Brand Mentions are usually counted as if each occurrence meant roughly the same thing.
A company is named. A link appears. Someone recommends a product. Another person asks whether a brand is trustworthy. A sentence is repeated across several pages. A discussion connects the same company with a problem, a feature, a product and a result.
In a conventional report, all of those events can collapse into one simple number:
one more mention.
The FunkyMEDIA Research Lab study was built because that number was no longer enough.
Over approximately ten years, FunkyMEDIA has operated a publishing infrastructure in which brands, products, recommendations, questions and commercial discussions accumulated across publisher-owned websites. The current research inventory contains 1,779,805 unique discussion URLs.
These are not randomly selected pages collected from the open web.
They come from a long-running publishing environment owned and operated within the FunkyMEDIA ecosystem.
That history makes it possible to ask a different kind of question:
What does a Brand Mention actually contain once we stop treating every link, recommendation, question and repeated phrase as the same signal?
The dataset is large, but the study does not pretend that scale alone creates insight
The full inventory contains 1,779,805 unique discussion URLs.
That number defines the scale of the source material. It does not mean that the corpus contains 1.77 million Brand Mentions.
A discussion URL is the research inventory unit. A Brand Mention is a much narrower analytical unit that has to be identified inside the text.
To avoid confusing scale with depth, the research program was divided into separate layers.
The first is the Full Research Inventory: 1,779,805 unique discussion URLs.
The second is a Representative Validation layer built from 5,000 randomly sampled threads containing 8,708 posts. This layer was used to inspect the structure and chronology of the wider inventory.
The third is Gold Dataset v1, a deep-analysis layer containing 1,000 threads, 4,447 posts and 794 outbound links.
Gold Dataset v1 is where the research becomes much more granular. It supports the Brand Mention taxonomy, repetition analysis, intent calibration and semantic relationship extraction.
This separation is central to the study.
The large inventory tells us how much material exists.
The Gold Dataset tells us what selected Brand Mentions actually look like.
Those are different questions and should remain different datasets.
Finding 1: A link is not always a Brand Mention
One of the first assumptions tested in Gold Dataset v1 was simple:
If a discussion links to a company website, does that mean the company was explicitly mentioned?
The answer is no.
The taxonomy separated three structures:
416 Explicit Brand Mentions — Linked
In these records, the brand is visibly named and a matching link is present.
328 Unlinked Brand Mentions
The brand is visibly named, but no matching outbound link appears.
378 Linked Entity References
The text links to an entity-owned destination, but the visible wording does not necessarily name the brand.
This distinction changes the interpretation of link data.
If an analyst counted only links, the 328 unlinked Brand Mentions would disappear from the measurement.
If every brand-owned link were counted as an explicit Brand Mention, the 378 Linked Entity References would inflate visible naming.
The same raw export can therefore tell two very different stories depending on what is actually being measured.
A link shows where a page points.
A Brand Mention shows whether the brand itself appears in the visible language.
The two can overlap.
They are not interchangeable.
Why this matters
Brand monitoring, Digital PR and future AI Search research increasingly depend on evidence that exists outside a company's own website.
If link presence and visible naming are mixed together, it becomes difficult to tell whether the market is actually talking about the brand or simply linking to an entity-owned destination.
The taxonomy solves that problem by keeping the units separate.
Finding 2: Almost half of explicit Brand Mentions in Gold Dataset v1 had no link
The 744 explicit Brand Mentions in Gold Dataset v1 consist of:
- 416 linked explicit mentions
- 328 unlinked explicit mentions
That means 44.1% of explicit Brand Mentions in this calibrated dataset were unlinked.
This figure belongs to Gold Dataset v1. It is not presented as the percentage of all Brand Mentions in the 1.77-million-URL inventory or across the wider internet.
But within the calibrated layer it demonstrates something important:
explicit brand visibility does not depend on hyperlink presence.
A company can be named, recommended, questioned or discussed without receiving a link.
For researchers interested in brand presence, that makes unlinked mentions impossible to ignore.
Finding 3: Mention volume can hide repeated language
A second question concerned repetition.
If a brand appears many times, how much of that volume represents new language and how much repeats an existing formula?
Among the 744 explicit Brand Mentions in Gold Dataset v1:
- 63 were classified as exact formulaic mentions
- those mentions formed 21 exact formulaic clusters
- 681 remained unique-like under the current method
The phrase unique-like is intentionally cautious.
It does not mean organic.
It does not prove that the wording was independently written.
It means only that the current repetition method did not classify the text as an exact or near-text duplicate after the duplicate-detection process was applied.
Why repetition changes the meaning of a count
Imagine two brands with 100 mentions each.
For the first brand, most mentions repeat a narrow set of formulations.
For the second, the 100 mentions appear inside many different questions, use cases, product discussions and experiences.
The headline number is identical.
The information environment is not.
That is why the FunkyMEDIA research separates frequency from language diversity.
Mention volume still matters.
It simply needs context.
Finding 4: Brands often enter the conversation before they are recommended
Many brand reports focus on recommendations because recommendations appear close to a purchase decision.
But the Gold Dataset calibration shows that brands can become visible earlier.
Inside the 744 explicit Brand Mentions, the calibration layer identified:
- 248 question / consideration signals
- 179 recommendation signals
- 19 negative / warning signals
- 8 strong first-hand experience signals
- 4 comparison signals
These categories are calibration results, not estimates of full-corpus prevalence.
Their importance is conceptual.
A person can mention a brand while asking whether anyone has used it.
Another user can ask whether it is trustworthy.
A brand can appear as one option in a problem-solving discussion.
None of those situations requires a recommendation.
Yet the brand has already entered the decision process.
The decision journey begins before endorsement
This means Brand Mention research should not begin and end with the question:
“How many people recommended the brand?”
A more complete analysis asks:
When did the brand enter the conversation?
Was it being verified?
Was it being considered?
Was it being compared?
Was it recommended?
Was a result or experience described?
These are different stages of brand visibility.
Finding 5: One Brand Mention can contain several relationships at once
The study also examined how much structured information can exist inside one mention.
Consider a hypothetical form of sentence:
A user describes a problem, names a brand, refers to a specific product, explains a feature and reports a result.
That is still one explicit Brand Mention.
But semantically it contains several different relationships.
The semantic calibration extracted 536 relationship records from Gold Dataset v1.
After normalization, those records represented 276 unique Brand → Relation → Entity pairs.
The calibration includes 139 brands with at least one detected semantic relationship and 98 entities or relationship terms.
The observed relation families include:
- Use Case
- Product
- Service
- Feature
- Problem
- Result
- Location
- Competitor / Co-Mention
Use Case was the most frequent relationship family in the current calibration, with 288 records.
Other observed counts include:
- 64 Competitor / Co-Mention
- 48 Product
- 45 Feature
- 32 Service
- 23 Problem
- 23 Result
- 13 Location
From mention count to Brand Context Graph
This is where the research moves beyond counting.
A mention tells us that the brand appeared.
A semantic relationship tells us what the brand appeared with.
Across many mentions, those relationships begin to form a semantic territory around the brand.
That territory can contain products, problems, features, results, use cases, locations and other companies.
The Brand Context Graph is a way to visualize that structure.
It is not a universal knowledge graph.
It is a graph of relationships observed in the analysed text.
Finding 6: A Brand Mention can continue creating visibility without being owned by the brand
The infrastructure behind the research reveals another important property of Brand Mentions.
When a brand appears on a publisher-owned page, the brand can benefit from the visibility even though it does not own the underlying asset.
The domain belongs to the publisher.
The URL belongs to the publisher.
The surrounding editorial context belongs to the publisher.
The internal linking environment belongs to the publisher.
The brand benefits from the mention, but does not fully control the place where the mention exists.
This makes a Brand Mention different from owned media.
It can behave like a distributed brand asset: an external piece of brand visibility hosted inside somebody else's publishing infrastructure.
Why the ten-year infrastructure matters
Because the FunkyMEDIA research corpus comes from a long-running publisher environment, Brand Mentions can be observed as persistent objects rather than only publication events.
A mention may continue to exist years after publication.
Its surrounding page may continue to be crawled.
Its brand context may remain available to search engines, researchers and AI systems.
The current study does not prove that Brand Mentions cause AI recommendations, and it does not claim that persistence itself causes AI recommendations.
It does show that Brand Mentions can remain part of the public information environment long after the moment they were originally published.
That persistence deserves to be measured separately from publication volume.
The research contains real commercial brands
The corpus is not populated with synthetic company names or abstract examples.
The analysed material contains discussions referencing recognised consumer and B2B brands.
Examples observed in the research inventory include:
Allegro, Samsung, BMW, Toyota, IKEA, Lidl, Media Expert, Medicover, Volvo and Castorama.
These brands are shown because they were observed in the source research data.
Their presence demonstrates that the publishing infrastructure contains real commercial brand context.
The corpus by itself does not prove the exact contractual relationship behind every individual publication.
That distinction matters.
The research can demonstrate that a brand appeared in the analysed publishing environment without inventing a commercial relationship that the dataset itself does not document.
The macro layer shows how strongly the corpus is oriented toward decisions, opinions and recommendations
A separate macro subset contains 1,493,085 offline threads used for title, category and repetition analysis.
This is not the same layer as Gold Dataset v1 and is not used to calculate Brand Mention taxonomy percentages.
It provides a structural view of the wider discussion environment.
Within this macro subset:
1,036,637 threads — 69.43% — were located in review/recommendation-oriented paths or categories.
Title analysis identified:
- 307,749 opinion-oriented titles — 20.61%
- 308,715 recommendation-oriented titles — 20.68%
- 57,609 question-mark titles — 3.86%
- 127,123 titles containing company/business wording — 8.51%
Repetition analysis of titles also found:
- 39,203 exact duplicate-title groups
- 39,866 normalized repeated-title groups
- 156,339 threads belonging to exact duplicate-title groups
- 168,663 threads — 11.30% — belonging to normalized repeated-title groups
These are structural title and path findings.
They are not counts of explicit Brand Mentions.
Why include the macro layer
The macro subset helps explain the environment in which the deeper Brand Mention analysis sits.
The corpus contains a large amount of content framed around opinions, recommendations, companies and decisions.
That makes it a useful environment for studying how brands enter commercial conversations.
What the study does not prove
Large datasets make it easy to overstate conclusions.
This research deliberately separates what was observed from what remains a hypothesis.
The study does not prove that Brand Mentions cause AI recommendations.
It does not prove that unlinked Brand Mentions are universally more valuable than linked mentions.
It does not prove that repeated language is spam, automated or inauthentic.
It does not prove that unique-like language is organic.
It does not prove that two co-mentioned brands are direct competitors.
It does not prove that a Result relationship represents verified causality.
It does not claim that Gold Dataset percentages describe all 1,779,805 discussion URLs.
These boundaries are not weaknesses in the study.
They are part of the result.
A useful research framework should make it clear not only what the data supports, but also where the evidence stops.
Why FunkyMEDIA is publishing the research
FunkyMEDIA is an AI Search and Brand Mentions agency.
The agency works in an area where commercial claims can easily move faster than evidence.
Terms such as AI visibility, citations, recommendations, mentions and brand authority are increasingly used as if their meaning were obvious.
They are not.
FunkyMEDIA Research Lab was created to publish a measurable evidence layer behind that work.
The goal is to make the underlying datasets, findings, graphs and limitations available outside the agency's internal work.
That allows journalists, marketers, researchers and AI systems to inspect where a claim came from instead of receiving only a commercial conclusion.
The methodology and research interpretation are authored by Rafał Cyrański.
The research is published by FunkyMEDIA / FunkyMEDIA Research Lab.
The funkymedia.space service is operated by BIGTECH IT Sp. z o.o.
What comes next
The current publication establishes a baseline:
Brand Mentions are not one uniform signal.
They differ in link status, repetition, conversational role, contextual depth and semantic structure.
The next research questions become more interesting because this baseline now exists.
How long does a brand remain visible inside publisher infrastructure?
Which brands appear repeatedly across years rather than once?
How concentrated are mentions across recurring publishing sources?
How does the semantic territory around a brand change over time?
Which Brand Mention structures are most likely to persist?
And, separately, how much of this human discussion context is later reflected in AI-generated answers?
Those questions require additional analysis.
They should not be answered by assumption.
They are the next stage of the research program.
Research links
Brand Mention taxonomy — Not All Brand Mentions Are Equal →
Links vs naming — A Link Is Not Always a Brand Mention →
Repetition — The Repetition Problem →
Decision stage — Brand Mentions Before the Decision →
Semantic layer — Brand Context Graph →