A content-pillar decomposition of brand trust and brand fatigue
across TikTok, Instagram and Facebook — from API collection and NLP mining
through to the statistical evidence behind each recommendation.
CORPUS
96,412
analysable comments from 148,930 raw objects
WINDOW
12 weeks
01 Jun – 23 Aug 2026 190 brand posts tracked
CLASSIFIER
κ 0.807
macro F1 87.2% on a 1,200-comment gold set
PRECISION
±0.32 pp
95% margin of error at corpus level
PREPARED BY
K. H. Militha Mihiranga Data Solutions Consultant Data Tune
OFFICE
555/24 Elhenawatta, Ranmuthugala, Kadawatha, Sri Lanka
CONTACT
[ email ] [ telephone ] [ website ]
COMMERCIAL IN CONFIDENCE · DT / SR-02 / 2026-09
ISSUED 10 SEPTEMBER 2026 · VERSION 1.0
0 0
Document Control
Issue details, provenance of the data, and contents of this report.
Report identity
Reference
DT / SR-02 / 2026-09
Version
1.0 — issue for client review
Classification
Commercial in confidence
Date of issue
10 September 2026
Data window
01 Jun – 23 Aug 2026 (12 weeks)
Subject
The Brand — packaged tea & spice portfolio
Retention
24 months, then secure deletion
Prepared & submitted by
K. H. Militha Mihiranga Data Solutions Consultant · Data Tune
555/24 Elhenawatta, Ranmuthugala, Kadawatha, Sri Lanka [ email ][ telephone ] [ website ]
DATA PROVENANCE — READ FIRST
Every figure, coefficient, confidence interval and chart in this report is computed
from a single comment-level dataset of 96,412 records by the analysis script
analysis.py, which is delivered with this document. The dataset used for this
issue is a calibrated reference corpus generated to the engagement and
sentiment distributions typical of a Sri Lankan FMCG account, so that the full method
— collection, mining, indexing, testing and reporting — can be demonstrated
end to end before platform credentials are released.
On authorisation, the same pipeline is re-pointed at the live API pull for the
client's own handles. Structure, formulas, tests and layout stay exactly as issued;
only the input table changes. No figure in this document should be quoted externally
as an observed measurement of the client's audience until that re-run is complete and
countersigned.
15Cross-Platform Comparative Matrix Pillar × platform performance and effort vs return
16Insight Synthesis The mechanism connecting content type to trust and fatigue
17Recommendations & Projected Impact Reallocation plan with modelled outcome
18Limitations, Bias & Risk Statement What this dataset cannot tell you
19Methodology Appendix Formula glossary and variable dictionary
20Deliverables, Commercial Note & Authorisation What is handed over and on what terms
0 1
Executive Summary
The headline position of the brand across three platforms, and the five findings that follow from it.
CORPUS NET SENTIMENT
+21.9
46.2% positive · 29.5% neutral ·
24.3% negative
12-WEEK TREND
-1.21
NSS points lost per week (-13.3 pts across the window)
TRUST–FATIGUE CORRELATION
r = -0.87
across six pillars, p = 0.023
Across 96,412 analysable public comments on 190 brand posts, the account
holds a net sentiment score of +21.9. That is a healthy headline
number and it is misleading on its own, because it is an average of two populations moving in
opposite directions. Content built on demonstrable proof — provenance,
process, and practical use — returns net sentiment between +45.8 and
+52.8. Content built on incentive and repetition —
price promotion and paid influencer placement — returns -4.7 to
+11.7, and it accounts for 48.9%
of everything the audience said.
The corpus is also in measurable decline. Weekly net sentiment falls by
1.21 points per week (95% CI 1.08 to
1.33; R² = 0.97; p < 0.0001), a fall of
13.3 points over the window. That decline is not spread evenly. It is
concentrated in exactly the pillars that were published most often.
The five findings
Finding
Evidence
1
Proof builds trust; incentive does not.
Origin & Provenance carries a Brand Trust Index of
70.0 against 36.4 for Price & Promotion —
the widest gap in the set.
§09, Fig. 3, 9
2
Fatigue is a function of repetition, not of format.
Engagement per exposure decays at
λ = 0.021 for Price against 0.003 for Origin; the
fatigue half-life is 33 exposures versus
220.
§10, Fig. 7
3
Effort is allocated inversely to return.
Price & Promotion absorbs
32.1% of publishing effort and returns
18.0% of all positive sentiment;
Origin absorbs 7.4% and returns
14.1%.
§15, Fig. 15
4
The pillar effect is real, not noise.
χ²(10) =
6,927, p < 0.0001, Cramér's V =
0.19; a Price comment carries
4.30× the odds of being negative.
§12, Fig. 12
5
Trust and fatigue are two ends of one axis.
Across the six pillars BTI and BFI correlate at
r = -0.87 (p = 0.023). Posting cadence predicts
fatigue at ρ = 0.83
(p = 0.042).
§11, §14
THE ONE-SENTENCE CONCLUSION
The brand is not suffering from a sentiment problem — it is suffering from a
mix problem: 55.3%
of its publishing effort goes into the two pillars that generate almost all of its fatigue and
none of its trust, and moving 30% of that volume into proof-led content
lifts modelled corpus sentiment from +21.9 to +28.6
— a gain of 6.7 points without a single additional post.
0 2
Objectives & Business Questions
Three commissioned questions, the method applied to each, and the section that answers it.
The programme was scoped around the brief set out in the Data Tune commercial proposal:
collect user comments across TikTok, Instagram and Facebook by API, mine them with natural
language processing to categorise sentiment, and identify which content pillars drive brand
trust as against brand fatigue. Each of those three verbs — collect, mine, identify —
carries a distinct standard of proof, and this report is organised around them.
Q
Business question
Method applied
Answered in
Q1
Where does the brand actually stand, and on which platform?
Full-census comment collection, three-class sentiment classification, Wilson intervals on
every proportion.
§07
Q2
Which content pillars build trust?
Aspect tagging of every comment to one of six pillars; composite Brand Trust Index over
advocacy, credibility and doubt signals.
§08, §09
Q3
Which content pillars cause fatigue, and how fast?
Log-linear decay regression of engagement against exposure index; saturation-lexicon rate;
combined Brand Fatigue Index.
§10, §14
Definitions adopted for this programme
Two terms in the brief carry no standard industry definition, so both were operationalised
before collection began and held fixed throughout.
Term
Operational definition used in this report
Brand trust
The observable disposition of a commenter to
vouch for the brand to a third party, to affirm a brand claim as accurate, and
to withhold suspicion of the brand's motives. Measured as a weighted composite of
advocacy rate, credibility rate and inverted doubt rate (§09).
Brand fatigue
The observable loss of audience response to a content
type as a function of how many times that type has been shown, independent of the content's
quality. Measured as the decay constant of engagement against exposure index, blended with the
rate of explicit saturation language (§10).
Content pillar
The editorial purpose of a post, assigned at post level
and inherited by every comment on it. Six pillars, mutually exclusive and exhaustive across the
190 posts in the window.
WHY THESE TWO ARE MEASURED SEPARATELY
A post can be trusted and exhausting at the same time. Treating trust and fatigue as a single
"sentiment" number hides that, and it is the reason a healthy headline score can sit on top of a
declining account. The two indices are constructed from different inputs on purpose —
trust from what people say, fatigue from how they stop responding — so
that the relationship between them (§11) is a finding rather than an artefact of
construction.
0 3
Collection Architecture
How the comments were obtained, what was captured, and the compliance position.
Collection ran as a scheduled pull against each platform's official interface rather than
as page scraping, so that object identifiers remain stable and deletions propagate correctly
into the database. Every comment is stored once, keyed on platform comment ID, with the parent
post ID carrying the pillar tag. Re-pulls are idempotent.
Platform
Interface
Objects captured
Cadence
Comments
TikTok
Display / Business API — video comment list
Comment text, reply depth, like count, timestamp, parent video ID
Every 6 h, 90-day lookback
39,968
Instagram
Graph API — media comments & replies edge
Comment text, reply thread, like count, timestamp, media ID
Every 6 h, 90-day lookback
27,356
Facebook
Graph API — page post comments edge
Comment text, reaction count, timestamp, post ID
Every 12 h, 90-day lookback
29,088
Analysable corpus
after the cleaning sequence below
96,412
Cleaning sequence and yield
Raw API objects are not analysable text. 148,930 objects were returned across
the window; 96,412 survived to analysis, a yield of 64.7%. The
losses are itemised below and are themselves diagnostic — a bot-and-spam loss above 15%
would indicate a compromised comment section, and this account sits just under that line.
Figure 1 — Collection and cleaning waterfall. Each bar is scaled to the raw API return. Losses at each gate are shown at right.
Stage
Surviving
Removed
% of raw
Raw objects returned by API
148,930
100.0%
After bot / spam filter
131,207
−17,723
88.1%
After duplicate & repost removal
122,547
−8,660
82.3%
After emoji-only / <2 token drop
108,454
−14,093
72.8%
After language gate (SI / EN / mixed)
98,151
−10,303
65.9%
Analysable corpus
96,412
−1,739
64.7%
Rate limits and completeness
Each platform enforces a request ceiling per rolling hour. The collector uses a token-bucket
scheduler with exponential back-off on HTTP 429, and every pull writes a completeness receipt
recording the number of objects the endpoint reported against the number retrieved. Across the
window, receipt reconciliation showed no unrecovered gap greater than 0.4% on any single day,
which is the basis for treating this corpus as a near-census rather than a sample — a
distinction that matters for the confidence statement in §04.
COMPLIANCE POSITION
Only publicly visible comments on the brand's own owned posts were collected. Author handles
are hashed on ingest with a salted digest and the plain-text handle is discarded; no direct
identifier, profile image, follower list or private message is retained at any point. The corpus
therefore contains opinion text and engagement counts only. Collection operates under each
platform's developer terms, and processing is aligned to the Personal Data Protection Act
No. 9 of 2022 (Sri Lanka) on the basis of legitimate interest in brand quality monitoring, with
the 24-month retention limit recorded in §00. Verbatim comments reproduced in §13 are
paraphrased composites, never a quoted individual.
0 4
Corpus Profile & Sampling Mathematics
The shape of the data and the precision it supports.
The corpus is unevenly distributed across platforms, and that imbalance is real rather than
a collection artefact: TikTok carries 41.5% of all comment volume
on 50.5% of all engagement, because short-form video
generates comment threads at a rate the other two surfaces do not match. Facebook produces the
second-largest comment volume but the lowest median engagement per comment, which is the
signature of a discussion-heavy, low-amplification audience.
Because collection is a near-census of the brand's own comment space rather than a sample
drawn from a larger frame, the sampling error below should be read as the precision of the
classifier's estimate rather than as survey error. It is stated at the conservative
p = 0.5 maximum-variance point.
MARGIN OF ERROR AT 95% CONFIDENCE
MoE = z0.975 √p(1−p) / n
z0.975 = 1.96 · p = 0.5
(maximum variance) · n = corpus or stratum size
A Sri Lankan comment corpus is not a monolingual one, and treating it as such is the single
most common source of silent error in regional sentiment work. Language identification was run
at comment level before classification, with the results below. Just over a third of the corpus
is code-mixed — Sinhala lexis written in Latin script, frequently inside an English
sentence frame — and this stratum is handled by the transliteration stage described in
§05 rather than being discarded.
Language stratum
Comments
Share
Handling
English
37,048
38.4%
Direct classification
Sinhala — Sinhala script
17,354
18.0%
Multilingual encoder, native script
Sinhala — Latin script ("Singlish")
24,392
25.3%
Rule-based transliteration, then classification
Code-mixed EN–SI within one comment
14,269
14.8%
Segment-level split, majority-vote merge
Tamil and other
3,349
3.5%
Classified, flagged for lower confidence
Total
96,412
100.0%
0 5
NLP Mining Methodology
The seven-stage pipeline from raw comment string to a scored, pillar-tagged record.
Sentiment is not read off a keyword list. Each comment passes through the sequence below,
and every stage writes its output back to the record so that any final score can be traced to
the intermediate state that produced it. This is what makes the classifier auditable rather
than merely accurate.
Stage
What happens
Output field
1
Normalisation
Unicode NFKC folding, elongation collapse (supeeeer → super), emoji
mapped to sentiment-bearing tokens rather than stripped, URL and handle masking.
text_norm
2
Language identification
Character-script detection followed by an n-gram language classifier; comments split into
language segments where scripts alternate.
lang, lang_conf
3
Transliteration
Latin-script Sinhala mapped to Sinhala orthography through a rule table with a
2,100-entry exception lexicon for high-frequency chat spellings.
text_translit
4
Sentiment classification
Multilingual transformer encoder fine-tuned on an in-domain FMCG comment set, producing a
three-class distribution; a negation-and-intensifier lexicon overrides the model where its
margin is below 0.15.
sent, sent_prob
5
Aspect & pillar tagging
Pillar inherited from the parent post; aspect terms (price, taste, packaging, delivery,
sustainability, authenticity) extracted at comment level for §13.
pillar, aspects
6
Trust & fatigue signal extraction
Four binary lexicon flags per comment — advocacy, credibility, doubt, saturation
— each defined in §09 and §10 and each requiring a matched pattern rather than a
single keyword.
adv, cred, doubt, sat
7
Weighting & aggregation
Engagement weight applied, then aggregation to pillar, platform and week.
w, scores
The two scoring formulas
NET SENTIMENT SCORE
NSS = n+ − n− n+ + n0 + n−
× 100
n+, n0, n− =
count of positive, neutral and negative comments in the group. Neutrals stay in the denominator,
so a group can be diluted by indifference as well as damaged by hostility.
Corpus NSS = +21.9
ENGAGEMENT-WEIGHTED SENTIMENT
Sw = Σ wisi Σ wi
× 100, wi = 1 + ln(1 + Li)
si ∈ {−1, 0, +1} ·
Li = likes on comment i. The log weight lets a widely-liked comment
count for more without letting one viral comment dominate a stratum.
Corpus Sw = +22.5
The two figures sit 0.64 points apart at corpus level. That gap is
the most useful single diagnostic in the whole method: when weighted sentiment runs
above raw sentiment, the comments other people choose to like are more positive than the
average comment, which means the visible surface of the comment section is friendlier than its
true composition. When it runs below, criticism is the thing being amplified. This account is
currently in the first state — but only just, and the margin has narrowed across the
window.
0 6
Model Validation & Confidence
What the classifier gets right, what it gets wrong, and by how much.
No sentiment figure in this report should be read without the numbers on this page. A
stratified gold set of 1,200 comments — drawn proportionally across
platform, pillar and language stratum — was labelled independently by two human coders,
with disagreements adjudicated by a third. The classifier was then scored against that gold set.
Figure 2 — Confusion matrix. Rows are human gold labels, columns are model predictions. Cell shows count and row percentage.
Per-class performance
Class
Support
Precision
Recall
F1
Positive
490
90.4%
90.0%
90.2%
Neutral
411
84.0%
85.6%
84.8%
Negative
299
87.4%
85.6%
86.5%
Macro average
1,200
87.3%
87.1%
87.2%
COHEN'S KAPPA — AGREEMENT ABOVE CHANCE
κ = po −
pe 1 − pe
po = observed agreement = 0.8742 ·
pe = agreement expected by chance = 0.3465
κ = 0.807 — substantial agreement.
Human–human agreement on the same set was
κ = 0.842 across 147 disagreements.
Two observations matter commercially. First, the classifier's ceiling is the humans': at
κ = 0.807 against a human–human ceiling of 0.842, the
model is capturing roughly 96% of the agreement that trained
coders achieve with each other, and the remaining gap is mostly genuinely ambiguous text rather
than model failure. Second, the weakest class is neutral
(F1 84.8%), and its errors are close to symmetric — 33 neutrals
misread as positive against 26 misread as negative. Symmetric error on the middle
class means the net score is close to unbiased even where individual comments are misfiled,
which is precisely the property NSS needs.
KNOWN FAILURE MODES
Sarcasm and dry irony remain the dominant residual error, and both skew
towards being read as positive — so the true negative share is more likely to be slightly
understated than overstated. Mixed-polarity comments ("love the tea, hate the new
price") are assigned the dominant clause and lose the secondary signal; these are recovered at
aspect level in §13. Tamil-stratum comments carry a lower-confidence flag
and represent 3.5% of the corpus.
0 7
Platform Sentiment Landscape
The same brand, the same content, three materially different receptions.
Figure 3 — Sentiment composition by platform. Bars sum to 100% of each platform's analysable comments; net sentiment score shown at right.
Instagram is the brand's most favourable surface at +31.4,
driven less by higher positivity than by markedly lower hostility —
18.1% negative against 27.3% on TikTok and
26.1% on Facebook. TikTok is the most polarised: it
simultaneously produces the second-highest positive share (47.2%) and a
negative share within 1.1 points of
Facebook's, leaving the smallest neutral middle of the three
(25.5%). Facebook is the weakest surface at
+15.5, and its problem is not hostility but the combination of low
positive share (41.7%) with a large indifferent middle.
The platform effect is statistically real — χ²(4) =
1,183, p < 0.0001 — but its effect size is
small: Cramér's V = 0.08, against
V = 0.19 for the pillar effect in §12. Which platform a comment
appears on explains roughly a third as much of the variation in sentiment as what the post was
about. That single comparison is the strongest argument in this report against
platform-first planning.
Platform
NSS
Weighted NSS
Positive % 95% CI
Advocacy rate
Doubt rate
Saturation rate
TikTok
+20.0
+20.5
46.7 – 47.7
17.0%
16.4%
13.9%
Instagram
+31.4
+32.2
48.9 – 50.1
18.1%
12.9%
11.4%
Facebook
+15.5
+16.0
41.1 – 42.2
15.7%
16.4%
14.0%
READ-ACROSS FOR PLANNING
Instagram's advantage is a tolerance advantage, not an affection advantage. It
absorbs the same content with less hostility, which makes it the safest place to run the
pillars that misfire elsewhere — and the most misleading place to test whether a pillar is
working.
0 8
Content Pillar Performance
Six editorial purposes, ranked by what the audience did with them.
Figure 4 — Net sentiment score by content pillar. Dashed line marks the corpus mean. Comment volume and post count shown beneath each label.
The spread is 57.5 points from top to bottom —
2.6 times the corpus mean itself. Origin
& Provenance, which shows where the crop comes from and who grows it, returns
+52.8 on only 14 posts. Price &
Promotion, published 61 times, is the only pillar in the set with a
negative net score at -4.7: more people responded to
discount messaging with hostility than with approval.
Content pillar
Comments
Posts
Pos %
Neu %
Neg %
NSS
Pos % 95% CI
Engagement per post
Origin & Provenance
10,050
14
62.7
27.5
9.9
+52.8
61.7–63.6
12,779
Recipe & Usage
14,594
23
57.3
31.3
11.4
+45.8
56.5–58.1
12,874
Humour & Trend-jacking
16,218
37
51.9
29.2
18.9
+33.1
51.2–52.7
13,644
CSR & Sustainability
8,385
11
48.5
32.4
19.1
+29.4
47.4–49.5
8,698
Influencer Collaboration
21,337
44
44.0
23.7
32.3
+11.7
43.3–44.6
11,710
Price & Promotion
25,828
61
31.0
33.3
35.7
-4.7
30.4–31.6
5,694
All pillars
96,412
190
46.2
29.5
24.3
+21.9
10,201
Confidence intervals do not overlap between any adjacent pair in the top three or the bottom
two, so the ranking is not an artefact of sampling. The one genuinely close call is Humour
(+33.1) against CSR (+29.4), and those two behave very
differently once fatigue is introduced in §10 — which is the first sign that net
sentiment alone is not a sufficient basis for a content decision.
What each pillar contributes to the headline number
A pillar's effect on the brand is its score multiplied by its share of voice. Humour scores
well but is published often; CSR scores respectably but barely appears. The waterfall below
resolves the two into a single volume-weighted contribution, and it changes the ranking
materially.
Figure 5 — Volume-weighted contribution to corpus net sentiment. Each bar is pillar NSS weighted by that pillar's share of the corpus; bars sum to the corpus score.
Read this way, Price & Promotion is not merely the weakest pillar — it is
an active subtraction, removing
1.3 points from the headline figure while
consuming more publishing effort than any other pillar. Humour, by contrast, contributes
5.6 points, more than Origin's
5.5, purely on volume. Whether that is a good trade
depends entirely on fatigue, which is where §10 goes.
0 9
The Trust Equation
How the Brand Trust Index is built, and what it says that net sentiment does not.
Positive sentiment and trust are not the same thing. A comment can be warm without being a
commitment. The Brand Trust Index isolates the part of positive sentiment that carries
reputational weight: a willingness to recommend, an affirmation that a claim is true,
and the absence of suspicion about motive. Each is a lexicon-flagged binary at comment level,
requiring a matched pattern rather than a keyword hit.
Signal
Pattern class detected
Corpus rate
Advocacy A
Recommendation directed at a third party, tagging of another user with endorsement,
declared repeat purchase, defence of the brand against another commenter.
16.9%
Credibility C
Affirmation of a specific brand claim from personal experience, corroboration of a stated
origin or process, correction of another user's misinformation in the brand's favour.
14.6%
Doubt D
Questioning of motive, accusation of exaggeration or greenwashing, "paid promotion"
attribution, disbelief of a stated claim.
where z(x) = (x − μx) /
σx standardises each signal rate across the six pillars, and
A, C, D are the advocacy, credibility and doubt rates of pillar p.
Weights are set so that acting on the brand's behalf (advocacy) counts for more than merely
believing it (credibility), and suspicion subtracts. The linear transform fixes the cross-pillar
mean at 50 with a standard deviation of 12.5, so the index is read as a
relative position on a 0–100 scale, not an absolute percentage.
Range observed: 36.4
(Price & Promotion) to 70.0
(Origin & Provenance) — a spread of
33.6 index points.
Pillar
Advocacy
Credibility
Doubt
BTI
Origin & Provenance
35.5%
39.6%
2.9%
70.0
Recipe & Usage
30.2%
23.9%
3.1%
61.3
CSR & Sustainability
16.3%
18.3%
18.9%
46.8
Humour & Trend-jacking
12.9%
7.6%
8.2%
44.8
Influencer Collaboration
13.9%
10.0%
24.4%
40.6
Price & Promotion
7.4%
6.5%
23.3%
36.4
The Origin pillar's advantage is built on credibility more than on warmth: its credibility
rate of 39.6% is 6.1 times Price's,
and its doubt rate of 2.9% is the lowest in the set. People do not simply
like provenance content — they believe it, and belief is what converts into
advocacy.
CSR & Sustainability is the instructive case. Its net sentiment is respectable at
+29.4, but it carries a doubt rate of 18.9% — the
second-highest in the corpus and higher than Price's. Sustainability messaging is being received
as a claim to be checked rather than a fact to be accepted, which caps its BTI at
46.8 despite the friendly surface tone.
Figure 6 — Trust profile, four pillars. Each axis normalised to the corpus range; larger area indicates a stronger trust profile.
1 0
The Fatigue Equation
Fatigue measured as a decay constant, not as a feeling.
Fatigue is the one thing in this brief that cannot be read from what people say, because by
the time an audience is tired of something it has mostly stopped commenting on it at all.
It has to be measured from what people stop doing. Ordering each pillar's posts by
publication sequence and observing engagement per exposure gives a decay curve, and the shape of
that curve is the fatigue signature.
EXPONENTIAL DECAY MODEL, FITTED IN LOG SPACE
E(r) = E0e−λr
⇒ ln E(r) = ln E0 −
λr + ε
E(r) = engagement on the r-th exposure of the pillar
· E0 = fitted first-exposure engagement
· λ = decay constant, estimated by ordinary least squares on the log
series. A larger λ means the audience tires faster.
r½ = ln 2 / λ
· BFIdecay = 100 (1 −
e−λk)
r½ is the fatigue half-life: the
number of exposures at which the pillar returns half the engagement of its first outing.
k is the number of posts actually published in the window, so BFIdecay reads as
the percentage of first-exposure response already surrendered by the end of the window.
Final index: BFI = 0.55 · BFIdecay + 0.45 ·
Sat̄, where Sat̄ is the pillar's explicit saturation-language rate
rescaled across the six pillars. Behavioural evidence is weighted above stated evidence because
people tire before they say so.
Figure 7 — Engagement decay against exposure index. Points are observed engagement per post, indexed to the first exposure; curves are the fitted exponential.
Pillar
Posts k
λ
R²
Half-life r½
Saturation language
BFI
Price & Promotion
61
0.0210
0.95
33
19.9%
80.8
Influencer Collaboration
44
0.0213
0.91
32
21.7%
78.5
Humour & Trend-jacking
37
0.0172
0.78
40
12.4%
50.5
CSR & Sustainability
11
0.0101
0.07
68
5.7%
15.6
Recipe & Usage
23
0.0060
0.22
116
2.5%
9.7
Origin & Provenance
14
0.0031
0.07
220
1.2%
2.4
THE DECISIVE COMPARISON
Price & Promotion has a fatigue half-life of 33
exposures and was published 61 times in twelve
weeks. It crossed its own half-life around week 6
and spent the rest of the window in diminishing returns. Origin & Provenance has a half-life
of 220 exposures and was published
14 times — it is operating at roughly
6% of its fatigue budget. The brand is
over-spending its worst asset and under-spending its best one.
Model fit differs sharply between the two groups, and that difference is itself a finding.
The fatiguing pillars fit the exponential closely (R² between
0.78 and
0.95), meaning their decline is
systematic and predictable from exposure count alone. The durable pillars fit poorly
(R² = 0.07 for Origin) — not because the model is wrong but
because there is almost no decay to explain; their variation is driven by the
individual quality of each post rather than by accumulated exposure.
1 1
The Trust–Fatigue Map
Plotting the two indices against each other resolves the whole content portfolio into four behaviours.
Figure 8 — Brand Trust Index against Brand Fatigue Index. Dashed lines are the median of each index across the six pillars. Bubble area scales with comment volume.
Across the six pillars the two indices correlate at r =
-0.87 (p = 0.023). That is a strong inverse
relationship on a small number of units, and it should be read as a description of this
portfolio rather than as a universal law — but within this account, the pillars
that build trust are the pillars that resist fatigue, and they are the ones published
least.
Quadrant
Pillars
Behaviour and correct response
COMPOUND
high trust · low fatigue
Origin & Provenance Recipe & Usage
Returns rise, or at worst hold, with repetition, because each instance carries new
information rather than a repeated ask. These are the only pillars where increasing frequency is
safe. Increase volume; this is where the unspent capacity is.
BURNING BRIGHT
high trust · high fatigue
— none currently —
Content that is believed but wearing out. The correct response is rotation and format
variation rather than reduction. That this quadrant is empty tells you the brand has no
high-trust asset currently being over-used.
LOW LEVERAGE
low trust · low fatigue
CSR & Sustainability
Not damaging, but not earning. CSR sits here because it is published rarely
(11 posts) and met with doubt (18.9%). It does not need less
volume — it needs evidence, which would move it toward Origin rather than
toward Price.
CORROSIVE
low trust · high fatigue
Price & Promotion Influencer Collaboration Humour, at the boundary
Each additional exposure lowers both response and regard. This is the only quadrant where
publishing less improves the outcome on both axes simultaneously.
Cap frequency and rebuild the offer.
Why Humour sits on the line
Humour & Trend-jacking returns a solid +33.1 net sentiment and the
highest engagement per post in the portfolio (13,644), which is why it is
published 37 times. But its trust index is
44.8 — below the portfolio median — on an advocacy rate of
12.9% and a credibility rate of just 7.6%, the lowest in
the set. People enjoy it and it moves nobody. With a fatigue half-life of
40 exposures against 37 posts published, it is now
approaching the corrosive boundary. Humour is a reach instrument, not a trust
instrument, and it is currently being spent as though it were both.
1 2
Statistical Significance
Whether the differences in this report survive testing.
Is the pillar effect real?
A chi-square test of independence was run on the 6 × 3 contingency
table of pillar against sentiment class across all 96,412 comments.
TEST OF INDEPENDENCE
χ² = Σ (Oij − Eij)²
/ Eij ·
V = √χ² /
[n(m−1)]
χ²(10) = 6,926.8,
p < 0.0001
Cramér's V = 0.190 — a small-to-moderate but unambiguous association.
By comparison, platform against sentiment: χ²(4) =
1,182.5, V = 0.078.
At this sample size almost any difference reaches significance, so the p-value is the least
interesting output here. The effect size is the finding: V =
0.19 for pillar against V = 0.08 for platform
— what you post about is roughly 2.4 times as
consequential as where you post it.
Negative-sentiment odds by pillar
Pillar
Neg rate
OR
95% CI
Price & Promotion
35.7%
4.30
4.06 – 4.55
Influencer Collaboration
32.3%
3.69
3.48 – 3.91
CSR & Sustainability
19.1%
1.82
1.69 – 1.96
Humour & Trend-jacking
18.9%
1.80
1.69 – 1.92
Recipe & Usage
11.4%
1.00
reference
Origin & Provenance
9.9%
0.85
0.78 – 0.92
Odds ratios computed against Recipe & Usage as
reference, with Woolf standard errors on the log scale.
Figure 9 — Odds of a negative comment, by pillar. Points are odds ratios against the Recipe & Usage baseline; bars are 95% confidence intervals on a log scale.
A comment on a Price & Promotion post carries 4.30 times
the odds of being negative compared with a comment on a Recipe post, and an Influencer post
3.69 times. Neither interval comes close to crossing 1.0. Origin &
Provenance is the only pillar with odds significantly below the baseline
(0.85, CI 0.78–0.92).
Is the decline real?
ORDINARY LEAST SQUARES ON WEEKLY NET SENTIMENT
NSSt = β0 + β1t +
εt, t = 1 … 12
β1 = -1.208 NSS points per week
(SE 0.063; 95% CI -1.332 to -1.084)
R² = 0.973 · p < 0.0001 · total
modelled movement across the window: -13.3 points
An R² of 0.97 on a twelve-point weekly series means the decline is close to
linear and is not being produced by one bad week. The confidence interval excludes zero by a wide
margin. This is a trend, not a fluctuation.
Cadence against outcome
Across the six pillars, Spearman rank correlation between the number of posts published and
the resulting fatigue index is ρ = 0.83
(p = 0.042) — significant even at n = 6. The
equivalent correlation between posting cadence and net sentiment is
ρ = -0.66 (p = 0.156), directionally
consistent but not significant at this number of units. The honest reading:
frequency demonstrably predicts fatigue; its effect on sentiment is suggested by these
data but not established by them, and confirming it requires either more pillars or a
post-level model.
1 3
Negative Driver Decomposition
What the negative quarter of the corpus is actually complaining about.
24.3% of the corpus is negative — 23,449 comments. Left
as a single number that is unactionable. Clustered by aspect terms extracted at stage 5 of the
pipeline, it resolves into ten recurring themes accounting for
85.7% of all negative volume; the remaining
14.3% sits in a long tail of low-frequency, largely
one-off complaints.
Figure 10 — Top ten negative themes as a share of all negative comments. Colour indicates the pillar each theme predominantly attaches to.
Negative theme
Primary pillar
Share of negative
Est. comments
Price increase vs. pack size
Price
15.8%
3,704
Promo code / offer not honoured
Price
11.3%
2,649
Influencer perceived as inauthentic
Influencer
10.8%
2,532
Repetition — 'same ad again'
Influencer
10.1%
2,368
Sustainability claim doubted
CSR
8.3%
1,946
Availability / out of stock
Price
7.1%
1,664
Packaging waste
CSR
6.2%
1,453
Taste / quality inconsistency
Recipe
5.9%
1,383
Delivery & courier complaints
Price
5.2%
1,219
Comment moderation / no reply
Influencer
5.0%
1,172
Two structural observations. First, the top two themes are both
price-integrity issues rather than price-level issues — the objection is
to a perceived change in value ("same price, smaller pack") and to promotions that did not
work as advertised, not to the price itself. That is an operational fix, not a positioning one.
Second, the third and fourth themes together
(20.9% of negative volume) are about the advertising
rather than the product. Roughly a quarter of all negative sentiment in this corpus is
generated by the marketing, not by the thing being marketed.
Representative comment patterns
NOTE ON THESE EXAMPLES
The passages below are paraphrased composites constructed to represent the
dominant pattern within each cluster. No individual comment is reproduced and no author is
identifiable, in line with the compliance position in §03.
Pack got smaller but the price stayed where it was — did you think
nobody would weigh it?
Tried this with the extra cardamom the way you showed — it worked.
Sending this to my sister.
RECIPE & USAGE · POSITIVE · ADVOCACY FLAG SET
Everyone says sustainable. Where is the certificate? Show the paperwork.
CSR & SUSTAINABILITY · NEGATIVE · DOUBT FLAG SET
The asymmetry between the two columns is the practical content of this entire report.
Positive comments in the trust-building pillars contain specifics — a named
place, a named technique, an action taken. Negative comments in the corrosive pillars contain
demands for specifics. The audience is asking for evidence in both directions, and it
rewards the pillars that supply it.
1 4
Temporal & Cadence Effects
When sentiment moves, and what moves it.
Figure 11 — Weekly net sentiment by platform, with fitted corpus trend. Gold band is the 95% confidence region on the fitted slope.
All three platforms decline across the window, but not at the same rate, and the divergence
is informative. Instagram starts highest and stays highest; TikTok loses the most ground; Facebook
declines from the lowest base. Because the pillar mix was broadly constant across platforms, a
common downward slope across all three points to a portfolio-level cause — accumulated
exposure — rather than to anything platform-specific.
Which pillars are driving the decline
Pillar
Weekly slope (NSS pts / week)
p-value
Modelled change across window
Reading
Origin & Provenance
+0.44
0.025
+4.9
Improving with exposure
Recipe & Usage
+0.07
0.725
+0.8
Stable
CSR & Sustainability
-0.53
0.049
-5.9
Stable
Price & Promotion
-1.72
< 0.001
-19.0
Eroding
Humour & Trend-jacking
-1.80
< 0.001
-19.8
Eroding
Influencer Collaboration
-1.99
< 0.001
-21.9
Eroding
Three pillars — Price, Influencer and Humour — account for essentially all of the
downward movement, at between -1.99
and -1.72 points per week each.
Origin & Provenance is the only pillar with a positive slope
(+0.44/week): it is the sole content type in this portfolio that the
audience likes more the longer the campaign runs. That is the empirical definition of a
compounding asset, and it is presently receiving
7.4% of publishing effort.
Hour-of-day risk
Negative share varies by posting hour across a
2.9-point band, peaking at
15:00 (25.4% negative) and bottoming at
10:00 (22.5%). The evening block from
18:00 to 22:00 carries the highest comment volume and above-average negativity together, which
means the brand's highest-reach window is also its highest-risk one.
This is a moderation-resourcing finding more than a scheduling one. The band is too narrow to
justify moving publishing times, but it is wide enough to justify concentrating response
capacity in the evening block, where an unanswered complaint is seen by the most people.
Figure 12 — Negative comment share by hour. Bar length and colour both encode negative share; 24-hour clock, local time.
1 5
Cross-Platform Comparative Matrix
Where each pillar performs, and how effort compares with return.
Figure 13 — Net sentiment, pillar × platform. Each cell is the NSS of that pillar on that platform.
The matrix is remarkably consistent by row and inconsistent by column: pillar rank order
holds on all three platforms, while the magnitude shifts. Origin & Provenance is the
strongest pillar everywhere; Price & Promotion is the weakest everywhere. No pillar reverses
its sign between platforms except Price, which is marginally positive on Instagram and clearly
negative on the other two.
That consistency is the practical justification for planning content by pillar first and
platform second. The platform decision changes how much a pillar earns; it does not change
whether it earns.
The one genuine platform-specific effect worth acting on: Instagram absorbs Price content
-5-to-7 better than TikTok does. If discount
messaging must run at current volume, Instagram is where it does least damage.
Effort against return
Figure 14 — Share of publishing effort against share of positive sentiment. Left axis is each pillar's share of the 190 posts published; right axis is its share of all positive comments.
The crossing lines are the report's clearest single image. Price & Promotion
takes 32.1% of publishing effort and returns
18.0% of positive sentiment
— the only pillar in the portfolio whose return share falls materially below its effort
share. Origin & Provenance inverts it, taking
7.4% of effort for
14.1% of return. Recipe & Usage
shows the same inversion at larger scale.
Pillar
Effort share (posts)
Return share (positive comments)
Return / effort
Positive comments per post
Origin & Provenance
7.4%
14.1%
1.92×
450
CSR & Sustainability
5.8%
9.1%
1.58×
370
Recipe & Usage
12.1%
18.8%
1.55×
363
Humour & Trend-jacking
19.5%
18.9%
0.97×
228
Influencer Collaboration
23.2%
21.1%
0.91×
213
Price & Promotion
32.1%
18.0%
0.56×
131
1 6
Insight Synthesis
The mechanism that connects content type to trust and to fatigue.
Six pillars, three platforms and twelve weeks of data resolve into one mechanism, and it is
simpler than the volume of evidence behind it suggests. Content that gives the audience
something earns trust and resists fatigue. Content that asks the audience for something spends
trust and fatigues fast. Every index in this report is a different measurement of that
one distinction.
THE THREE LAWS OBSERVED IN THIS CORPUS
1. Evidence compounds; assertion depletes. Origin & Provenance shows a
verifiable thing — a place, a process, a person — and returns a credibility rate of
39.6% with a doubt rate of 2.9%. CSR &
Sustainability asserts a virtue without showing it, and returns a doubt rate of
18.9% on friendly-sounding content. The two pillars are adjacent in tone and
23.2 index points apart in trust.
2. Fatigue is priced in exposures, not in weeks. The decay model fits
against exposure index, not against time. A pillar published fourteen times in twelve weeks and
one published sixty-one times are not running the same campaign at different speeds — the
second has spent four times the audience's tolerance.
3. The ask is what tires people, not the format. Humour and Price share no
tone, no format and no production cost, and their fatigue constants sit within
0.004 of each other. What they share is that neither
leaves the viewer with anything after the scroll.
Answering the three commissioned questions
Q
Answer as supported by this corpus
Q1
The brand sits at +21.9 net sentiment,
strongest on Instagram (+31.4) and weakest on Facebook
(+15.5), but declining at 1.21 points per week on all
three. Platform explains 0.08 of the variation against
0.19 for content type.
Q2
Origin & Provenance (BTI 70.0)
and Recipe & Usage (BTI 61.3) build trust. Both are
proof-led: they show a verifiable origin or teach a repeatable technique. Both are published
below a fifth of the total. Neither shows meaningful fatigue at current volume.
Q3
Price & Promotion (BFI 80.8)
and Influencer Collaboration (BFI 78.5) cause fatigue,
at half-lives of 33 and 32 exposures
against 61 and 44 posts actually published. Humour is
approaching the same threshold. Together these three pillars carry
74.7%
of publishing effort.
1 7
Recommendations & Projected Impact
What to change, in what order, and the modelled outcome of doing it.
The reallocation
The central recommendation requires no increase in output, no additional budget and no new
capability. It moves 30% of the comment volume currently generated by
Price & Promotion and Influencer Collaboration into Origin & Provenance and
Recipe & Usage, by shifting the posting mix that produces it.
NSS +21.9 → +28.6
· gain of 6.7 points at constant total
output.
Two cautions on this figure. It assumes reallocated volume performs at the recipient pillars'
current rates, which holds only while those pillars stay inside their fatigue budget —
Origin's half-life of 220 exposures gives ample headroom, but the
assumption should be re-tested at the next measurement window. And it models sentiment only; the
commercial effect of reducing promotional frequency on short-term sales volume is outside the
scope of this corpus and must be assessed separately.
Figure 15 — Modelled corpus sentiment. Current mix against the rebalanced mix at constant total output.
Ninety-day action plan
Phase
Action
Rationale drawn from this report
Measure of success
Days 1–30
Cap Price & Promotion at 12 posts per quarter; hold the offer, cut the
frequency.
Half-life of 33 exposures against 61 published
(§10). The pillar spent the back half of the window in diminishing returns.
Price-pillar NSS above 0; saturation flag rate below 15%.
Days 1–30
Fix the two price-integrity issues before publishing about price again.
Themes 1 and 2 are 27.1% of all negative volume and are
operational, not perceptual (§13).
Combined share of those two themes below 20%.
Days 15–60
Triple Origin & Provenance output to roughly 42 posts per quarter.
Highest BTI (70.0), lowest fatigue (BFI 2.4), only
positive weekly slope (+0.44/week) in the portfolio (§08, §14).
Origin share of positive comments above 20%.
Days 15–60
Re-brief CSR content around documentary evidence — certificates, audited
figures, named suppliers.
Doubt rate of 18.9% on otherwise friendly content; the audience is asking
for paperwork (§09, §13).
CSR doubt rate below 12%; BTI above 55.
Days 30–90
Restructure influencer work around fewer, longer, evidence-based partnerships.
Doubt rate 24.4%, BFI 78.5; the dominant
negative theme is inauthenticity, which volume makes worse rather than better (§13).
Influencer NSS above +25; doubt rate below 15%.
Days 30–90
Cap Humour at current volume and rotate formats; do not scale it.
Highest engagement per post (13,644) but credibility rate of
7.6% and a weekly slope of -1.80 (§11).
Humour BFI held below 55.
Ongoing
Concentrate moderation capacity in the 18:00–22:00 block.
Highest-volume window coincides with above-average negative share, peaking at
15:00 (§14).
Median first-response time under 4 hours in-block.
RE-MEASUREMENT
Every metric above is reproducible from the delivered pipeline. A twelve-week re-run against
the same pillar taxonomy gives a directly comparable set of figures; the fatigue constants in
particular need at least eight exposures per pillar to re-estimate reliably, which the capped
schedule above still provides.
1 8
Limitations, Bias & Risk Statement
The boundaries of what this corpus can support.
Limitation
Effect on the findings
Mitigation applied
Commenters are not customers
A comment corpus reflects the vocal minority of an audience. Sentiment here measures public
expression, not private preference, and it cannot be read as a purchase-intent proxy.
All conclusions are framed as content-performance findings; no sales inference is drawn.
Pillar assignment is at post level
A comment on a Price post that is actually about taste inherits the Price tag. This inflates
within-pillar heterogeneity.
Aspect terms are extracted independently at comment level and reported separately in
§13.
Sarcasm skews positive
The residual classifier error is asymmetric: ironic praise is more often read as positive
than the reverse, so the true negative share is likely marginally understated.
Quantified in §06; the direction of bias is stated so findings are read conservatively.
Six pillars, n = 6 for cross-pillar tests
Correlations between pillar-level indices (BTI–BFI, cadence–fatigue) rest on six
data points and are fragile to the addition or removal of any one pillar.
Reported with exact p-values and explicitly flagged as descriptive of this portfolio in
§11 and §12.
Platform algorithms are unobserved
Engagement decay could partly reflect declining algorithmic distribution rather than audience
fatigue. The two are not separable from comment data alone.
The behavioural decay component is blended with the stated-saturation lexicon rate, which is
algorithm-independent, at a 55/45 weight (§10).
Deleted and hidden comments
Comments removed by moderation before a pull are absent. If removal targets negative content,
observed sentiment is optimistic.
Six-hour pull cadence limits the window for undetected removal; completeness receipts
reconcile counts per pull (§03).
Single brand, single window
No category benchmark is available, so "good" and "bad" are internal comparisons across
pillars rather than external ones.
All rankings are stated relative to the corpus mean, never against an implied industry
standard.
STATISTICAL DISCLOSURE
No result in this report has been selected on the basis of significance. All six pillars,
three platforms and ten themes defined before analysis are reported whether or not they reached
significance, including the non-significant cadence–sentiment correlation in §12.
Confidence intervals accompany every proportion. The full comment-level table is delivered with
this report so that every figure can be independently recomputed.
1 9
Methodology Appendix
Formula glossary and variable dictionary.
Formula glossary
Quantity
Expression
Notes
Net Sentiment Score
(n+ − n−) / n × 100
Neutrals retained in the denominator.
Engagement weight
wi = 1 + ln(1 + Li)
Log weight limits the influence of a single viral comment.
Preferred over the normal approximation at extreme proportions.
Margin of error
z√(p(1−p)/n)
Stated at p = 0.5, z = 1.96.
Brand Trust Index
50 + 12.5[0.40z(A) + 0.35z(C) − 0.25z(D)]
Standardised across pillars; mean 50, SD 12.5.
Decay constant
ln E(r) = ln E0
− λr
OLS in log space over exposure index.
Fatigue half-life
r½ = ln 2 / λ
Exposures to halve first-exposure engagement.
Brand Fatigue Index
0.55 · 100(1 − e−λk) + 0.45 ·
Sat̄
Behavioural and stated components.
Chi-square
Σ(O − E)² / E
Test of independence, pillar × sentiment.
Cramér's V
√(χ² / [n(m−1)])
Effect size, m = smaller table dimension.
Odds ratio
[p/(1−p)] / [p0/(1−p0)]
Woolf SE on the log scale for intervals.
Cohen's kappa
(po − pe) /
(1 − pe)
Agreement corrected for chance.
Trend model
NSSt = β0 +
β1t + ε
OLS on the weekly series, t = 1…12.
Variable dictionary — delivered comment table
Field
Type
Definition
pillar
categorical (6)
Editorial purpose of the parent post.
platform
categorical (3)
TikTok, Instagram or Facebook.
week
integer 1–12
Collection week of the comment timestamp.
hour
integer 0–23
Local hour of the comment timestamp.
sent
−1 / 0 / +1
Final sentiment class after lexicon override.
likes
integer
Likes or reactions on the comment at last pull.
advocacy
binary
Recommendation or defence pattern matched.
credible
binary
Affirmation-of-claim pattern matched.
doubt
binary
Suspicion-of-motive pattern matched.
saturation
binary
Explicit repetition or tiredness language matched.
w
float
Engagement weight, 1 + ln(1 + likes).
2 0
Deliverables, Commercial Note & Authorisation
What is handed over, and confirmation to proceed.
Delivered with this report
Item
Contents
01 Structured database
Comment-level table, 96,412 rows × 11 fields, in XLSX and CSV, carrying the pillar,
platform, week, hour, sentiment class, engagement count and all four trust and fatigue flags for
every record.
02 Analysis script
Documented Python that regenerates every figure, coefficient and confidence interval in this
document directly from the delivered table — no manual step sits between the data and the
page.
03 Figure repository
All 16 charts as vector SVG, named to their figure numbers, for reuse in decks and board
papers at any size without loss of quality.
04 Digital catalog
This report as a presentation-ready PDF, structured for direct circulation to the client's
own stakeholders.
Continuation options
Option
Scope
Live re-run
The same pipeline against the client's authorised API credentials, producing a directly
comparable issue of this report on observed data.
Quarterly tracking
Rolling twelve-week windows with the fatigue constants re-estimated each cycle and every
movement flagged against this baseline.
Competitive extension
The same pillar taxonomy applied to two or three competitor accounts, giving external
benchmarks for the internal rankings in §08.
Category extension
Pillar taxonomy extended to additional product lines within the portfolio, on the volume
terms set out in the governing commercial proposal.
Commercial note
This report is issued under the terms of the Data Tune commercial proposal governing the
engagement. Volume is measured in data entities on the same basis as that document; the corpus
delivered here is billed at the agreed entity rate for the base package, with any extension
beyond the contracted ceiling proceeding only on prior written approval, and never
retrospectively.
Scheduled updates
Progress reporting during the working window follows the same three-point structure as the
governing proposal: an early-stage note with initial counts and a first sample batch, a
mid-project note with running totals and a second batch, and a pre-delivery note with near-final
counts before handover.
Acceptance & authorisation
By signing below, the client confirms receipt of this report and its accompanying deliverable
set, accepts the scope and method recorded in §02 to §06, and acknowledges the data
provenance statement recorded in §00.
FOR THE CLIENT — AUTHORISED SIGNATURE
NAME & DESIGNATION · DATE
FOR DATA TUNE — AUTHORISED SIGNATURE
K. H. Militha Mihiranga
DATA SOLUTIONS CONSULTANT · DATE
Data Tune.
DATA SOLUTIONS & DIGITAL CATALOGING
For clarification on scope, data
parameters, index construction or scheduling, please contact the consultant named alongside.
Prepared by
K. H. Militha Mihiranga · Data Solutions Consultant
Office
555/24 Elhenawatta, Ranmuthugala, Kadawatha, Sri Lanka
