A webpage can be:
- indexed by Google;
- directly relevant to the query;
- detailed;
- genuinely useful;
- written specifically to answer the user’s problem;
and still never be selected as a source in an AI-generated answer.
We saw this clearly in a recent Google AI Mode experiment.
We created a resource specifically addressing a question we had already tested, confirmed that Google had indexed it, and then repeated the original query ten times.
The new page was cited 0 times out of 10.
Only when we made the user’s information need much more specific did AI Mode eventually find and cite the resource.
That suggests an important distinction:
Relevance gets you into the competition. It does not necessarily give an AI system a reason to choose you.
So what else might matter?
Our experiments increasingly point towards five connected factors:
- Candidate volume — how many plausible alternatives are competing?
- Competition quality — how strong are those alternatives?
- Query specificity — how narrowly has the user’s actual need been defined?
- Distinctiveness — what does your page contribute that the alternatives do not?
- Evidence — why should that distinctive contribution be trusted?
These are not claimed ranking factors.
They are a practical way of thinking about what we are observing in AI visibility testing.
A Relevant Page That Wasn’t Selected
The resource we created was:
How Many Times Should You Repeat an AI Search When Testing Brand Visibility?
Before it existed, we asked Google AI Mode how many times someone should repeat a query when testing whether their business was consistently recommended.
Five consecutive responses recommended 30 searches.
We then published a resource explaining why there is no universal number that automatically makes an AI visibility test statistically reliable.
The page distinguished between:
- a quick repeatability check;
- broader visibility monitoring;
- formal statistical estimation.
Google Search Console confirmed that the page had been indexed within roughly ten minutes of our indexing request.
We then repeated the original broad query ten more times.
The page was never cited.
When we later made the information need much more specific, however, AI Mode eventually found and cited the exact resource.
You can read the full experiment here:
We Built a Page for an AI Search Query — Then Tested What It Took for AI Mode to Cite It
The useful lesson for this post is simpler:
Indexed + relevant did not automatically equal selected.
Why not?
A Useful Working Model: The Candidate Environment
We cannot see Google’s internal retrieval process.
So when we talk about a candidate pool, we are not claiming that Google creates one literal fixed list of webpages and chooses from it.
It is simply shorthand for the wider competitive environment of sources that could plausibly contribute to an answer.
A detailed AI query may involve several connected information needs, and different sources may potentially contribute to different parts.
But the underlying competitive problem remains.
If your page contains useful information, how many other pages could provide something equally useful?
For a broad query such as:
How should I measure AI visibility?
there may be a huge number of plausible sources.
For a much narrower question involving a very specific customer situation, the number of genuinely good matches may be far smaller.
That difference matters.
1. Candidate Volume Matters
Suppose your resource is one of five pages that genuinely answer a particular question.
Now suppose it is one of 5,000.
Those are very different competitive environments.
This does not mean AI systems simply count webpages and follow whichever opinion has the most votes.
But more plausible sources mean more substitutes.
This is particularly important for broad topics such as:
- What is AI visibility?
- What is generative engine optimisation?
- How do I monitor my brand in ChatGPT?
- What is AI Share of Voice?
There is already a large amount of material answering those questions.
Publishing another accurate page may still be useful for your own visitors.
But from a citation perspective, you have joined a large group of possible alternatives.
2. Competition Quality Matters
Numbers alone are not enough.
Competing against 100 weak pages is different from competing against 100 strong, established resources.
AI search has changed how answers are presented, but many familiar search principles remain relevant.
A source still has to look useful and dependable compared with the alternatives.
That means questions such as these still matter:
- Is the information supported?
- Is there first-party evidence?
- Is the methodology transparent?
- Does the source demonstrate real experience or expertise?
- Are unusual claims corroborated?
- Does the page answer the complete requirement?
- Are stronger existing sources already making the same point?
A new page can therefore be highly relevant and still face difficult competition.
3. Query Specificity Changes the Competition
This may be one of the most important ideas to come out of our testing.
Compare:
Accountants in Birmingham
with:
Birmingham accountant for a 25-person manufacturer using Xero that needs payroll and monthly management accounts on a fixed fee
Both queries concern Birmingham accountants.
But the second introduces several real customer requirements.
Every additional requirement potentially removes businesses and webpages that only partially fit.
We saw something similar in our nervous-patient dentist experiment.
The customer was not merely looking for a dentist in Leeds.
They needed a practice particularly suitable for someone with severe dental anxiety, a previous bad experience, interest in sedation and a desire for clear information before treatment.
That is a much narrower requirement.
You can read that experiment here:
Would AI Mode Find the Dental Practice With the Strongest Evidence? A Five-Run Test
This is one reason we increasingly prefer testing customer-need themes rather than concentrating only on broad industry phrases.
You Don’t Need to Discover Every AI Search Query — Test Customer-Need Themes Instead
The topic matters.
But the customer’s circumstances may matter much more.
4. Distinctiveness Reduces Substitutability
Suppose 100 webpages all explain the same concept in essentially the same way.
You publish page 101.
Even if your article is excellent, what does it contain that an AI system cannot obtain from the other 100?
That is the problem with commodity content.
Commodity information is not inherently bad.
Most useful websites necessarily contain information that is already known.
But if your information is easily interchangeable, your webpage may also be interchangeable as a source.
That gives us another useful principle:
The goal is not merely relevance. It is reducing substitutability.
A less substitutable resource might contain:
- original testing;
- first-party data;
- specialist experience;
- a genuinely useful comparison;
- a new framework;
- a detailed case study;
- an important exception;
- a customer scenario that is poorly covered elsewhere;
- evidence challenging an overly simple conclusion.
None of those guarantees an AI citation.
But they give the page something more than generic topical relevance.
They give it information that is harder to replace.
5. Distinctive Claims Need Evidence
There is an obvious danger here.
If repeating what everyone else says creates commodity content, should a website deliberately disagree?
Not necessarily.
Being different is not the same as being useful.
Consider the question of how many times someone should repeat an AI visibility query.
A weak dissenting page might say:
Everyone else says 30, but I think five is enough.
Why should anyone believe it?
A more useful argument is:
A quick repeatability check and a formal statistical study answer different questions. Five runs can reveal obvious immediate fluctuation without being statistically significant. Formal statistical estimation requires a method designed around the question being measured.
That is not merely disagreement.
It explains the distinction.
The more your page departs from an established consensus, the more important the supporting evidence becomes.
That evidence might come from:
- original experiments;
- first-party experience;
- external research;
- transparent calculations;
- documented examples;
- corroborating sources.
Distinctiveness gets attention.
Evidence gives that distinctiveness weight.
Consensus Can Still Be Powerful
Our recent experiment demonstrated this quite well.
Before our new resource existed, five consecutive AI Mode responses recommended 30 searches.
After publication, the answers became more variable, but 30 remained the dominant recommendation.
Our one dissenting page did not suddenly overturn that existing information environment.
That makes intuitive sense.
If large numbers of credible sources reinforce essentially the same conclusion, that conclusion has significant informational weight.
But there is also a force pulling in the opposite direction.
If those sources all repeat substantially the same information, another copy contributes relatively little that is new.
That creates an interesting tension:
Consensus has strength through repetition. Distinctive information has value through uniqueness.
We will explore that tension separately because it deserves a post of its own.
For now, the practical lesson is simple:
Do not be different merely for the sake of being different. Be useful in a way that existing sources are not.
Sometimes the Opportunity Is a Gap Rather Than a New Fact
Distinctiveness does not always require discovering something completely new.
Sometimes the useful information already exists, but there is a problem with how it is currently presented.
One possibility is what we have called an AI Synthesis Gap.
The pieces of the answer exist, but they are scattered across several sources.
A genuinely useful page may join those pieces together around the customer’s actual problem.
There may also be another type of opportunity.
Sometimes the information exists, but the prevailing interpretation is too simplistic.
Our 30-search example fits that pattern.
The facts around sample size, repeatability and AI fluctuation already existed.
The useful contribution was identifying that several different testing objectives were being treated as though they were the same thing.
That distinction between a synthesis gap and an interpretation gap is something we will explore separately.
The important point here is that simply repeating the established explanation is not the only content opportunity.
But AI Can Synthesize Fragmented Information Too
There is an important complication.
AI systems can join information from several sources themselves.
We saw this during our own experiment.
When we asked a narrow question about conducting a quick non-statistical repeatability check, AI Mode produced methodology very similar to our new resource.
But it did so without citing us.
It assembled the answer from other webpages.
That means joining fragmented information is useful, but it may not automatically make your page essential.
A stronger resource may combine completeness with something more difficult to substitute:
- your own evidence;
- your own data;
- your own experience;
- information specific to your business;
- a distinction other sources have overlooked.
The more easily AI can reconstruct your answer from alternatives, the less it needs your particular page.
First-Party Evidence Can Be Harder to Substitute
This is where business websites may have a particular advantage.
Consider two dental-practice webpages.
One says:
We welcome nervous patients.
Another explains:
- what happens at the first appointment;
- how the practice deals with severe anxiety;
- what sedation options are available;
- whether the patient can stop treatment;
- how preparation works;
- what costs are involved;
- what happens afterwards.
The second page contains much more specific first-party evidence.
Another dentist cannot simply copy those facts unless its own service genuinely works the same way.
That makes the information less interchangeable.
The same principle applies to other industries.
An accountancy firm could simply say:
We work with manufacturers.
Or it could explain:
- the type and size of manufacturers it typically helps;
- which accounting systems it supports;
- whether payroll is included;
- how monthly management reporting works;
- how stock and margins are handled;
- whether fees are fixed;
- what onboarding actually involves.
That is much more useful to a customer with a specific requirement.
And it gives an AI system much more concrete evidence to work with.
Before You Create Another Page, Ask Five Questions
Rather than publishing another page simply because the topic is relevant to your business, consider these questions first.
1. How many plausible alternatives already exist?
Is this a very broad subject with thousands of pages covering essentially the same information?
Or a narrower problem with relatively few complete answers?
2. How strong are those alternatives?
Are you competing with generic content, or established specialist resources backed by substantial evidence?
3. Can the customer need be defined more precisely?
What circumstances narrow the real problem?
For example:
- industry;
- location;
- budget;
- urgency;
- business size;
- existing software;
- previous experience;
- technical requirements;
- unusual constraints.
4. What can you contribute that is difficult to substitute?
Could you provide:
- first-party evidence;
- original testing;
- real data;
- specialist experience;
- a useful framework;
- an important exception;
- a detailed case study?
5. What evidence supports that contribution?
If you are making a distinctive claim, explain why it should be trusted.
Show the evidence.
Explain the methodology.
Link to supporting material where appropriate.
A different opinion without evidence is weak.
A distinctive contribution supported by evidence is much more useful.
Relevance Is the Starting Point
None of this means relevance has stopped mattering.
If your page does not address the user’s information need, there is little reason to expect it to appear.
But once several sources are relevant, the competitive question changes.
It becomes:
Which source fits this particular requirement most closely?
Which source provides the strongest evidence?
Which information can already be obtained from dozens of alternatives?
Which page contributes something genuinely useful?
Which source becomes difficult to substitute?
That is a much richer question than:
What keyword should I target?
The Bigger Principle
AI search has changed how answers are assembled, but it has not removed competition between sources.
A relevant webpage is still only one possible source among many.
The stronger opportunity may therefore be to solve narrower problems, provide evidence competitors cannot easily reproduce and contribute information that becomes particularly useful when the customer’s circumstances become specific.
Relevance makes you eligible. Distinctiveness reduces substitutability. Evidence gives that distinctiveness weight.
None of that guarantees an AI citation.
Our own experiments demonstrate that clearly.
But it gives the system something commodity content often does not:
a reason to choose your page rather than another one.