Formulaic Language v1

Why repetition needs its own dataset

Raw Brand Mention volume can hide very different language structures.

Imagine two brands with 100 mentions each.

For one brand, the same sentence may repeat dozens of times.

For the other, nearly every mention may be phrased differently.

The total mention count is identical.

The wording diversity is not.

Formulaic Language v1 exists to make that difference measurable.

Exact formulaic mentions

The current calibration contains:

63 exact formulaic mentions

These are explicit Brand Mentions that fall into exact repeated wording patterns.

The classification is about textual repetition.

It does not automatically tell us why the wording repeated.

Possible causes can include:

reused editorial phrasing,

templates,

republication,

copying,

or other forms of repeated language.

The dataset records the repetition pattern.

It does not infer motive.

Exact formulaic clusters

The 63 exact formulaic mentions are organized into:

21 exact formulaic clusters

A cluster groups mentions that share the same exact repeated structure.

The cluster count is useful because it distinguishes:

how many repeated mentions exist

from

how many repeated patterns produced them.

Those are different questions.

Unique-like mentions

The current calibration contains:

681 unique-like mentions

That equals:

91.5% of the 744 explicit Brand Mentions

The term unique-like is intentionally cautious.

It does not mean organic.

It does not mean independently written.

It does not mean human-authored.

It does not mean original in an absolute sense.

It means only that the current repetition analysis did not classify the wording as an exact or near-formulaic duplicate.

That distinction must remain visible whenever the result is reported.

Why "unique-like" is not the same as "unique"

The research does not use the stronger label "unique" because text can differ without being fully independent.

Two passages can express the same idea with different wording.

A rewritten sentence may not be an exact duplicate.

A templated structure may contain enough variation to avoid exact matching.

That is why the safer term is:

unique-like

The dataset describes the observed wording pattern.

It does not make a provenance claim.

Near-formulaic calibration

After exact duplicates were removed, the current calibration found:

0 near-formulaic candidates at the 0.90 threshold

This result should be interpreted carefully.

It does not prove that no semantically similar wording exists.

It means that under the current calibration threshold, no additional near-formulaic candidates were identified after removing exact duplicates.

The threshold is part of the calibration result.

It should not be treated as a universal definition of similarity.

What 8.5% exact formulaic means

Within the 744 explicit Brand Mentions:

63 mentions were classified as exact formulaic.

That produces:

8.5%

This number describes exact repeated wording inside Gold Dataset v1.

It does not describe the entire Full Research Inventory.

It also does not tell us whether repeated wording is good or bad.

It tells us only that part of the calibrated Brand Mention layer contains exact repetition.

What 91.5% unique-like means

The remaining:

681 mentions

represent:

91.5%

of the explicit Brand Mention layer.

These mentions did not fall into the exact formulaic classification.

Again, this does not mean 91.5% of the mentions were organic, spontaneous or independently authored.

It means the wording was not classified as exact or near-formulaic under the current calibration.

Why repetition changes the meaning of volume

Suppose a brand has 500 Brand Mentions.

If 400 repeat the same sentence, that volume carries a different language-diversity profile from 500 differently phrased mentions.

The number 500 is still correct.

But the interpretation changes.

That is the central reason Formulaic Language v1 exists.

It allows mention volume and wording diversity to be reported separately.

Repetition does not automatically mean manipulation

Repeated wording can have many causes.

Some are deliberate.

Some are operational.

Some are editorial.

Some may result from syndication or reuse.

Formulaic Language v1 does not assign intent.

It does not label repeated language as spam.

It does not classify the source as manipulated.

It records a textual pattern.

Interpretation requires additional evidence.

Relationship to the Macro Subset

The Macro Subset contains a separate large-scale repetition layer based on titles.

That analysis found:

39,203 exact duplicate-title groups

39,866 normalized repeated-title groups

156,339 threads in exact duplicate-title groups

168,663 threads in normalized repeated-title groups

The normalized repeated-title groups account for:

11.30% of the 1,493,085-thread Macro Subset

These are title-level structural findings.

They are not the same as the Formulaic Language classification inside the 744 explicit Brand Mentions.

The two layers should remain separate.

Why the two repetition layers should not be merged

Formulaic Language v1 measures repeated wording inside explicit Brand Mentions.

The Macro Subset measures repeated title structures across a much larger thread population.

One is a text-level Brand Mention analysis.

The other is a structural title-level analysis.

Both can reveal repetition.

They do not measure the same thing.

Why this matters for Brand Mention reporting

A useful Brand Mention report should answer at least two separate questions:

How many mentions exist?

and

How much of that language repeats?

Without the second question, raw volume can create a misleading impression of diversity.

Formulaic Language v1 provides the repetition layer needed to interpret that volume more carefully.

Why this matters for AI Search research

External information environments can contain both repeated and varied descriptions of brands.

The current dataset does not prove that one pattern is more influential on AI systems than the other.

It creates measurable variables for future research:

exact repetition,

cluster size,

wording diversity,

and recurring phrasing.

Future experiments can test whether different repetition structures are reflected differently in AI-generated brand descriptions or recommendations.

That remains a future research question.

What this dataset does not claim

Formulaic Language v1 does not claim that repeated wording is manipulation.

It does not claim that unique-like wording is organic.

It does not claim that unique-like wording is independently authored.

It does not claim that repetition is inherently negative.

It does not claim that diverse wording is inherently better.

It does not claim that the 8.5% / 91.5% split applies to the Full Research Inventory.

And it does not claim that repetition automatically affects AI visibility.

The dataset has a narrower purpose:

to measure how much exact wording repetition exists inside the calibrated explicit Brand Mention layer.

The practical takeaway

A Brand Mention count measures frequency.

Formulaic Language v1 adds another dimension:

how much of that frequency comes from repeated wording?

For Gold Dataset v1:

744 explicit Brand Mentions

become:

63 exact formulaic

and

681 unique-like

That does not determine quality.

It makes repetition visible.

Related datasets

Gold Dataset v1
The source layer for the repetition analysis.

Brand Mention Taxonomy v1
The classification that defines the explicit Brand Mention layer.

Semantic Relationships v1
The semantic relationship layer.

Research Limitations
The boundaries that apply to repetition interpretation.