ORIGINAL RESEARCH REPORT
AN EXPLORATORY INVESTIGATION BY LLUMO
SEPTEMBER 2026
THE SIGNAL
Persistent candidates. Variable ranking.
AI systems can repeatedly surface the same businesses while changing the order, explanation, and final shortlist.
EXECUTIVE SUMMARY
AI can compress a market into a shortlist.
The question is no longer only whether a customer can find a business. It is whether an AI system will choose it when a customer asks for a recommendation.
THE QUESTION
“What are the 10 best restaurants in London?”
Tested across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode — then repeated within the same systems.
SECTION 01
From the Director’s Desk
Digital visibility has acquired a new problem
For years, businesses have understood visibility in relatively familiar terms.
A business wanted to appear in search results. It wanted to rank for important keywords. It wanted reviews, traffic, backlinks, mentions and strong positions in search.
The objective was relatively simple:
Be found.
But the search environment is changing.
A consumer can now ask an AI system a question such as:
“What are the 10 best restaurants in London?”
The consumer does not receive a page of links to investigate.
They receive an answer.
Ten businesses may be selected. Each may be described. Each may be given a reason for consideration. Some may be ranked above others.
And the businesses that do not appear may never enter the consumer’s consideration at all.
That changes the nature of digital visibility.
The question is no longer only:
Can a customer find my business?
It is increasingly:
Will an AI system choose my business when a customer asks for a recommendation?
And then another question follows:
Why?
That is where this investigation began.
At first, we expected to find differences between AI platforms. We asked five systems the same question and compared their answers.
They disagreed.
Then we asked a more uncomfortable question:
What happens if we ask the same AI the same question again?
The answer became more interesting.
The same system could return a different list.
That led us further.
If an AI system can give us a ranked list of “the best” businesses, where does that ranking come from?
We asked the systems themselves.
They named sources: Michelin, Time Out, SquareMeal, The Good Food Guide, national publications, reviews, business websites and other information sources.
So we went to those sources.
What we found was not a hidden master list of the ten businesses the AI had just recommended.
We found hundreds of restaurants, separate rankings, editorial selections, awards, reviews, guides and different methodologies.
The source could explain some of the information available to the AI.
It could not, by itself, explain the exact final answer.
That gap is the subject of this report.
The Invisible Shortlist is not necessarily a hidden list.
It is the unseen process between:
the question a consumer asks
and
the answer an AI system decides to give.
SECTION 02
Executive Summary
AI is becoming a new layer of business discovery.
Consumers increasingly use AI systems to discover restaurants, hotels, services, products and companies. Recent external research suggests that this behaviour is moving quickly: BrightLocal’s 2026 research reported that 45% of consumers had used AI tools for local-business recommendations in the previous year, compared with 6% the year before.
But recommendation is different from search.
Traditional search can expose a large market.
AI can compress that market into a shortlist.
That means the AI system is not simply helping the consumer find information. It is helping determine which information reaches the consumer first.
To investigate this, Llumo began with a deliberately simple consumer question:
“What are the 10 best restaurants in London?”
The same question was tested across ChatGPT, Claude, Gemini, Perplexity and Google AI Mode.
The resulting lists differed.
That finding alone was not surprising enough to be the conclusion. There is no objectively measurable universal answer to “the best 10 restaurants in London.”
The investigation therefore moved deeper.
We tested repeated conversations within the same AI platform.
We tested Google Search across different accounts.
We examined recurring businesses.
We asked AI systems where their recommendations came from and what factors they said influenced selection and ranking.
Then we investigated the sources they named.
This produced a more complicated picture.
Certain businesses repeatedly appeared within individual AI systems, suggesting that recommendation outputs were not simply an entirely new list every time. At the same time, rankings could move substantially between runs.
The most useful description of this pattern is therefore:
Persistent candidates, variable ranking.
The source investigation produced another important observation.
AI systems could identify authoritative sources that informed their recommendations, but those sources did not provide a simple one-to-one explanation for the final AI rankings.
Michelin, for example, provides a large curated selection of London restaurants rather than a single “Top 10 London” ranking. Time Out publishes its own ranked lists. SquareMeal publishes another. The National Restaurant Awards has another. The Good Food Guide has its own methodology and ranking.
Yet an AI may combine information associated with several such sources and produce a new ranked answer.
This does not prove exactly how any model internally works.
It does reveal something important for businesses:
Being present in an information source and being recommended by an AI are not necessarily the same thing.
A source may contribute to the information environment surrounding a business without determining whether that business appears, where it appears, or how it is described.
This leaves an important unresolved layer:
What causes a business to move from being information available on the web to being a business selected and ranked in an AI answer?
That is the Ghost in the Prompt.
SECTION 03
01 — The Question That Started Everything
The investigation began with a question that appeared almost too simple to be useful.
“What are the 10 best restaurants in London?”
There was no request for:
Michelin-starred restaurants
affordable restaurants
romantic restaurants
restaurants for tourists
restaurants near a particular location
restaurants with the best reviews
restaurants with the best food
restaurants ranked by critics
There was simply one broad request.
What are the best 10?
That matters because the consumer did not define “best.”
The AI did.
The user supplied the question.
The AI supplied the interpretation, candidate selection, ranking and explanation.
That distinction became central to the research.
We were not trying to determine what the user meant by “best.”
We were asking:
When the user leaves “best” undefined, what does the AI actually do?
And more importantly:
What evidence can we observe for how it arrives at that answer?
SECTION 04
02 — Experiment One: Five AIs, One Question
The first experiment used the same consumer-style question across five major AI/search systems:
ChatGPT
Claude
Gemini
Perplexity
Google AI Mode
The complete outputs were recorded and compared.
The results were different.
Among the businesses appearing in multiple systems were:
Restaurant
ChatGPT
Claude
Gemini
Perplexity
Google AI
St. John
7
—
4
7
6
CORE by Clare Smyth
—
2
—
2
5
Fallow
9
—
3
—
2
OMA
—
7
—
—
3
Gymkhana
—
—
2
—
9
Ikoyi
—
5
7
5
—
Camille
1
1
—
1
—
Legado
—
—
—
9
10
Bouchon Racine
—
—
—
—
1
A dash means that the restaurant did not appear in that system’s top ten.
This was an exploratory comparison, not a ranking of the restaurants themselves.
The pattern was immediately visible.
Some businesses appeared across several systems.
Others appeared in only one.
And even where the same restaurant appeared in several systems, its position could be different.
St. John, for example, appeared in four of the five outputs but did not occupy the same position.
CORE by Clare Smyth appeared in three.
Fallow appeared in three.
OMA and Gymkhana appeared in two.
This established the first layer of the investigation:
Different AI systems can produce different versions of the same market.
But this was only the beginning.
SECTION 05
03 — The Question Changed
At this point, the obvious conclusion would have been:
“Different AIs give different answers.”
But that was not enough.
A journalist raised a better question:
What happens if you ask the same AI the same question repeatedly?
That changed the direction of the research.
If five different systems disagree, perhaps each system simply has its own methodology.
But what if one system does not consistently reproduce its own answer?
That is a different question.
So we opened fresh conversations and repeated:
“What are the best restaurants in London?”
The question remained substantially unchanged.
The experiment was designed to hold the prompt constant and observe what changed.
SECTION 06
04 — Experiment Two: Does the Same AI Agree With Itself?
ChatGPT
Ten valid repeated ChatGPT runs produced 31 distinct restaurants across 100 recommendation slots.
Some businesses appeared repeatedly.
Others appeared only once.
The most persistent were:
The Ledbury — 9/10 runs
Fallow — 9/10
Dinner by Heston Blumenthal — 7/10
St. John — 7/10
The Barbary — 6/10
CORE by Clare Smyth — 5/10
Kitchen Table — 5/10
Restaurant Story — 5/10
Pavyllon London — 5/10
Dishoom/Dishoom Covent Garden — 5/10
But persistence did not mean ranking stability.
The Ledbury appeared nine times and was almost always ranked at or near the top.
Fallow also appeared nine times, but its position moved considerably, from #2 in one run to #10 in another.
Dinner by Heston Blumenthal appeared seven times, with positions ranging from #2 to #5.
St. John appeared seven times, ranging from #3 to #8.
The observation was therefore more nuanced than “ChatGPT is random.”
The evidence suggests:
Some businesses remain persistent candidates while their relative positions can change.
This distinction matters.
A business can be consistently present without being consistently ranked.
Presence and position are different measurements.
SECTION 07
05 — Claude: A Different Pattern
Claude produced an even more concentrated pattern in the valid repeated runs.
Across four valid ranked runs, the following businesses repeatedly appeared:
Camille — 4/4
CORE by Clare Smyth — 4/4
Tiella — 4/4
Restaurant Story — 4/4
Mountain — 4/4
Ikoyi — 4/4
OMA — 4/4
Impala — 4/4
Several remained in almost exactly the same positions across those runs.
Camille repeatedly appeared at #1.
CORE repeatedly appeared at #2.
Tiella repeatedly appeared at #3.
Restaurant Story repeatedly appeared at #4.
Mountain remained around #5–6.
Ikoyi around #5–6.
OMA remained #7.
Impala remained #8.
This was a striking contrast with the wider movement observed in ChatGPT.
But it also provided an important methodological lesson:
Different systems may display different levels of internal stability.
Two Claude runs were anomalous and did not return a complete ranked top ten. Those runs were preserved as part of the research record but excluded from quantitative ranking calculations, in accordance with the research protocol.
The anomalies themselves are important because real-world AI research does not always produce clean laboratory-style outputs.
SECTION 08
06 — Gemini: Stability at the Top
Gemini produced five valid runs.
Three businesses appeared in all five:
CORE by Clare Smyth — 5/5
Positions:
1, 1, 2, 1, 1
Restaurant Gordon Ramsay — 5/5
Positions:
3, 3, 1, 2, 2
The Ledbury — 5/5
Positions:
2, 2, 3, 5, 3
Again, we see a recurring candidate pool.
But we also see movement.
The top three businesses were not always in the same order.
This is important because it demonstrates that:
High presence does not automatically mean fixed rank.
A system may repeatedly identify a business as one of its strongest candidates while changing the relative ordering of those candidates.
SECTION 09
07 — Perplexity: Recurring Candidates, Moving Positions
Perplexity produced another interesting pattern.
The Ledbury appeared in all five valid runs:
6, 4, 4, 2, 3
Bouchon Racine appeared in four of five and occupied #1 in each of the first four runs.
Other recurring businesses included:
Mountain
Dorian
Trinity
Legado
AngloThai
Impala
The Ritz Restaurant
Ikoyi
Gymkhana
CORE
Hélène Darroze at The Connaught
Again, there was a recurring universe of candidates.
But there was no single permanent ordering.
The result reinforces a pattern seen across the experiments:
The system may repeatedly encounter or favour certain candidates without reproducing a fixed ranking every time.
SECTION 10
08 — The Google Account Experiment
The investigation then moved outside conventional chatbot conversations.
Google Search was tested across different Google accounts on the same computer.
The purpose was not to prove that account identity alone caused the differences.
The purpose was to ask a narrower question:
Can recommendation results differ when the environment is kept broadly similar but the Google account changes?
They did.
One account produced a list including Camille, OMA, The Devonshire, Impala, Mountain, Gymkhana, St. John, Bouchon Racine, Fallow and Circolo Popolare.
Another produced:
Camille, Tiella, St. John, Belly Bistro, Normah’s, Agora, Barrafina, Josephine, Kiln and Dinner by Heston Blumenthal.
Another account’s AI Mode produced:
CORE by Clare Smyth, Restaurant Gordon Ramsay, Sketch, Gymkhana, The Ritz Restaurant, Dinner by Heston Blumenthal, Evelyn’s Table, Fallow, Ekstedt at The Yard and OMA.
Other accounts produced still different combinations.
The correct conclusion is not:
“Changing accounts causes Google to change its rankings.”
There were too many uncontrolled factors to establish that causation.
The defensible finding is:
Different account contexts produced different observed recommendation outputs under broadly similar search conditions.
That was enough to justify another question:
How much of what a consumer sees is determined by the information environment surrounding the query, rather than by a universal fixed list?
SECTION 11
09 — We Asked the AI Where Its Recommendations Come From
This became one of the most important stages of the investigation.
Instead of guessing, we asked the AI systems directly.
The question was:
“Where do your recommendations come from? What sources do you consider, what factors influence which businesses you recommend, and how do you decide which businesses to include or rank?”
The answers were revealing.
ChatGPT identified sources including:
Michelin Guide
Time Out London
World’s 50 Best Restaurants
The Good Food Guide
SquareMeal
national newspapers and food publications
restaurant websites
community reviews and review platforms
It described factors such as reputation, critical recognition, experience, distinctiveness, current relevance and fit with the question.
Gemini similarly described a mixture of authoritative guides, review platforms, business metadata and contextual relevance. It specifically stated that frequency of appearance across critical top lists could influence selection and ranking.
Claude gave an even more concrete explanation for the restaurant test.
It said it had searched for “best restaurants in London 2026” and synthesized information from publications including:
Time Out
Balance Journal
DesignMyNight
TheFork
It also said that it selected restaurants that appeared across multiple sources or were described with specific details such as chefs, dishes and awards.
That statement was particularly interesting.
Because now we had something testable.
If repeated appearance across sources helps a restaurant become a candidate, then perhaps the source ecosystem leaves a measurable footprint.
So we went looking.
SECTION 12
10 — Following the Source Trail
This was the point where the investigation became much more interesting.
We did not simply accept what the AI told us.
We went to the sources.
And what we found created a new problem.
Michelin
Michelin’s London selection contains hundreds of restaurants.
It is a curated universe of restaurants with Michelin distinctions, cuisine information, price levels, locations and other structured signals.
But it does not simply say:
Here are the 10 best restaurants in London.
The AI therefore cannot simply be reproducing a Michelin Top 10.
Time Out
Time Out has its own ranked list of London’s best places to eat.
Its September 2026 list contains 50 restaurants and is independently reviewed and updated.
It therefore provides something much closer to a ranking.
But again:
The AI’s list is not simply the Time Out list.
SquareMeal
SquareMeal publishes its own London Top 100.
Its methodology combines in-house critic opinion with public votes and considers cooking, hospitality, diner experience, consistency and other factors.
Yet again:
The AI does not simply reproduce SquareMeal’s Top 100.
National Restaurant Awards
The National Restaurant Awards provides another ranked ecosystem.
Bouchon Racine was #1 in 2026.
The Ledbury was #4.
Mountain was #7.
Legado was #10.
Camille was #20.
Ikoyi was #24.
Kitchen Table was #48.
And many other restaurants followed.
But again:
The AI did not simply reproduce this ranking.
SECTION 13
11 — The Source Does Not Equal the Answer
This became one of the most important findings of the investigation.
Consider Fallow.
SquareMeal ranked Fallow at #43.
Yet Fallow appeared repeatedly in our AI experiments and reached as high as #2 in one ChatGPT run.
St. John appeared at #61 in SquareMeal’s Top 100.
Yet it repeatedly appeared in AI outputs and reached #3 in one ChatGPT run.
Camille provides another striking example.
Camille was:
#1 on Time Out’s current top-50 list
#20 in the National Restaurant Awards
#53 in SquareMeal
Claude repeatedly put Camille at #1.
ChatGPT included it only three times in ten runs.
These examples show something critical:
A source’s ranking does not automatically become an AI’s ranking.
The source may be part of the information environment.
But something else happens between the source and the answer.
SECTION 14
12 — The Candidate Pool
At this point, the investigation began to suggest a model.
Not a proven description of how an AI internally works.
A research model.
We observed something resembling:
01
Information across the web
02
Source presence and information availability
03
Potential candidate businesses
04
AI selection
05
AI ranking
06
AI description
07
Consumer decision
The first part of this chain is observable.
We can see the websites.
We can see editorial coverage.
We can see awards.
We can see reviews.
We can see business profiles.
We can see what sources AI cites or names.
But the middle remains partly opaque.
Why does Restaurant A become a candidate?
Why does Restaurant B not?
Why does Restaurant A become #1 today and #6 tomorrow?
Why does one AI repeatedly recommend a restaurant another AI almost never mentions?
Why can a business appear in the same source ecosystem as its competitors but receive very different AI visibility?
These are not questions the source lists alone answer.
SECTION 15
13 — The Invisible Shortlist
This is where the title finally becomes clear.
The Invisible Shortlist is not the consumer’s hidden preference.
The consumer asked a straightforward question.
“What are the 10 best restaurants in London?”
The consumer did not define “best.”
The AI did.
It decided what information was relevant.
It decided which businesses were worth presenting.
It decided how to describe them.
It decided their order.
And when we repeated the experiment, the answer sometimes changed.
Yet the AI can give the consumer a polished explanation of the result.
That creates an unusual information problem.
We can observe:
Prompt → Answer
We can sometimes observe:
Answer → Sources
But we cannot fully observe:
Sources → selection → weighting → ranking → answer
That missing layer is the Ghost.
It is the part of the recommendation process that sits between the user’s simple question and the AI’s apparently authoritative answer.
SECTION 16
14 — The Invisible Editor
AI increasingly behaves like an invisible editor.
Not because it necessarily intends to suppress businesses.
Not because it necessarily has a secret list.
And not because its answers are necessarily random.
Rather, because it filters a large information environment into a small number of choices.
Traditional search might show a consumer dozens of pages.
An AI recommendation might give the consumer ten businesses.
The consumer therefore encounters a heavily compressed version of the market.
That creates three different visibility states:
1. Present
The AI knows enough about the business to mention it.
2. Recommended
The AI actively selects it for the consumer’s question.
3. Well represented
The AI accurately explains what the business is, what it is known for and why it may be relevant.
These are not the same thing.
A business can be known without being recommended.
It can be recommended without being accurately represented.
And it can be accurately represented in one system while being poorly represented in another.
This is why AI visibility cannot be reduced to a single “mentioned/not mentioned” metric.
SECTION 17
15 — What We Think May Be Happening
The evidence allows us to propose several hypotheses.
It does not allow us to present them as proven facts.
Hypothesis 1: Source presence may influence candidate availability
Businesses that appear across authoritative information ecosystems may have more opportunities to enter an AI’s consideration set.
But presence alone does not guarantee recommendation.
Hypothesis 2: Repeated source presence may strengthen visibility
A business appearing across multiple independent sources may create a richer information footprint.
That may make it easier for AI systems to encounter, identify or describe.
Again, this requires further testing.
Hypothesis 3: Source ranking does not necessarily determine AI ranking
A restaurant being #10 on one source does not mean AI will rank it #10.
The AI may combine information from several sources.
Hypothesis 4: AI platforms may use different source ecosystems
Our experiments showed platform-specific recommendation clusters.
The systems also described different source mixes when asked how they generated recommendations.
Hypothesis 5: AI ranking involves an additional layer beyond source presence
Something happens between:
“This business exists in the information environment”
and
“This business belongs at #3.”
The nature and weighting of that layer remain an open research question.
SECTION 18
16 — The Most Important Distinction
The investigation has led us to separate five stages that are often treated as one.
1. Availability
Is information about the business available to the system?
2. Candidate formation
Does the business enter the set of businesses potentially relevant to the question?
3. Selection
Does the AI choose the business for the answer?
4. Ranking
Where does the business appear relative to other selected businesses?
5. Representation
What does the AI tell the consumer about that business?
These stages may be connected.
But they are not identical.
That distinction changes how businesses should think about AI visibility.
SECTION 19
17 — What This Means for Businesses
The lesson is not:
“Get listed on Michelin and AI will recommend you.”
We do not have evidence to support that claim.
It is also not:
“Get mentioned by five publications and you will become #1.”
Again, we cannot support that.
The more defensible conclusion is:
Businesses seeking AI visibility need to think about their entire information footprint, not simply their own website.
That includes the places where information about the business exists:
editorial coverage
respected industry guides
reviews
awards
business directories
third-party publications
authoritative databases
relevant specialist sources
first-party business information
But there is a second requirement.
The information must also make sense.
A business needs a clear identity.
Its name, location, category, services, products, expertise, reputation and distinguishing characteristics need to be represented consistently enough that an AI system has something coherent to work with.
This moves AI visibility beyond traditional SEO.
The question becomes:
What does the internet collectively say about this business?
And then:
What does AI make of that information?
SECTION 20
18 — From Ranking to Representation
Traditional SEO has given businesses familiar measurements:
rankings
impressions
clicks
traffic
conversions
backlinks
AI search introduces additional questions.
Mention
Does the AI know the business exists?
Recommendation
Does it recommend the business?
Position
Where does it place the business?
Reason
Why does it say the business is worth considering?
Representation
How does it describe the business?
Accuracy
Are those claims correct and current?
Consistency
Would another AI system tell the same story?
These dimensions create a broader picture of AI visibility.
The next generation of search visibility will therefore not be measured only by whether a business ranks.
It will also need to consider whether AI systems can:
find it, select it and represent it accurately.
SECTION 21
19 — What Surprised Us
The most important discovery in this investigation was not the first one.
At the beginning, we thought the story might be:
“Five AIs, five different lists.”
Then we discovered:
“The same AI can produce different lists.”
Then we discovered:
“Different Google account contexts can produce different observed outputs.”
Then we asked:
“Where do the recommendations come from?”
The AI gave us sources.
So we followed the sources.
And that created the next question:
“If these are the sources, how do they become this particular answer?”
The sources contain hundreds of candidates.
Some rank restaurants.
Some do not.
Some use critics.
Some use public votes.
Some use awards.
Some use reviews.
Some emphasise recency.
Some emphasise reputation.
Some emphasise geography.
And the AI can combine information from several of them.
The deeper we went, the less the answer looked like a simple list.
It looked like a system of information being filtered.
And that is why the investigation changed.
We did not begin with the final question.
We discovered the question while trying to answer the first one.
SECTION 22
20 — What We Cannot Claim
A credible investigation needs boundaries.
This research does not prove:
that AI recommendations are random;
that AI deliberately hides businesses;
that a particular publication causes a business to rank;
that Michelin is directly responsible for an AI recommendation;
that source frequency is definitely a ranking factor;
that all AI platforms use the same retrieval process;
that one platform is objectively better than another;
that account changes alone caused Google’s observed differences;
that a business can guarantee AI visibility by appearing in a particular publication.
The research also does not provide access to the private internal mechanisms of the AI systems tested.
When an AI tells us which sources it considers, that is evidence of what the system says about its methodology, not proof of every internal retrieval or ranking mechanism.
That distinction is essential.
SECTION 23
21 — Limitations
This is an exploratory investigation.
The restaurant sample was deliberately narrow.
The repeated-platform experiments were not conducted under a laboratory environment capable of controlling every possible variable.
AI systems can change models, retrieval systems, interfaces and source access.
Google account environments can contain personalization and contextual differences that are difficult to eliminate completely.
Some AI responses were incomplete or anomalous.
Those responses were preserved but excluded from quantitative calculations where they did not provide a complete ranked list.
The restaurant category also has unusual characteristics: it is highly editorialized, geographically concentrated and heavily reviewed.
Therefore, the findings should not automatically be generalized to every business category.
The purpose of this first investigation is not to claim that the mystery has been solved.
It is to demonstrate that the mystery is measurable.
SECTION 24
22 — The Business Question
The biggest change brought by AI search may not be that consumers have another way to find information.
It may be that information is being filtered before the consumer sees it.
A customer asking:
“What are the best restaurants in London?”
does not see the entire restaurant market.
They see the market after an AI system has made a series of decisions.
That means businesses are entering an environment where their first interaction with a potential customer may not be their website.
It may be an AI-generated description of them.
The fundamental questions therefore become:
Can AI find my business?
Will AI recommend my business?
Why does it recommend my business?
Where does the information supporting that recommendation come from?
Why does another AI give a different answer?
And when AI finds my business, what version of my business does the customer actually meet?
SECTION 25
23 — The Invisible Shortlist
A simple consumer question can produce a surprisingly complex answer.
01
The consumer asks.
02
The AI interprets.
03
Information is retrieved or considered.
04
Candidates become available.
05
Businesses are selected.
06
Positions are assigned.
07
Descriptions are generated.
08
The consumer receives ten names.
The consumer may never see the hundreds of businesses that were not selected.
And even the ten that appear may not appear in the same order the next time the question is asked.
We can see the question.
We can see the answer.
We can investigate the information surrounding the answer.
But there remains a layer we cannot fully see:
the process by which the AI turns an enormous information environment into a small, ranked recommendation.
That is the Ghost in the Prompt.
SECTION 26
Conclusion
AI search is changing the meaning of visibility.
The old question was:
“Where does my business rank?”
The new question may be:
“When a customer asks an AI to choose, does my business enter the answer—and what happens if it does?”
Our investigation suggests that AI recommendation visibility is more complicated than simply being present on the web.
A business may have strong authority and still be absent from a particular AI answer.
A business may rank modestly in one authoritative source and appear near the top of an AI recommendation.
A business may repeatedly appear within one platform while being absent from another.
And the same platform may change the ordering of its recommendations even when the consumer asks essentially the same question.
The evidence does not yet reveal a single formula.
That is precisely why the question matters.
The future of digital visibility may depend less on being present in one place and more on how consistently a business is discovered, interpreted, selected and represented across the information ecosystems that AI systems use to construct answers.
For businesses, that means the work is no longer finished when the website is optimised.
The next challenge is understanding the information environment around the business.
Because when the customer asks AI:
“Who should I choose?”
the business does not control the answer.
It can only influence the information from which that answer is constructed.
And somewhere between the prompt and the recommendation sits the Ghost.
SECTION 27
About Llumo
Llumo is an AI Search & Digital Marketing company founded by Musa Aykac, with more than 20 years of experience across digital marketing and technology.
Musa’s expertise spans SEO, PPC, AEO, GEO, AI search visibility, analytics, automation and digital growth.
Llumo focuses on helping businesses understand how they are discovered, interpreted and represented across emerging AI-driven search environments.
SECTION 28
Research Note
This report combines Llumo’s original exploratory experiments with external 2026 research into AI-assisted business discovery, recommendation and business accuracy.
The original experiments included cross-platform comparisons, repeated conversations, Google account observations and investigation of the information sources identified by AI systems.
The core prompt used in the restaurant experiments was:
“What are the 10 best restaurants in London?”
The experiment was designed to record outputs rather than predetermine conclusions. The research protocol specifically requires complete outputs to be recorded, anomalous runs to be preserved but excluded from quantitative calculations where necessary, and cited sources not to be treated as proof of causation.
No lock-in, just AEO for FREE.
