan so that it can be wrong cheaply.

Four buckets for everything on your list

Before you score anything, put every item on your list into one of four buckets based on evidence, not on how sensible it sounds.

Bucket What goes in it, for a nonprofit Why
Do Titles and answering paragraphs that match the real questions your audience asks; crawler access and key facts in plain HTML; consistent facts on the third-party sources AI already cites for your area Held up in within-site comparisons, or is a basic entry requirement for being found at all
Only if cheap Dated statistics with sources, answer-first openings, visible "last updated" dates, author boxes Plausible and sometimes correlated with citations, but unproven. Do them when they cost almost nothing
Nice to have FAQ schema, llms.txt, heading tweaks, chunking articles into small pieces No within-site lift in the best data so far. Keep them if you have them, don't schedule them as priorities
Skip Paying for mentions, fake reviews, coordinated Reddit posting, swapping links for coverage Risky for any organisation, and for a charity a reputational problem long before it is a search problem

Two notes on the middle buckets. I am not telling you to remove schema or an llms.txt file you already have. They are harmless and sometimes useful for other reasons. I am telling you not to promise your director that they will get you into ChatGPT, and not to let them push a "Do" item off this quarter's plan.

For the "Do" bucket, my paragraph-level checklist covers how to write the answering paragraph, and the about page piece covers the single page most worth fixing first.

Score what survives

Everything in "Do" and "Only if cheap" then gets two quick scores from 1 to 3.

Impact: 1 if it affects one question or one page, 2 if it affects several questions in one topic, 3 if it affects a priority topic or how your organisation is described everywhere.

Effort: 1 if one person can do it in under a week, 2 if it fits in a month, 3 if it needs several people or longer than a month.

Divide impact by effort and work from the top. There is one exception, and it is the one small teams most often get wrong: anything in "Do" that is slow, usually getting your facts corrected or added on external sites like registers, umbrella bodies or local press, starts immediately even if its score is low. Those things take months, and you don't control the timing.

One gap, one hypothesis

Here is how this looks in practice. Take a fictional organisation, Northgate Food Network, a food bank with two staff and forty volunteers.

Their monthly check of ten questions shows three gaps. They never appear for "food bank near me that delivers", although they deliver. When they do appear for volunteering questions, the answer quotes opening hours from 2024. And for "how to donate food in Northgate", a council page and a national directory are cited, but not them.

The temptation is to throw five fixes at all three. Don't. Write one hypothesis per gap:

  • Delivery gap: the delivery service is mentioned only in a PDF leaflet, so nothing on the site answers the question in a paragraph. Fix: one paragraph on the services page that answers it directly.
  • Wrong hours: the national directory still lists the old hours, and AI repeats it. Fix: correct the directory entry, not the website.
  • Donation gap: the council page and directory are the sources AI trusts for this question. Fix: get listed correctly on both. One hypothesis per gap feels slow. It is the only way that, three months from now, you will know which change did anything.

Test it like you mean it

This is the step almost every AI visibility plan leaves out, including, to be honest, the 30-day sub-query plan I published earlier, which tells you how to run a monthly panel but not how to judge it.

Three rules make the difference between a test and a feeling.

Run each question more than once. AI answers vary from day to day. In one example Similarweb published, the same brand's appearance rate was 4.4% over 28 days and 7.5% over the last two days. Anyone checking only those two days would have drawn the wrong conclusion. If your panel runs each question once a month, that is an anecdote. Run each one at least three times in the same week, and ideally build your first baseline from daily runs over a few weeks.

Keep control pages. When you change three pages, pick two or three similar pages you leave alone. If the changed pages gain citations and the unchanged ones don't, you have a signal. If both gain, something else happened, probably a model update, and you should not take credit for it.

Decide in advance what counts. At the end of each cycle, every action gets one of four decisions:

Decision When
Keep The changed pages gained citations or mentions; the control pages did not
Expand The gain appears on two or more AI systems or topics
Revise The pages gained, but the overall result for that group of questions didn't move. The hypothesis was probably wrong
Stop Nothing moved

A sensible rhythm for a small team is a 90-day cycle: two weeks to measure and plan, two blocks of about a month to do the work, and two weeks at the end to re-run the same questions in the same way. Commit to one cycle at a time. What goes into the next one depends on what the re-test tells you.

Measuring without fooling yourself

The biggest measurement mistake I see is judging AI visibility by the traffic that analytics labels as AI. Similarweb's 2026 research found that people who got a brand recommendation from ChatGPT were two and a half times more likely to visit that brand within a week than a direct competitor. But only about 9% of those visits arrived through a channel labelled as AI. The rest showed up as direct visits, ordinary search and other referrals.

For a nonprofit, that means a volunteer who heard your name from ChatGPT and then typed it into Google will look like branded search in your reports. So report three things separately and don't add them up:

  • Presence: are you named, recommended and described accurately for your tracked questions.
  • Readiness: which fixes actually shipped.
  • Impact: AI referral visits and actions (donations, sign-ups, enquiries) as a floor, plus the trend in searches for your organisation's name. Aleyda Solis suggests labelling each number as observed, proxy or modelled. I'd add that habit to every board report on this topic. It stops a guess from becoming a target.

One more pattern worth knowing: AI systems tend to cite deeper pages, two or three levels into a site, while most of the traffic they send lands on the homepage. Strengthen the specific pages that get cited, and make sure your homepage gives an AI-referred visitor an obvious next step.

Finally, set no numeric targets in your first cycle. There is no credible public benchmark for a charity's "share of AI mentions". Build your own over two or three cycles first.

My view

Most of the money currently being spent on AI visibility is being spent on the "Nice to have" bucket, because those tactics are easy to sell, easy to check off and easy to show in a screenshot. The work that the evidence actually supports is less glamorous: writing a clear paragraph that answers a real question, keeping your facts consistent on sites you don't own, and making sure the crawlers can get in.

None of this needs a tool subscription. It needs a spreadsheet, ten to thirty honest questions, and the discipline to test one thing at a time. A wrong bet then costs you one cycle, not a year, and after a year you will have something almost nobody in the sector has: your own evidence about what works for your organisation, in your language, for your audience.

Frequently asked questions

Should I remove schema markup if it doesn't help AI citations?

No. It still supports rich results in Google and costs nothing to keep. Just don't count it as AI visibility work or expect citations from it.

How many questions should a small nonprofit track?

Between ten and thirty, each tied to something a real person does: donate, volunteer, refer someone, find a service. Thirty questions that matter beat three hundred that look good on a dashboard. Generic questions like "what is food insecurity" will mostly produce zeros that look like gaps but aren't.

Is 90 days long enough to see results?

It is long enough to learn whether a hypothesis holds, which is what a cycle is for. Changes to your own pages can show up within one cycle. Corrections on external sites often take longer, so they start first and run across cycles.

Does this apply to AI answers in languages other than English?

The method does. The evidence may not. The research so far is almost entirely English-language. In smaller languages there are fewer sources for AI to choose from, so a single accurate register or directory listing may matter even more. Test it rather than assume it.

Sources

Questions people ask about this

Should I remove schema markup if it doesn't help AI citations?

No. Schema markup still supports rich results in Google and costs nothing to keep. Just don't count it as AI visibility work or expect AI citations from it, because the best current evidence shows no within-site lift from structured data.

How many questions should a small nonprofit track for AI visibility?

Between ten and thirty, each tied to something a real person does: donate, volunteer, refer someone, find a service. Thirty questions that matter beat three hundred that look good on a dashboard. Generic questions like "what is food insecurity" mostly produce zeros that look like gaps but aren't.

Is 90 days long enough to see results from AI visibility work?

It is long enough to learn whether a hypothesis holds, which is what a 90-day cycle is for. Changes to your own pages can show up within one cycle. Corrections on external sites such as registers and directories often take longer, so they should start first and run across cycles.

Does evidence-first AI visibility work apply to answers in languages other than English?

The method does, but the evidence may not, because the research so far is almost entirely English-language. In smaller languages there are fewer sources for AI to choose from, so a single accurate register or directory listing may matter even more. Test it rather than assume it.

Where this comes from

  1. Chen, M. and Koudas, N. (2026). What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO). arXiv preprint 2610.15033 arxiv.org
  2. Barenholtz, L. (2026). How to Build an AI Search Optimization Roadmap. Similarweb aisearch.similarweb.com
  3. Solis, A. (2026). State of AI Search Optimization survey. SEOFOMO hub.seofomo.co
  4. Solis, A. A 3-layer framework to measure AI presence, readiness and business impact www.aleydasolis.com
  5. Similarweb (2026). The downstream impact of AI visibility www.similarweb.com
  6. Similarweb (2026). 2026 Generative AI Landscape www.similarweb.com

Next in AI SEO & GEO