Attached Research · 2026 Edition

What 11,793 people working on their attachment reveal about whether it can change

An original look at 51,676 journaling-based measurements from a help-seeking community — where attachment anxiety tends to shift over time, while avoidance mostly holds.

By Gabriel Uribe, Luis Goicouria, and Sophie Cheung · June 24, 2026

Same points, eased by the measured average change.

The attachment map — the real aggregate distribution of all 11,793 members (binned cells of 5+ users, not individuals). Pressing the button slides every point by the measured average anxiety change among consistent journalers. Observational, not causal.

11,793
people analyzed
51,676
attachment measurements
90%
score high on anxiety
38%
reliably improved

01 / Overview

Most attachment statistics come from small surveys answered once. We looked instead at how patterns show up across tens of thousands of real journaling sessions — and how they shift within the same person over months. Members journal about twice a week (for those who stick with it), and also take lessons, meditate, and work through a near-daily set of guided activities; this analysis scores the journaling.

One caveat shapes everything below: these are people who came to Attached because they're working on insecure attachment. This is a portrait of a help-seeking community, not the general public — so read it that way.

Key findings

  • Anxiety dominates. Average anxiety is 0.74/1.0; ~9 in 10 score above the midpoint.
  • Security is rare here — between 3% and 14% depending on the cutoff, the mirror image of the general population.
  • Patterns soften with practice. Anxiety fell ~0.15 on average among consistent journalers; 38% showed a reliable drop. (Observational.)
  • The two dimensions behave distinctly — anxiety tracks daily mood; avoidance doesn't — as theory predicts.

02 / The two dimensions

A community running high on anxiety

Attachment lives on two dimensions: anxiety (fear of abandonment) and avoidance (discomfort with closeness). Here, anxiety is near-universal while avoidance varies widely.

How we measure it: each journal entry is read by an AI rater (Google Gemini 3 Flash) prompted to score attachment on the established ECR-R scale (1–7), averaged over three passes and rescaled to 0–1 — so 0.5 is the scale's midpoint. Full method in the whitepaper.

Attachment anxiety90.4% above midpoint
Attachment avoidance43.1% above midpoint

Share of users above the 0.5 midpoint (n = 11,793).

03 / The four styles

The labels depend on where you draw the line

Cross the two dimensions and you get the familiar styles — but the cutoff matters enormously. Drag the slider and watch the picture change. That fragility is why we treat the continuous scores as the real measure and the labels as a rough map.

Cutoff0.50 (standard)
Anxious51.3%
Fearful-avoidant39.1%
Secure5.6%
Avoidant4.0%

Current pattern at the chosen cutoff (n = 11,793). Values at 0.40 / 0.50 / 0.60 are measured; in-between points are interpolated.

“Attachment anxiety tends to shift over weeks of practice. Avoidance mostly holds.”
The central finding

04 / Can it change?

Can attachment actually change?

Observational finding

Averaging each person's first six vs. last six measurements (single entries are too noisy alone), attachment anxiety drops — and the more someone practices, the larger the drop.

2–4 journaling sessions−0.070
5–9 journaling sessions−0.144
10–19 journaling sessions−0.189
20–39 journaling sessions−0.255
40+ journaling sessions−0.281

Average decrease in anxiety by amount of practice. Effect sizes (Cohen's d) reach 0.82.

−0.15
average drop in anxiety (95% CI −0.166 to −0.139)
38%
showed a statistically reliable decrease
21%
moved from an insecure to a secure pattern
Why we won't call this “the app lowers anxiety.” There's no control group, heavier users started more anxious, and people often journal most during hard stretches and recover naturally. The pattern is real and consistent — but it's an association, not proof of cause.

05 / Mood by pattern

How each pattern feels, day to day

Members also log how they feel on a simple mood grid (pleasant vs. unpleasant, high vs. low energy). People whose current pattern is secure report pleasant moods about 69% of the time — versus roughly 55% for anxious and fearful-avoidant patterns — and report high-arousal distress the least.

Secure69% pleasant
Avoidant59% pleasant
Fearful-avoidant56% pleasant
Anxious55% pleasant
Calm / contentEnergizedDown / depletedAnxious / agitated

Self-reported mood check-ins, by current journaling-derived pattern; cross-sectional association (not causal). The avoidant group is small (~100 users).

06 / Methods & trust

How much should you trust this?

The detail behind every figure — expand what you want to dig into.

What people say vs. what shows up over time+

At sign-up, people describe themselves as even more anxious-leaning than their behavior later suggests, and almost no one calls themselves secure. Self-image and lived pattern line up only moderately (r ≈ 0.36).

PatternSelf-describedJournaling-derived
Anxious64%51%
Fearful-avoidant12%39%
Secure2%6%
Avoidant5%4%

Intake also included ~18% “not sure” (omitted above).

Attachment by relationship situation+

People in a breakup show the highest anxiety; single users are the only group where fearful-avoidant outnumbers anxious. Cross-sectional associations, not causes.

SituationAnxiousFearful-av.SecureAvoidant
Breakup55%35%7%3%
Long-term52%37%6%5%
Dating48%45%4%3%
Single38%48%7%7%
How much should you trust these numbers?+

Reliability. A single measurement is noisy (r ≈ 0.29), so we average: six entries reach ~0.71 reliability, ten reach ~0.80. Every change figure uses averaged measurements.

Validity. The anxiety signal tracks daily mood distress (r ≈ 0.22–0.25) while avoidance doesn't — the expected pattern. But mood isn't attachment; the definitive test (an ECR-RS questionnaire measured concurrently) is on our roadmap, and until then we don't claim this is a validated clinical instrument.

Who it represents. A help-seeking, mostly English-speaking, mostly adult community — not a random population sample. What we don't claim: no population prevalence rates, no diagnoses, no causal effect, no individual-level verdicts.

Methodology, in brief+

Each journal entry is scored by an AI rater (Google Gemini 3 Flash) on the ECR-R attachment scale (1–7), run three times and averaged, then rescaled to 0–1 (so 0.5 is the ECR-R midpoint). A separate intake assessment provides each person's starting point. “Current pattern” is a 30-day half-life average of someone's journaling scores; categories apply a 0.5 midpoint reported with a 0.4–0.6 sensitivity analysis. Change analyses average first vs. last scores (primary six per side; sensitivity three and ten) and flag reliable change with a Reliable Change Index. Self-reported mood and wellbeing check-ins are used only to triangulate, not to compute the score. Scores are AI-generated, not human-rated, and aren't yet validated against a concurrent ECR-RS questionnaire. Confidence intervals are 95%; dose–response reports Cohen's d vs. the least-active group. Subgroup breakdowns require 500+ people, with cells under 30 suppressed. All outputs are aggregate and de-identified. Full methods in the whitepaper.

Work on your own attachment patterns

Attached turns these insights into daily, personalized practice.

For educational purposes; not medical advice. Figures are aggregate and de-identified. © 2026 SkyPorch LLC.