Brox.AI Research
Social Media
Algorithm Mapping
How 409 AI-generated digital twins respond to content across four distinct algorithmic strategies — and what it reveals about platform design, engagement, and the gap between what people say and what they do.
409
Digital Twins
4
Algorithms
5
Content Per Algo
8,180
Total Responses
Study Design
Methodology
409 digital twins — AI-generated personas built from real interview transcripts — were exposed to 5 pieces of content from each of 4 algorithm types. Each twin predicted which reactions they would have to each piece of content.
Algorithm A
Engagement Bait
Emotionally charged, sensationalised content designed to provoke reactions through outrage, fear, or controversy.
- High-Emotion Political Clip
- Outrage-Inducing News
- Fear-Based Health Content
- Celebrity Controversy
- Divisive Opinion Piece
Algorithm B
Credible Information
Factual, verified content from authoritative sources designed to inform rather than inflame.
- Verified News Article
- Expert Analysis Piece
- Investigative Report
- Data-Driven Infographic
- Public Interest Story
Algorithm C
Identity & Community
Content centred on shared identity, community belonging, and cultural pride.
- Community Pride Post
- Local Achievement Story
- Cultural Heritage Feature
- Community Event Promotion
- Grassroots Initiative
Algorithm D
Balanced Perspectives
Multi-perspective, nuanced content presenting different viewpoints on issues without taking sides.
- Both Sides Policy Breakdown
- Comparative Analysis
- Debate Format Feature
- Stakeholder Perspectives
- Nuanced Explainer
Process
Three-Phase Design
Phase 1: All 409 twins respond to all 4 algorithms (20 content pieces total). Scores calculated and turning point determined.
Phase 2: Each twin is assigned to their best-match algorithm. New content is served to validate the assignment.
Phase 3: Algorithm A is re-tested asking what people would say versus what they would do to measure the say-do gap.
Phase 2: Each twin is assigned to their best-match algorithm. New content is served to validate the assignment.
Phase 3: Algorithm A is re-tested asking what people would say versus what they would do to measure the say-do gap.
Framework
Reaction Scoring System
Each reaction type is assigned a weighted score reflecting its value as a signal of genuine engagement (positive) or rejection (negative). Scores are summed across all 5 content pieces per algorithm to create a cumulative score per person per algorithm.
| Reaction | Score | Signal |
|---|---|---|
| Share with Positive Comments | +8 | Strongest advocacy |
| Follow the Poster | +7 | Long-term commitment |
| Share | +6 | Active amplification |
| Share in Direct Message | +5 | Private endorsement |
| Comment | +4 | Active participation |
| Favourite / Save | +4 | Content valued for later |
| Like | +3 | Passive approval |
| Profile Click | +3 | Curiosity about source |
| Click / Watch | +2 | Basic interest |
| Scroll Past / Ignore | -1 | Mild disinterest |
| Hide | -5 | Active suppression |
| Share with Negative Comments | -6 | Hostile amplification |
| Report | -8 | Strongest rejection |
Turning Point
Score = 0
If a person's cumulative score for an algorithm is positive, the platform shows them more of that content. If negative, it hides the content. Zero is the boundary between show and suppress.
Phase 1
Round 1 Results
How all 409 digital twins scored across the four algorithms. Each score is the cumulative total across 5 content pieces.
-23.5
Algo A Mean
+16.6
Algo B Mean
-9.8
Algo C Mean
+14.3
Algo D Mean
Mean Scores by Algorithm
Diverging bars show the mean cumulative score. Left of centre is negative (suppress), right is positive (show).
Algo A
-23.5
Algo B
+16.6
Algo C
-9.8
Algo D
+14.3
Sentiment Split
Percentage of the 409 respondents scoring positive vs negative for each algorithm.
Positive (Score > 0)
Negative (Score < 0)
Neutral (Score = 0)
Algo A
90%
8%
Algo B
25%
72%
Algo C
77%
21%
Algo D
24%
73%
Key Finding
42%
of respondents scored negatively across both A and C algorithms. These individuals reject both engagement bait and community-identity content, but actively embrace credible information and balanced perspectives.
Why does Algorithm A perform so poorly despite being the dominant platform strategy?
Algorithm A — engagement bait — produces a mean score of -23.5, the worst of any algorithm by a wide margin. 90% of respondents scored negatively. This is striking because engagement bait is the dominant strategy on most social media platforms: content is ranked by clicks, watch time, and shares, all of which favour sensationalised material. The disconnect is that platforms measure the guilty click (short-term engagement) but miss the hide and report signals (long-term rejection). When you weight reactions properly — penalising hide (-5) and report (-8) while crediting share (+6) and follow (+7) — the true picture emerges: people overwhelmingly reject this content even though they occasionally click on it.
Assignment
Algorithm Distribution
Each person was assigned to the algorithm where they scored highest — but only if that score was positive. People with no positive scores are classified as “None”.
Click a segment to explore its demographic breakdown
Algorithm B — Credible Information
178 people (43.5%)
Gender
Generation
Algorithm D — Balanced Perspectives
130 people (31.8%)
Gender
Generation
None — All Scores Negative
75 people (18.3%)
Gender
Generation
Algorithm C — Identity & Community
20 people (4.9%)
Gender
Generation
Algorithm A — Engagement Bait
6 people (1.5%)
Gender
Generation
Headline
75% prefer credible or balanced content
Algorithms B and D together capture 308 out of 409 respondents (75.3%). Only 6 people (1.5%) are best-served by engagement bait. The remaining 18.3% reject all four approaches.
Reasoning
Click to explore the analytical reasoning behind these distribution patterns.
Why does Algorithm B dominate the distribution?
Algorithm B — Credible Information — captures 43.5% of all respondents because it activates the broadest set of positive engagement signals across demographics. Credible, fact-based content avoids triggering negative reactions (hide, report) while consistently generating passive-to-moderate engagement (click, like, save). Its 78.7% female skew is notable — women in this study disproportionately prefer verified, authoritative sources. Baby boomers also index high here (23.0% of Algo B vs 19.6% overall), suggesting that older generations particularly value source credibility.
What differentiates Algorithm D people from Algorithm B?
Algorithm D people value nuance and multiple perspectives over pure credibility. The key demographic difference is gender: D has a much higher male share (34.6% male vs B's 16.9%). Men in this study are roughly twice as likely to prefer balanced, debate-style content over straightforward verified news. Generationally, D also skews slightly younger — millennials are 36.9% (vs 30.9% in B), suggesting that younger cohorts prefer content that presents all sides rather than a single authoritative take.
Who are the 75 people assigned to “None” and why?
These 75 people (18.3%) scored negatively across all four algorithms — they didn't engage positively with any content type in the study. Their demographic profile mirrors the overall sample (72% female, 41.3% Gen X), suggesting this isn't a specific demographic rejecting content but rather a cross-cutting segment of digital sceptics. These are likely people with strong avoidance behaviours (high hide/report rates) who reflexively disengage from algorithmically-served content regardless of its nature. For platform designers, this 18.3% represents the audience that may need entirely different engagement approaches, or may simply be unreachable through feed-based content.
Why is Algorithm C small but powerful?
Algorithm C captures just 20 people (4.9%), but these are the most deeply engaged respondents in the entire study. The group skews towards millennials (45%) and has a notable baby boomer presence (30%). Zero Gen Z members appear here. The community-identity content resonates with people who have established social identities and local ties — which explains the millennial/boomer split (established community members) and the absence of Gen Z (still forming their social identities). In Round 2, this group showed 100% retention and 80% score improvement, making them the most loyal audience segment.
Engagement Tiers
Assigned respondents grouped by the strength of their engagement with their best-match algorithm.
High (40+)
106 (31.7%)
Moderate (15–39)
94 (28.1%)
Low (1–14)
66 (19.8%)
Disengaged (0)
9 (2.7%)
Fully Rejected (<0)
59 (17.7%)
Phase 2
Round 2 Validation
New content was served to respondents in their assigned algorithm. 177 out of 334 assigned respondents appeared in Round 2. Did they stay positive?
86.4%
Overall Retention
85%
Algo B Retention
100%
Algo C Retention
86%
Algo D Retention
Score Change: Round 1 vs Round 2
Mean scores before and after distribution. Higher in R2 means the assignment is working.
Round 1
Round 2
Algo B
Algo C
Algo D
Standout
Algorithm C: 100% retention, 80% improved
Every single person assigned to Algorithm C stayed positive in Round 2, and 80% actually increased their score. Community-identity content creates the deepest loyalty among those it resonates with. Mean score jumped from +12.1 to +25.0 — a +12.9 point increase.
Algorithms B & D
Stable performance with slight decline
Both credible information (B) and balanced perspectives (D) showed ~85% retention with modest score declines. This is expected — Round 1 scores were already among the highest possible, so regression to the mean is natural. The high retention rate confirms the distribution was successful.
Deep Dive
The Say-Do Gap
Algorithm A was re-tested on 350 respondents, but instead of predicting what they would do, we predicted what they would say. The gap reveals how self-presentation distorts behavioural data.
-22.2
DO Mean
-2.6
SAY Mean
+19.6
Mean Gap
90.3%
SAY More Positive
Volume Gap
Reaction volume: DO 2.29 vs SAY 0.69 per person per content
When people actually engage with content, they perform far more actions than they report. They multi-react (click, watch, then hide) but when asked, they simplify their behaviour to a single stated response.
Reaction-by-Reaction Comparison
Paired bars showing the DO rate (blue) vs SAY rate (orange) for each reaction type.
DO (Behavioural)
SAY (Stated)
Scroll Past
-2.2pp
Hide
-10.9pp
Click / Watch
-8.0pp
Report
+0.1pp
Comment
+13.4pp
Share in DM
+6.1pp
Like
+0.3pp
Profile Click
-1.7pp
Share w/ Neg
+1.7pp
Share w/ Pos
+1.2pp
Three Key Findings
Finding 1
People under-report avoidance
Hide drops 10.9 percentage points from DO to SAY. People actively suppress content in practice but present themselves as more tolerant when asked. The act of hiding is private, so there is no social cost — but admitting it feels like admitting intolerance.
Finding 2
People over-claim active engagement
Comment jumps 13.4 percentage points from DO to SAY. People say they would voice their opinions, but in practice stay silent. The aspiration to engage is strong, but the friction of actually typing a comment is a barrier that self-reports fail to capture.
Finding 3
People under-report passive curiosity
Click/Watch drops 8.0 percentage points from DO to SAY. People consume far more content than they would admit. The guilty click — watching something you know is engagement bait — is one of the most common behaviours that goes unreported.
Why Does This Happen?
Click to explore the psychological mechanisms behind the say-do gap.
Why do people hide content privately but deny it publicly?
Hiding content is a zero-cost, zero-visibility action. Nobody sees you hide a post. But when asked “would you hide this?”, admitting avoidance feels like admitting intolerance or closed-mindedness. The 10.9pp drop from DO to SAY reflects social desirability bias: people want to appear open-minded, even though their behaviour shows they actively curate their feeds to avoid content they find distasteful. This is why behavioural data is essential — the private act of hiding is invisible in surveys.
Why do people say they’d comment but actually stay silent?
The 13.4pp over-claim on commenting is the largest gap in the study. It reveals the difference between intention and friction. When asked, people genuinely believe they would express their opinion. But in practice, the effort of formulating a public response, the risk of backlash, and the simple laziness of scrolling past all create barriers. The intention is real, but the execution gap is enormous. This is why comment counts are such poor predictors of stated engagement — people overestimate their own willingness to participate publicly by a factor of nearly 8x.
What does the “guilty click” tell us about engagement metrics?
People click and watch engagement bait at 12.3% in practice but only admit to 4.3%. This is the “guilty click” — consuming content you know is low-quality because curiosity overrides judgement. For platform designers, this creates a dangerous feedback loop: click metrics suggest people want engagement bait, but people would never say so. The behavioural data (clicking) and the attitudinal data (hiding) are both real, but they tell different stories. People simultaneously consume and reject the same content. Relying on click-through rates alone massively overestimates genuine interest.
Gap Distribution
How large is the gap between SAY and DO scores? Nearly half of all respondents have a gap exceeding +20 points.
Gap > +20
49.7%
Gap +10 to +20
27.7%
Gap +5 to +10
8.3%
Gap 0 to +5
5.1%
Gap < 0
9.1%
Turning Point Analysis
Do respondents cross the zero threshold (positive vs negative) when comparing their SAY score to their DO score?
64.6%
Both Negative
24.6%
SAY Positive / DO Negative
0.6%
SAY Negative / DO Positive
10.3%
Both Positive
Critical Implication
24.6% overstate their engagement
These respondents say they would engage positively, but their actual behaviour is negative. If you relied on surveys alone, you would incorrectly classify nearly a quarter of your audience as engaged when they are actually rejecting the content. This is the core risk of survey-based content strategy.
Case Study
Individual Journey Map
Follow a single digital twin through every algorithm, every content piece, every reaction — and see how the right algorithm transforms behaviour entirely.
DT
Digital Twin #691f3555
Female · Gen X · Assigned to Algorithm D · Score range: 115 pts
A: -26
B: +79
C: +21
D: +89
Algorithm A
-26
Engagement Bait
Algorithm B
+79
Credible Info
Algorithm C
+21
Community
Algorithm D
+89
Balanced — Winner
Cumulative Score Across All Content
Each dot is one content piece. The line traces how engagement builds or collapses across each algorithm’s feed — and then into the assigned Round 2.
Round 1
Testing All Algorithms
Algorithm A
Engagement Bait
-26
1
High-Emotion Political Clip
click/watch +2
profile click +3
comment +4
hide -5
+4
2
Breaking Crime Post
scroll past -1
hide -5
report -8
-10
3
Celebrity Scandal Thread
scroll past -1
click/watch +2
hide -5
-14
4
Highly Shareable Meme
scroll past -1
hide -5
-20
5
Reaction Video
click/watch +2
share in DM +5
hide -5
report -8
-26
Pattern
Curiosity meets rejection
She clicks and watches — even comments once — but hides every single piece of content. The guilty click in action: she can’t resist looking, but actively suppresses it from her feed. By content 4, she’s reporting as well as hiding.
Algorithm D
Balanced Perspectives — Best Match
+89
1
“Both Sides” Policy Breakdown
click/watch +2
like +3
comment +4
+9
2
Cross-Ideology Dialogue Clip
click/watch +2
like +3
comment +4
share in DM +5
+23
3
Data Visualization Post
click/watch +2
like +3
favourite +4
+32
4
Bridge Influencer Commentary
click/watch +2
like +3
follow +7
comment +4
share w/ pos +8
share in DM +5
favourite +4
+65
5
Reflective Long-Form Post
click/watch +2
like +3
comment +4
share w/ pos +8
follow +7
+89
The Shift
Zero negative reactions. Escalating engagement.
Not a single hide, scroll past, or report across all 5 pieces of balanced content. Every content piece triggers more reactions than the last. By content 4, she’s performing 7 positive actions on a single post. This is what algorithmic match looks like.
Round 2
Assigned to Algorithm D
Algorithm D — New Content
Validation Round
+96
1
“Two Perspectives” Video Split Screen
click/watch +2
like +3
comment +4
+9
2
Community Dialogue Event Clip
click/watch +2
like +3
comment +4
share w/ pos +8
+26
3
Contextualized Incident Breakdown
click/watch +2
like +3
comment +4
share in DM +5
+40
4
Long-Form Reflective Essay
click/watch +2
like +3
comment +4
share w/ pos +8
share in DM +5
+62
5
Data-Driven FAQ Post
click/watch +2
profile click +3
like +3
share +6
share w/ pos +8
share in DM +5
follow +7
+96
Validation Confirmed
Round 2 score (+96) exceeds Round 1 (+89)
On completely new content she has never seen before, her engagement is even higher than Round 1. The final content piece triggers 7 positive actions including share, share with positive comments, share in DM, follow, and profile click. The algorithm assignment didn’t just hold — it deepened.
Say-Do Gap
Algorithm A: What She Would Say
For the High-Emotion Political Clip — the first piece of Algorithm A content — compare what she did versus what she said she would do.
What She Would Do
High-Emotion Political Clip
- +2 Click / Watch
- +3 Profile Click
- +4 Comment
- -5 Hide
Score: +4
What She Would Say
High-Emotion Political Clip
- +4 Comment
Score: +4
The Gap in Action
Same score, completely different behaviour
Both DO and SAY produce a score of +4, but the reality is entirely different. When she does, she performs 4 actions: clicks, checks the profile, comments, then hides. When she says, she claims just one: commenting. She erases the guilty click, the curious profile visit, and crucially the hide.
The Full Picture
From -26 to +96
The same person who reports and hides engagement bait becomes one of the most engaged respondents in the study when served balanced, multi-perspective content. The algorithm doesn’t just matter — it transforms behaviour entirely.
Conclusions
Key Takeaways
Four headline conclusions from the study, each with direct implications for platform design, content strategy, and audience research methodology.
1
Credible and balanced content wins
Algorithms B and D together capture 75% of respondents. The majority of people actively prefer fact-based, multi-perspective content over sensationalism or identity-driven material. Platform strategies built around quality content are not just ethical — they are commercially optimal.
2
Engagement bait fails
Algorithm A was rejected by 90% of respondents. Only 6 out of 409 people preferred engagement bait as their best-match algorithm. The dominant platform strategy of optimising for clicks and outrage is actively alienating the vast majority of users.
3
Community content is niche but powerful
Algorithm C captures few people (4.9%), but those it captures show 100% retention and 80% improvement in Round 2. Community-aligned content creates intense loyalty among a narrow audience — ideal for targeted strategies.
4
The say-do gap is massive
Survey-based content strategy would overestimate engagement by +19.6 points on average. 90.3% of people present more favourably than they behave. 24.6% cross the turning point entirely — claiming engagement while actually rejecting the content.
Bottom Line
Build for trust, not outrage.
The data is unambiguous: people engage most with credible, balanced content. They reject engagement bait overwhelmingly. And they tell you the opposite when you ask them in surveys. The only reliable way to understand audience behaviour is to observe it, not ask about it.