While looking more closely at NOLA Marketing after our recent discussion of the NOLA/OtterlyAI case study, I came across an article by Emily Matthews that connected particularly strongly with several things we have been observing in our own AI visibility tests.
The article is What Is a Query Fan-Out in Answer Engine Optimization (AEO)?.
Emily provides a particularly clear practical explanation of query fan-out: what it is, the different forms it can take, how related questions can be identified and how those information needs can be organised across a website. NOLA defines query fan-out as an answer engine expanding one question into related sub-questions and drawing on supporting information to build the eventual response.
NOLA approaches the subject mainly from an Answer Engine Optimization perspective: how marketers can understand these fan-outs and structure useful information around them.
Our interest at AI Visibility Testing is complementary.
What does query fan-out mean when we are trying to decide what to test and how to interpret the results?
That question leads to an interesting possibility:
Perhaps the individual prompt is not always the most useful unit of AI visibility measurement. The customer problem behind the prompt may matter more.
One Customer Question Can Contain Several Information Needs
Consider one of the queries we have previously used when testing recommendations for accountancy firms:
I run a small manufacturing company in Birmingham with 25 employees. I need an accountancy firm to handle year-end accounts, corporation tax, payroll and monthly management accounts. I use Xero and would prefer a firm offering fixed monthly fees. Which three Birmingham accountancy firms should I consider, and why?
To the person asking it, that is one question.
But a useful answer potentially needs to establish several things.
For example:
- Which firms serve Birmingham businesses?
- Which have relevant manufacturing experience?
- Which provide the required accounting and payroll services?
- Which support Xero?
- Which offer a suitable fee structure?
The final recommendation then has to bring those requirements together and identify which firms appear to be the best overall fit.
That is very different from asking:
Which accountants are good in Birmingham?
The first prompt is not necessarily more meaningful simply because it contains more words.
It contains more circumstances capable of changing what a suitable answer looks like.
Query fan-out gives us a useful framework for thinking about those circumstances.
Specificity Matters When It Changes the Problem
This fits particularly well with something we found while building AI visibility query sets around real customer situations.
In one experiment, we started with CRM software for a small business and progressively added more information about the customer.
Changing the business to a small recruitment agency materially changed the recommendations.
Adding that the agency mainly placed temporary workers changed the requirements further. Features such as worker availability, scheduling, timesheets, compliance and pay-and-bill suddenly became more relevant.
But adding that the recruitment agency had simply been trading for five years produced a much smaller change.
That led us to a working principle:
The kind of specificity may matter more than the amount of specificity.
Query fan-out provides another way to think about why.
“Five years in business” adds detail.
“Recruitment agency specialising in temporary workers” adds requirements.
The second piece of information potentially changes what the answer needs to investigate before it can identify a suitable product.
For AI visibility testing, that suggests we should not create query sets simply by making prompts progressively longer.
We should look for customer circumstances that genuinely change the problem being solved.
NOLA’s Different Types of Fan-Out Help Make This Clearer
One useful part of Emily’s article is that NOLA does not treat every expansion of a question as being the same.
The article distinguishes between semantic, temporal, entity-based, contextual and hierarchical fan-outs. For example, a contextual fan-out might change according to industry, role or use case, while an entity-based fan-out may introduce relevant companies, products or tools.
We do not need to reproduce NOLA’s full framework here — Emily’s original article is the better place to explore it.
But the distinction is useful for testing.
Adding “temporary recruitment agency” is not merely another phrase attached to a CRM query. It introduces a different operating context.
Adding “manufacturing company” to an accountancy query potentially introduces a different industry context.
Adding requirements around software, budget, location or a particular customer circumstance can introduce further dimensions.
This is why two prompts that look superficially similar can potentially represent quite different tests.
Query Fan-Out and the AI Synthesis Gap
There is also a strong connection with another idea we have explored: the AI synthesis gap.
Sometimes the information needed to solve a customer problem exists online, but no single resource brings all of it together.
One page supplies one part.
Another supplies something else.
A third fills another gap.
The AI system may then be capable of synthesising those pieces into a useful response.
We encountered this when we built a page specifically around an AI search query and tested whether Google AI Mode would cite it.
Our new page was highly relevant to the question.
But AI Mode was already capable of producing a reasonable answer by combining information available elsewhere.
Query fan-out and the synthesis gap therefore give us useful language for two related but different things.
Query fan-out concerns how a question may expand into connected information needs.
The synthesis gap concerns how completely the available resources satisfy those needs.
That distinction is important.
If a detailed question creates several information needs, there does not necessarily have to be one webpage that answers all of them before an AI system can produce a useful response.
The system may be able to assemble the answer from several resources.
That has obvious implications for visibility testing.
A resource can be highly relevant without becoming indispensable.
Query Fan-Out Also Gives Us a Useful Way to Think About Citation Variation
Another behaviour we have repeatedly encountered is changing citations.
Run the same AI search several times and you may sometimes receive:
- a similar overall conclusion;
- similar recommendations or advice;
- but different supporting sources.
Emily describes query fan-out as involving questions expanding into related information needs that can draw from different sections, pages and content formats.
That gives us a useful model through which to think about the citation variation we observe.
If one customer question requires several pieces of information, and there are several sources capable of supplying each piece, it becomes easier to understand how broadly similar answers might potentially be constructed from different source combinations.
We should be careful about going further than that.
Our tests cannot expose the internal retrieval process behind an individual Google AI Mode response, so they cannot establish that query fan-out caused a particular citation change.
But the concept does help explain why expecting one permanent set of sources behind a complex AI-generated answer may be unrealistic.
Perhaps the Prompt Is Not the Real Unit of Measurement
This is where query fan-out becomes especially interesting for AI visibility testing.
AI visibility tools naturally work with prompts.
Prompts can be saved, rerun, compared and counted.
But the commercially important thing may sit one level above the individual wording.
It may be the customer situation the prompt represents.
Consider these four prompts:
- best Birmingham accountant;
- recommended Birmingham accountant;
- top Birmingham accountant;
- good accountant Birmingham.
We have technically tracked four prompts.
But they may represent almost the same broad customer need.
Now compare them with:
- accountant for a Birmingham manufacturing company;
- accountant for a property investment business;
- accountant for a construction company operating CIS;
- accountant for a growing company needing outsourced finance director support.
Again, there are four prompts.
But the underlying customer situations are considerably more diverse.
Each potentially requires different services, expertise and evidence.
This means the raw number of prompts being monitored tells us surprisingly little unless we also understand what those prompts actually cover.
One hundred slight variations around a narrow requirement are not automatically a better visibility test than 20 carefully chosen customer situations.
NOLA’s Approach to Finding Fan-Out Questions Supports This
Another part of Emily’s article that particularly fits with this approach is how NOLA suggests finding fan-out questions.
The article includes search-led methods such as People Also Ask, AI Overview themes and related searches. But it also recommends looking at public discussions, speaking with customers or subject-matter experts and simply considering what a person would naturally ask next.
I particularly like that mixture.
It prevents query research from becoming entirely tool-led.
If we are trying to test whether AI understands when a business is relevant to a real customer, then actual customer questions are an obvious place to start.
A useful AI visibility query set might therefore begin with:
What problems bring customers to this business?
Then:
Which circumstances materially change what a suitable solution looks like?
And then:
What would someone naturally need to know next?
Only after answering those questions do we need to start converting the situations into prompts.
Three Different Types of Variation Are Worth Separating
This also helps clarify three different things we can test.
1. Repeated-prompt variation
Run exactly the same prompt several times.
Do the businesses, recommendations or sources change?
This is the issue explored in our guide to how many times you should repeat an AI search when testing brand visibility.
This tells us something about the stability of the generated response.
2. Customer-situation variation
Change something that materially affects the customer’s requirements.
For example:
CRM for a small recruitment agency
becomes:
CRM for a small recruitment agency specialising in temporary workers.
Does a different recommendation set emerge?
This tells us something about how visibility changes across different customer needs.
3. Wording variation
Express substantially the same underlying requirement in different natural ways.
We explored this when we asked Google AI Mode the same buying question five different ways.
The recommendations varied.
Because identical prompts can also vary between runs, that small experiment could not establish that wording alone caused each change.
But it highlighted an important measurement problem:
Visibility for one saved prompt is not necessarily the same thing as visibility across the wider customer need that prompt represents.
Those are three different dimensions of testing, and it is useful not to confuse them.
Start With the Customer Problem, Not the Prompt List
This changes how I think about building an AI visibility test.
Instead of beginning with:
What prompts should we track?
start with:
What customer problems matter to this business?
Then:
Which circumstances could materially change the answer?
Then:
What connected information needs sit inside those situations?
Only then do we create representative prompts.
The sequence becomes:
Customer problem
↓
Important customer circumstances
↓
Connected information needs
↓
Representative prompts
↓
Repeated testing
↓
Observed visibility
That moves the methodology away from treating prompts as though they were simply a new version of keywords.
The phrase is measurable.
But the problem behind the phrase may be what really matters.
Query Fan-Out Is Not the Same Thing as Creating More Pages
There is another point in NOLA’s article that is worth highlighting because it avoids a very easy misinterpretation.
Query fan-out and hub-and-spoke content are not the same thing.
NOLA describes query fan-out as the expansion of the question. A hub-and-spoke model is one possible way marketers can organise content around those resulting questions.
That distinction matters.
Understanding that a customer problem contains several questions does not mean creating another URL for every one of them.
NOLA explicitly says fan-out questions can sit on dedicated pages or within sections of the same page, with the clarity and scope of the answer being more important than simply increasing the number of URLs. Closely connected questions may belong together, while genuinely distinct intents may justify deeper standalone resources.
That is an aspect of the guidance I particularly agree with.
The objective should not be:
Find every conceivable AI question and create a page for it.
The starting point should remain the customer.
A prospective customer may genuinely want to know:
- whether you work with businesses like theirs;
- whether you provide the particular service they require;
- which software you support;
- which areas you cover;
- what the service costs;
- how the service works;
- whether you have experience with their circumstances.
Those are useful questions regardless of whether an AI system ever cites the page answering them.
Implementation and Measurement Answer Different Questions
This is where NOLA’s work and AI visibility testing fit together particularly well.
NOLA’s article is primarily concerned with implementation: identifying fan-outs and structuring useful information around the questions that emerge. Its broader AEO guidance also covers content structure, internal linking, FAQs, structured data and technical implementation.
Our focus is primarily measurement.
Once information exists, what actually happens when relevant customer questions are asked?
Does the business appear?
How frequently?
For which situations?
Which competitors appear instead?
Which sources are cited?
Does the result remain stable when the same prompt is repeated?
Does the answer change when the customer’s circumstances change?
Those approaches are complementary.
A sensible content rationale does not remove the need for measurement.
Equally, measurement becomes much more useful when we understand the customer questions and information needs we are actually trying to evaluate.
There is therefore a perfectly reasonable hypothesis that a business should make information customers repeatedly need clear, complete and easy to find.
Whether a particular change then affects AI visibility is something we can test rather than assume.
Commercial Relevance Still Comes First
There is one further filter.
A customer problem can fan out into many related questions.
That does not mean every one deserves equal attention.
This is where the RAMP Framework remains relevant.
Before worrying about whether a business is visible, we should consider whether the customer situation itself matters commercially.
A broad educational question could create numerous related information needs while sitting a long way from any purchasing decision.
A much narrower question may describe a problem the business is particularly well placed to solve.
Visibility across those situations should not necessarily be valued equally.
The objective is not to appear across every question that might conceivably be asked.
It is to understand visibility across the customer problems that matter.
NOLA’s Wider AEO Resource Center Is Worth Exploring
Emily’s query fan-out article goes considerably further into practical implementation than we have here, and it sits within a much broader NOLA AEO Resource Center.
The resource center currently brings together guidance covering AEO foundations, content creation, buyer-journey content, query fan-out, FAQs, internal linking, company and product information, structured data and technical implementation.
If your interest extends beyond measuring AI visibility into actually planning and implementing an AEO programme, it is a useful next place to explore.
That is also why I found Emily’s query fan-out article valuable.
NOLA is approaching the problem from the implementation side: what questions surround a topic, and how can a business organise useful answers around them?
AI Visibility Testing approaches it from the measurement side: when those customer questions are actually put to an AI system, what gets returned?
Bringing those two perspectives together raises an important possibility.
Maybe We Need to Measure Customer Problems, Not Just Prompts
A prompt is easy to count.
A customer situation is more difficult.
But the customer situation may be much closer to what actually matters.
If one customer problem can expand into several connected information needs, and different formulations of that problem can produce changing answers, then asking only:
How many prompts are we visible for?
may be too narrow.
A better question might be:
Across the customer problems that matter to us, how consistently does AI understand when our business is relevant?
That does not make individual prompt tracking unimportant.
We still need prompts in order to perform the test.
But query fan-out helps us see the difference between the thing we use to measure and the thing we are ultimately trying to understand.
And that may be an important distinction for AI visibility testing.