If the retrieval paths behind AI-generated answers are partly hidden, personalised and variable, what does an AI visibility score actually prove?
AI search has created a measurement problem that cannot be solved simply by replacing keyword rank tracking with prompt tracking.
In traditional search, the relationship between query and outcome was comparatively easy to conceptualise. A person entered a query, a search engine returned a ranked set of results, and marketers could monitor positions, impressions, clicks and eventual conversions. Personalisation, location and other variables existed, but the basic object of measurement remained recognisable: a query against a ranked result set.
AI-driven search introduces additional stages between the question a person asks and the answer they ultimately receive.
Google confirms that AI Mode uses query fan-out, dividing a question into subtopics and issuing multiple searches across those subtopics and data sources before combining the information into a response. This means the prompt entered by the user is not necessarily the complete set of retrieval queries used by the system.
The implications extend beyond query expansion. A retrieved source is not necessarily a selected source. A selected source does not necessarily become a visible citation. A company can be mentioned without its own website being cited, while a third-party source may provide evidence about that company without the source itself becoming the recommended entity. The same initial question can also operate under different contexts and produce different retrieval pathways.
The AI Search industry has already recognised much of this. Query fan-out, synthetic queries, probabilistic visibility, passage-level retrieval, personalisation and citation variability are increasingly being studied, modelled and incorporated into AI Search methodologies and platforms. iPullRank, for example, has documented query expansion, retrieval and selection-for-synthesis as distinct parts of the generative search process, while also exploring how personalisation can influence the fan-out produced from the same initial query.
Once those characteristics of AI Search are accepted, however, a harder question follows: what claims are we actually entitled to make from the measurements we produce?
If query fan-out is partly hidden, human demand is represented through samples of prompts, retrieval can vary between runs, personalisation can alter the context surrounding a question and different AI systems operate differently, then the challenge is no longer simply how to measure AI visibility.
It is determining what the resulting measurement can legitimately be said to represent.
When a dashboard reports that a company has 42%, 58% or 71% AI visibility, the obvious question should therefore be:
42%, 58% or 71% of what?
Query fan-out has already broken the idea of a single rank
Traditional SEO conditioned marketers to think in positions. A page ranked first, third, tenth or twentieth for a defined query. Even where results varied by geography, device or personalisation, the basic unit of measurement remained reasonably stable.
Generative search complicates that model considerably.
A user’s question can trigger related questions, comparisons, entities, attributes and subtopics that were never explicitly typed. Different retrieval processes can then gather information from multiple sources before a smaller set of information is used to construct the final answer.
iPullRank’s AI Search Manual describes this broader process in terms of query expansion, retrieval, source aggregation and a later selection for synthesis stage, in which the amount of information available to the system is narrowed before the answer is generated. Its work on personalised query fan-out extends the issue further by examining how different user contexts can influence the questions generated on a user’s behalf.
The implication is that there may no longer be one stable answer to the traditional question:
Where does this company rank for this query?
Visibility is better understood as something that can vary across expressions of intent, user contexts, retrieval pathways, sources, AI systems and answer generations.
That observation is increasingly accepted. But accepting probabilistic visibility immediately creates another measurement requirement:
An empirical AI visibility percentage needs a clearly defined denominator and measurement population.
Without those definitions, an apparently precise percentage can remain surprisingly difficult to interpret.
The denominator problem
Imagine an AI visibility platform reports:
Brand A: 63% AI Visibility
At first glance, the number appears precise. Yet several completely different methodologies could produce that same percentage.
Brand A may have appeared in 63 of 100 manually selected prompts. The prompts may instead have been synthetically generated by an LLM. The test could include only recommendation queries, or mix commercial questions with informational ones. ChatGPT, Gemini, Google AI Mode and Perplexity may have been averaged together. Every engine may have been weighted equally. Each prompt may have been tested once, or repeated several times. Any brand mention might count as visibility, or citation and recommendation might be treated as separate events.
Each approach could produce a mathematically valid percentage.
They would not necessarily produce percentages that mean the same thing.
The issue is therefore not automatically that the underlying data is inaccurate. It is whether the scope of the claim matches the scope of the measurement.
A monitored prompt set is not automatically equivalent to the market. It is a sample designed to represent some part of that market, and the quality of the resulting visibility estimate depends on what was sampled, how it was sampled, under which conditions it was tested and what event was ultimately counted.
We believe AI visibility presents at least two distinct sampling problems
Much of the current discussion around AI Search measurement focuses on which prompts should be tracked. Query fan-out introduces a second layer because marketers are effectively dealing with two large and only partially observable question spaces.
1. The human intent universe
No company, agency or visibility platform knows every question every potential customer could ask an AI system.
Consider accounting software. One person may ask:
What is the best accounting software?
Another may ask:
What accounting software would you recommend for a 20-person construction business in the UK?
Another:
What are the best alternatives to Xero if I need project costing?
Another:
Which accounting platform works well with Shopify and handles VAT?
A fifth buyer may express substantially the same commercial need in a way nobody included in the monitoring set.
The number of possible formulations expands rapidly once industry, geography, company size, use case, product requirements, pain points, previous products, comparisons, buying stage and conversational follow-ups are introduced.
An AI visibility programme therefore has to sample human intent.
The first measurement question is consequently:
How well does our tracked prompt universe represent the commercial demand we actually care about?
2. The machine retrieval environment
The human prompt is only the beginning.
Google says AI Mode can divide a question into subtopics and search for them simultaneously. A software recommendation could therefore require evidence relating to price, features, integrations, suitability, security, reviews, implementation, alternatives or other attributes before an answer can be formed.
Marketers do not receive a complete universal log of every internal expansion generated for every user. Context introduces further variation. Google documents that AI Mode can use previous searches and activity saved in Search Services History to provide personalised responses when the relevant settings are enabled.
iPullRank has similarly explored how different user contexts can result in different fan-out behaviour and different information pools.
This leaves two separate uncertainties:
Are we adequately sampling what people ask?
And:
Are our tests adequately representing the conditions under which AI systems interpret and investigate those questions?
These are related questions, but they are not the same measurement problem.
Prompt tracking is useful. It is not market truth.
None of this diminishes the value of prompt tracking.
A stable set of commercially relevant prompts can reveal whether a company’s visibility is improving, which competitors repeatedly enter answers, which sources influence responses, where the entity remains absent and whether changes to content or authority coincide with changes in observed visibility.
iPullRank itself describes AI Search measurement as a fundamentally different challenge from conventional SEO and advocates a layered approach rather than one traditional source of truth.
The problem begins when the boundary of the sample disappears from the language used to describe the result.
There is a meaningful difference between:
Brand A appeared in 31% of our representative prompt sample.
and:
Brand A has 31% visibility in AI Search.
The first tells us what was observed within a defined measurement environment.
The second can easily be interpreted as a statement about the wider market.
AI Search does not need less measurement. It needs better-defined measurement boundaries.
Seven questions we should ask before accepting an AI visibility score
1. What is the actual unit of visibility?
The word visibility risks becoming too broad.
A company can appear in an AI-generated response in several fundamentally different ways. It can be mentioned, enter a consideration set, be actively recommended, have its own website cited, or be discussed in a third-party source that receives the citation. Information relating to the company could also contribute to retrieval without the company appearing visibly in the final answer.
These events should not automatically be treated as interchangeable.
Consider two companies. Company A appears in 70% of monitored answers but is recommended in only 8%. Company B appears in 45% of monitored answers but is actively recommended in 35%.
Which company has stronger AI visibility?
The answer depends on what visibility is intended to mean. A single percentage conceals that distinction.
2. How was the prompt universe constructed?
More prompts do not automatically produce better measurement.
An LLM can generate thousands of plausible questions very quickly, but synthetic plausibility is not equivalent to market representation.
A stronger prompt universe might draw from customer research, sales conversations, first-party query data, existing search demand, support questions, known product use cases, competitor research, buyer personas and synthetic expansion.
The methodology matters because a large sample can still be systematically unrepresentative.
Precision is not the same as validity.
It is entirely possible to measure the wrong population extremely precisely.
3. Should every prompt count equally?
Consider two prompts:
Best CRM software
and:
Best CRM for a four-person Danish biotechnology company using Notion with an outsourced finance department
Both questions are legitimate, but it does not automatically follow that each should contribute exactly one unit to an organisation’s headline visibility score.
Some prompts represent broad category demand. Others describe highly specific circumstances. Some occur early in the buyer journey, while others indicate much stronger purchase intent.
This raises an increasingly important measurement question:
Should AI visibility eventually be weighted by intent?
There may need to be a distinction between prompt share and intent-weighted visibility.
Without weighting, an obscure synthetic variation could influence a headline metric as much as a major category-level buying question. That does not make an unweighted score invalid, but it does change what that score can reasonably be said to represent.
4. How many times was each prompt tested?
AI-generated answers can vary between runs.
One prompt run therefore gives us one observation. It does not necessarily establish the stability of that observation.
If a company appears when a recommendation query is tested today, we know that it appeared in that response. We do not yet know whether it would appear nine times out of ten, five times out of ten or once out of ten under comparable conditions.
Those represent very different competitive states.
Repeated testing introduces a dimension traditional rank tracking did not have to confront in quite the same way:
variance.
An entity that appears consistently has a different visibility profile from one that surfaces occasionally. Both could look identical if somebody happens to check the prompt only once.
The next generation of AI visibility measurement should therefore consider not only whether an entity appeared, but how stable that appearance is under comparable conditions.
5. Was the fan-out observed, inferred or simulated?
This distinction is critical.
Not every list of fan-out queries produced by a marketing tool is necessarily a literal record of queries executed internally by the target AI system.
There are at least three useful evidence classes.
Observed
The target system exposes retrieval behaviour that can directly be recorded.
Inferred
Likely retrieval branches are reconstructed from observable behaviour such as responses, citations, query perturbations or source overlap.
Simulated
Another model is used to approximate the types of subqueries the target system may generate.
All three approaches can provide value, but they are not equivalent evidence.
iPullRank, for example, has documented query perturbation as a method for investigating hidden retrieval branches, while Qforia uses Gemini to simulate likely query fan-out for analysis and content planning.
That is sophisticated and useful methodology. It also illustrates exactly why evidence labels matter.
What the target system exposes is different from what an analyst infers, and both are different from what another model simulates.
A convincing simulation does not automatically become an observation of Google’s internal process.
6. Can visibility across different AI systems simply be averaged?
Suppose a company records:
ChatGPT: 64%
Google AI Mode: 42%
Perplexity: 58%
The arithmetic mean is 54.7%.
But what precisely does 54.7% AI Visibility represent?
Different systems do not necessarily share identical retrieval architectures, indexes, citation behaviour, query expansion, personalisation, freshness mechanisms or user populations. They may also differ substantially in commercial importance for a particular business or market.
An aggregate metric can still be useful, but the weighting needs a reason.
Equal weighting is itself a methodological decision. Weighting based on platform usage is another. Weighting by buyer behaviour, geography, conversion value or strategic importance produces different composite measures.
The engine-specific results should therefore remain visible even where an overall benchmark is useful. Otherwise, several conditional measurements can be combined into a single figure that appears more universal than its components justify.
7. Where does retrieval stop and selection begin?
This may be the most important question for diagnosis.
Query fan-out concerns how a system expands and investigates an information need, but retrieval is not the end of the process.
iPullRank’s own query fan-out framework explicitly separates a later selection for synthesis stage from retrieval, describing a process in which a larger body of retrieved information is narrowed before information moves into the generated response.
That distinction matters.
A company can have strong content coverage but weak retrieval. It can be retrieved but fail to survive later selection. It can be selected as relevant but merely mentioned. It can be mentioned without being recommended. It can even be recommended while a third-party source receives the visible citation.
These are different competitive problems.
This is where the concept Netsleek refers to as the Selection Layer in AI Search becomes useful.
We use that term conceptually, not as a claim that every AI system contains a literal technical component carrying that name. It describes the point at which eligibility and retrieval are no longer sufficient explanations for the final observable outcome.
Query fan-out asks:
What might the system need to investigate?
Retrieval asks:
What information becomes available?
Entity resolution asks:
Which organisation, product, person or concept does that information relate to?
Corroboration asks:
What supporting or independent evidence exists around the entity and its claims?
Selection asks:
Which candidates and pieces of evidence ultimately survive into the response?
Recommendation asks:
Which entities does the generated answer actually put forward to the user?
We should be careful not to turn this into a deterministic formula. External corroboration does not allow us to predict selection with certainty, and the complete proprietary selection logic of these systems is not publicly observable.
The distinction between being retrievable and appearing in the final answer, however, remains critical.
Collapsing that entire chain into a single visibility percentage removes much of the diagnostic information marketers actually need.
The industry is already building sophisticated systems around fan-out
This measurement problem is not being ignored.
In July 2026, iPullRank formalised a partnership with AI Search intelligence platform Profound. The partnership combines Profound’s citation, source, prompt and visibility data with iPullRank’s Relevance Engineering methodology. iPullRank also contributed three agents to Profound’s Agent Template Marketplace: the Citation Gap Engine, Query Fan-Out Coverage Auditor and Explanatory Power Index Audit.
This development demonstrates where enterprise AI Search measurement is heading: increasingly sophisticated observational data combined with more sophisticated methodology around retrieval, fan-out, content and visibility.
It also introduces an important distinction.
The sophistication of a measurement system does not automatically expand the population its data represents. A highly rigorous measurement can still support a narrower claim than the headline metric suggests.
A platform can contain an enormous prompt dataset. A methodology can model likely fan-out with considerable sophistication. Repeated observations can improve confidence, and better citation and source data can improve interpretation.
None of those advances removes the need to ask:
What population does this measurement legitimately allow us to make claims about?
This is not an argument against sophisticated measurement. It is precisely why sophisticated measurement deserves equally sophisticated interpretation.
AI visibility needs evidence boundaries
At Netsleek, we believe AI Search reporting should make several distinctions explicit.
Observed refers to what was directly recorded from the target AI environment or another first-party observational source.
Sampled describes the prompts, intents, markets, engines, languages, dates, sessions and repeated runs that formed part of the measurement.
Inferred describes conclusions reconstructed from observable behaviour.
Simulated describes behaviour approximated through another model because the underlying process could not be directly observed.
Weighted describes observations that were deliberately given greater or lesser importance and the reasoning behind that decision.
None of these evidence classes is inherently weak. Simulation can be highly useful. Inference is fundamental to analysis. Sampling is unavoidable. Weighting can make a measurement more representative.
The problem begins when those boundaries disappear.
A simulated fan-out becomes the fan-out. A prompt sample becomes the market. An inferred retrieval path becomes what the model searched. A blended multi-engine score becomes AI visibility without qualification.
The language becomes more certain than the evidence.
The central risk in AI visibility measurement may not be inaccurate data. It may be making a broader claim than the data was designed to support.
Good measurement should do the opposite. It should make the limits of the evidence visible.
A better analytical model of AI Search visibility
Instead of treating AI Search as:
Prompt → Position
a more useful analytical model is:
Human Intent Universe
↓
Sampled Prompt Universe
↓
Context and Personalisation
↓
Query Expansion / Fan-Out
↓
Retrieval
↓
Entity Resolution
↓
Available Evidence and Corroboration
↓
Selection
↓
Answer Synthesis
↓
Mention / Citation / Recommendation
↓
Referral / Conversion / Commercial Outcome
This should not be interpreted as a claim that every AI Search system implements this exact architecture in this exact sequence. It is an analytical measurement model intended to distinguish stages that may influence the observable outcome.
Its value is diagnostic.
If a company is absent from an AI-generated recommendation, the question can no longer simply be:
Why don’t we rank?
Instead, we can ask whether the brand adequately covers the relevant intent, whether the required information exists, whether retrieval systems can access and interpret it, whether that information is clearly associated with the correct entity, what external evidence exists around the organisation and its claims, whether the company appears to enter relevant candidate sets, whether it surfaces in citations but not recommendations, and whether competitors are consistently being surfaced from evidence environments in which the company is absent.
Those questions point towards different interventions.
A headline visibility percentage does not.
Query fan-out should also change how we think about existing content
The implications are not limited to dashboards. They also change how organisations should evaluate the information they already possess.
Consider a semiconductor manufacturer.
It may have spent decades producing white papers, research studies, technical papers, application notes, engineering documentation, fabrication data, product specifications, testing results and highly specialised R&D resources.
A buyer, engineer or researcher asks:
Which US semiconductor manufacturers are strongest in advanced power semiconductor technology?
Behind that apparently straightforward question may sit a much broader information requirement. The system may need evidence relating to the organisation itself, fabrication capabilities, process nodes, wafer technologies, materials, chip architecture, advanced packaging, testing, performance characteristics, applications, manufacturing capability, technical research and external authority.
The manufacturer may already possess extraordinary evidence across most of these areas, but much of it may exist inside a 40-page technical white paper created primarily for specialist human readers.
The information exists.
That does not mean every important finding, capability and relationship is equally retrievable.
This is where content engineering becomes important. Highly technical information can be structured so that important facts, findings and relationships are easier for retrieval systems to locate and interpret without removing the scientific or technical depth required by specialist audiences.
A single white paper may contain evidence relevant to numerous potential fan-out branches.
The opportunity is therefore not simply to publish more content. It is to make important existing knowledge easier to find, interpret, attribute and contextualise.
Even excellent fan-out coverage, however, should not automatically be described as visibility.
Being relevant to a likely subquery is not the same as being retrieved. Being retrieved is not the same as surviving selection. Being selected is not the same as being cited. Being cited is not the same as being recommended.
Those distinctions should survive all the way from content strategy to reporting.
What should an AI visibility report actually disclose?
If AI visibility is to become a serious business metric, organisations should be able to understand how the number was constructed.
A credible report should explain how the prompt universe was selected, which intent families were represented, whether prompts were human-generated, synthetically generated or both, which AI environments were tested, the geographic and language conditions, whether sessions were personalised or controlled, how often prompts were repeated and over what period the testing occurred.
It should also make clear whether mentions, citations, consideration-set inclusion and recommendations were measured separately.
Where weighting is used, the weighting logic should be stated. Where fan-out is discussed, the report should clarify whether it was observed, inferred or simulated.
Most importantly, the report should define its numerator and denominator.
This does not weaken the metric.
It makes the metric auditable.
AI visibility is not unknowable. It is conditional.
There is a danger of taking this argument too far.
If AI answers vary, fan-out is partly opaque and the complete universe of human demand cannot be observed, it would be easy to conclude that AI Search cannot meaningfully be measured.
That is not our position.
AI visibility can be measured.
The important point is that the measurement is conditional on its design.
A carefully constructed prompt sample can reveal whether visibility is improving. Repeated testing can reveal whether recommendations are becoming more stable. Intent segmentation can show where a company performs strongly and where it disappears. Citation analysis can identify sources that repeatedly influence answers. Competitive monitoring can identify which entities consistently enter consideration sets. Referral and conversion data can connect at least some AI discovery activity to real business outcomes.
These are valuable observations.
Their value increases when we state precisely what they establish.
Compare:
Brand A has 46% AI visibility.
with:
Brand A was recommended in 46% of 300 sampled UK accounting-software buyer prompts tested across three AI environments during August, with each prompt repeated three times.
The second statement is longer, but it is also far more informative.
We know what was measured. We know the sample. We know the market. We know the event being counted. We know the test conditions.
The first asks the reader to assume all of those things.
The next generation of AI visibility measurement should measure uncertainty too
Traditional SEO trained marketers to expect a position.
AI Search requires a different mindset.
Future measurement should increasingly report not only how often an entity appears, but how stable that appearance is, which intent families produce it, which engines produce it, which sources support it, whether the entity is merely mentioned or actively recommended and how sensitive the outcome is to wording, context and repeated runs.
Uncertainty should not automatically be treated as noise that needs to be hidden behind a clean percentage.
Uncertainty is itself information.
A brand appearing inconsistently across an important commercial intent family has a different problem from a brand that never appears at all.
One may already be competitive but unstable. The other may be failing earlier in retrieval, entity association or selection.
A single average can make both appear weak while concealing completely different underlying causes.
Query fan-out changes the question we should be asking
Query fan-out has changed the architecture of search.
It should also change the language of measurement.
The future of AI Search reporting should not simply become a traditional rank tracker with prompts substituted for keywords. It should distinguish human demand from synthetic prompts, samples from populations, observed behaviour from inferred behaviour, inference from simulation, retrieval from selection, mentions from citations, citations from recommendations, engine-level performance from blended scores, and measurement precision from the scope of the claim being made.
The question is therefore no longer simply:
Can AI visibility be measured?
It can.
Nor is the most important question:
How do we track query fan-out?
The harder question is:
What exactly have we measured, how representative is it of the market we care about, and what claims does the evidence actually justify?
AI Search measurement will continue to improve. Prompt datasets will become larger. Retrieval analysis will become more sophisticated. Visibility platforms will collect more observations. Simulation techniques will improve.
Better tools, however, do not eliminate the responsibility to define what a measurement actually represents.
In many cases, the next major advance in AI visibility measurement may therefore not be another metric.
It may be greater precision about what our existing metrics are entitled to claim.
Frequently Asked Questions About Query Fan-Out and AI Visibility Measurement
What is query fan-out in AI Search?
Query fan-out is a retrieval technique in which an AI Search system expands an initial question into multiple related subtopics or searches so that information can be gathered across those branches before a response is synthesised. Google explicitly confirms that AI Mode divides questions into subtopics and searches across multiple data sources.
Can marketers see Google’s exact query fan-out?
Not completely. Some retrieval behaviour may be visible under certain systems or interfaces, while other likely branches can be inferred through observable outputs or simulated using other models. iPullRank, for example, describes query perturbation as a way of exploring hidden retrieval branches and also provides tooling for simulated fan-out analysis.
Observed, inferred and simulated fan-out should therefore not be treated as equivalent evidence.
Does query fan-out mean AI visibility cannot be measured accurately?
No. It means accuracy has to be defined relative to the measurement being attempted. Controlled prompt samples, repeated testing, intent segmentation, citation analysis, recommendation tracking and outcome data can provide valuable evidence when the boundaries of the sample are clear.
What does an AI visibility percentage actually measure?
That depends entirely on the methodology. It could represent the percentage of monitored prompts in which a brand appeared, recommendation frequency, citation frequency, consideration-set inclusion, share of voice or a blended score across several systems. The numerator and denominator should therefore be explicitly defined.
Why is one prompt run not enough?
One run records one outcome. AI-generated responses can vary, so repeated runs help determine whether an entity’s appearance is stable or occasional. This provides information about variance rather than treating every isolated observation as equally representative.
Should all AI Search prompts receive equal weighting?
Not necessarily. Different prompts can represent very different levels of commercial importance, market demand and buying intent. An unweighted prompt-share metric can still be useful, but intent-weighted visibility may provide a more meaningful representation for some business questions.
Can ChatGPT, Gemini, Google AI Mode and Perplexity visibility be combined?
They can be combined into a deliberately constructed benchmark, but the weighting methodology should be clear. These systems do not necessarily behave identically, so engine-level performance should remain available alongside any blended score.
What is the difference between an AI mention, citation and recommendation?
A mention means the entity appears in the generated response. A citation means a source is visibly attributed or linked as supporting information. A recommendation means the system actively puts the entity forward as an appropriate choice. A company can achieve one without necessarily achieving the others.
What is the difference between retrieval and selection?
Retrieval concerns information becoming available to the system. Selection refers conceptually to the later narrowing of candidate information for inclusion in the generated answer. Strong retrieval therefore does not automatically guarantee a mention, citation or recommendation.
Is simulated query fan-out useful?
Yes. Simulation can help explore likely subtopics, identify content gaps and model potential retrieval requirements. It becomes problematic only when the simulation is described as though it were a direct observation of the target system’s internal query process.
What should businesses measure instead of a traditional AI rank?
There is unlikely to be one universal replacement metric. Stronger AI Search measurement can consider mention rate, recommendation rate, citation rate, consideration-set inclusion, competitive share, intent coverage, source visibility, stability across repeated observations, AI referral traffic and eventual commercial outcomes.
References
Google Search Help. AI Mode in Google Search.
https://support.google.com/websearch/answer/16011537
Google Search Help. Personalisation in AI Mode.
https://support.google.com/websearch/answer/17212611
iPullRank. Query Fan-Out — The AI Search Manual.
https://ipullrank.com/ai-search-manual/query-fan-out
iPullRank. Query Fan-Out and Personalization.
https://ipullrank.com/query-fan-out-personalization
iPullRank. Measurement for Generative Engine Optimization — The AI Search Manual.
https://ipullrank.com/ai-search-manual/measurement-geo
iPullRank. AI Search Manual — Quick Start Guide.
https://ipullrank.com/ai-search-manual/quick-start-guide
PR Newswire. iPullRank Named a Profound Agency Partner, Formalizing Enterprise AI Search Collaboration. 31 July 2026.
https://www.prnewswire.com/news-releases/ipullrank-named-a-profound-agency-partner-formalizing-enterprise-ai-search-collaboration-302839544.html
Research Authors
Ruan Masuret and Juanita Martinaglia are the co-founders of Netsleek, an AI Search & Brand Discoverability agency focused on how brands, information and digital entities are discovered across AI-driven search and recommendation environments.
Their work examines the changing mechanics of modern discovery, including query expansion and fan-out, information retrieval, entity resolution, corroboration, AI visibility measurement, machine-readable content architecture, agentic search, and the selection processes that influence which information, sources and entities are ultimately surfaced in AI-generated answers.
Through Netsleek, their research explores the relationship between technical search infrastructure, semantic information architecture, entity evidence and AI-driven discovery, with particular attention to developing measurement and optimisation approaches that distinguish what can be directly observed from what must be sampled, inferred or simulated.