Contextual hints appear inline as you work. A chat drawer answers open questions about your model, all grounded in the numbers you already computed.
One glance per tab
Open Results, Editor, MGA or Advanced and see 1 to 3 suggestions right at the top. Each hint reads your live model and cites your own numbers.
Ask, get grounded answers
Open the robot in the bottom right to ask anything about your model, data, or results. Four read-only tools give the model access to your project structure, never to raw rows.
Hair, Ringle and Sarstedt's Primer on PLS-SEM runs 428 pages. Chin's variance-explained tiers, Henseler's HTMT, Cohen's f², Roemer's HTMT2, Shmueli's out-of-sample rules - every step of a rigorous analysis has a cut-off, a citation, and a caveat about domain variance. Even experienced authors keep the primer open next to their editor and re-check numbers between paragraphs.
The AI Companion doesn't remove that rigor. It removes the lookup tax. It reads your live project, knows the threshold that applies, cites its source, and flags where you might be about to publish something a reviewer will catch.
The last three phases close the loop. Pick the right advanced methods, draft the Method / Results / Discussion paragraphs from your live output, and let every claim trace back to a curated paper the assistant can name.
Reads your model shape, sample size and results, then proposes an ordered pipeline of advanced methods with a one-line rationale each. IPMA before MGA. MICOM before MGA. Skip FIMIX below n = 200. Every suggestion knows the threshold it just crossed.
Method, Results and Discussion paragraphs drafted from your live compute output as LaTeX or Word. Includes a reviewer-simulation pass: what would a skeptical reviewer object to, and which citation answers it.
Answers ground every claim in a vector-indexed library of 50+ PLS-SEM classics. Chin, Henseler, Hair, Sarstedt, Ringle, Shmueli, Sarstedt et al. Each citation chip opens the paper's abstract in the drawer.
Turn a research question into a starter model in seconds. Get a validated questionnaire back from the model's LVs. Collect data, then interpret it with the same assistant.
Type your research question. The assistant proposes 3–6 latent variables with modes (reflective / formative), hypothesised paths with citations, and 2–3 blueprint references from the PLS-SEM literature. One click writes it into a new model doc with LVs and paths pre-drawn.
Given the model, the assistant generates 4–6 validated survey items per LV with source citations (Davis TAM, Zeithaml SERVQUAL, Cronin & Taylor …). Or start from scratch and write items yourself. Edit inline, add demographics, export as CSV / Markdown / JSON for Qualtrics, LimeSurvey or SoSci Survey.
trust1 I feel confident this shop delivers on its promises. trust2 This shop appears reliable in its transactions. trust3 The shop takes reasonable care with my personal data. The Companion isn't a chatbot bolted onto a form. It's woven into the panels you already use, and it reads only what it needs.
While you drag latent variables and paths, the assistant reads the current structure and calls out reflective / formative mode mismatches, LVs with fewer than 3 indicators, and unlinked constructs. It suggests when a higher-order construct might tidy up a tangle of correlated LVs, and cites Sarstedt et al. (2019) on when disjoint two-stage is appropriate.
Sample-size adequacy against the 10-times rule, missing-value audit column by column, sentinel-code detection (−99, −999), scale-range consistency across indicators of the same construct. All from column summaries - never from your data.
R² tiers with named cut-offs (Chin 1998), HTMT violations called out by pair, reliability failures per LV, SRMR fit warning, and the single most useful next method to run given what you already have. This is the highest-leverage step - most manuscripts are lost or won here.
The assistant scans your dataset for categorical columns with 2–5 levels (gender, market, treatment) and proposes an MGA candidate. It enforces the MICOM-before-MGA sequence, warns when a group falls below n = 30, and suggests which structural paths to focus the between-group comparison on.
market has 4 levels (DE, US, UK, JP). Group JP has n = 27 - below the 30-observation floor for stable bootstrap. Consider merging JP with another Asian market, or drop it from the MGA.Which of the 15+ methods actually fits your model? The assistant reads your R²s, sample size, model shape and reliability metrics, then proposes an ordered pipeline. The question "what should I run next" gets a concrete, defensible answer with citations.
Method section, Results section, Discussion of limitations - drafted from your live compute output as LaTeX or Word paragraphs. Includes the reviewer-defense pass: what a skeptical reviewer would ask, and the citation you'd answer with.
The Companion refuses to invent numbers. When it cites a cut-off, the citation is baked into its system prompt and reflects the canonical PLS-SEM literature that OpenPLS validates against.
| Criterion | Cut-off | Source |
|---|---|---|
| R² tiers | 0.19 / 0.33 / 0.67 | Chin (1998) |
| Cohen's f² tiers | 0.02 / 0.15 / 0.35 | Cohen (1988) |
| HTMT discriminant validity | < 0.85 (conservative), < 0.90 (liberal) | Henseler et al. (2015) |
| HTMT2 refinement | geometric-mean correction | Roemer et al. (2021) |
| SRMR model fit | < 0.08 | Hu & Bentler (1999) |
| Cronbach's α, composite reliability | ≥ 0.70 | Nunnally & Bernstein (1994) |
| 10-times rule for sample size | n ≥ 10 × max paths into any LV | Barclay et al. (1995) |
| Bootstrap iterations (publication) | ≥ 5000 | Hair et al. (2022) |
| IPMA importance × performance | total effects × mean latent scores | Ringle & Sarstedt (2016) |
| PLSpredict k-fold panel | RMSE, MAE vs. linear-model benchmark | Shmueli et al. (2019) |
When the assistant is asked about a threshold it does not know for sure, it says so instead of guessing. This behaviour is enforced via its system prompt - a hard rule: "If asked about a threshold you don't know for sure, say so and don't invent a number."
Illustrative composites drawn from OpenPLS user sessions and common PLS-SEM research contexts. None depict real users.
Companion flags Loyalty with only 2 indicators and cites Hair et al. 2019 on minimum item counts. Adds a third validated item from Watanabe et al. 2018.
R² = 0.62 for Loyalty. HTMT Trust ↔ Satisfaction = 0.88 flagged. Suggests HTMT2 refinement - passes cleanly at 0.83. Reviewer defense secured.
Suggests IPMA on Loyalty. Trust ranks highest priority (importance 0.42, performance 61). Ready-made contribution for the Managerial Implications section.
Companion detects R² dropping suspiciously low. Notes −99 values not registered as missing. Recommends adding −99 to the dataset's missing-code list.
R² for Care-Quality → Satisfaction jumps from 0.09 to 0.44 after correct missing-value handling. Result set becomes publishable.
Hospital ID has 4 levels. Companion enforces MICOM first - partial invariance holds. MGA reveals significant Trust → Loyalty differences between two sites.
Companion suggests a formative-reflective HOC for Service-Quality. Cites Sarstedt et al. 2019 disjoint two-stage. Six items reduce to a cleaner two-dimensional construct.
PLSpredict k = 10. Model out-performs linear-model benchmark on RMSE. Explains the difference between in-sample R² and out-of-sample validity to the client.
Exports LaTeX tables and IPMA priorities directly. Report draft ready two days before deadline.
Every PLS-SEM paper is your name, your defense. The Companion is deliberately built to strengthen your judgement, not replace it.
Model shape, column summaries, R², p-values, HTMT matrices, reliability - that's all. Never dataset rows, never uploader identity, never file names. The system-level invariant is enforced in code, not in a prompt.
Every answer shows which of the four read-only tools the model called. If it claims a number, you can verify which tool surfaced it. No black-box interpretation.
The Companion can suggest "run IPMA on Loyalty" - it cannot execute the compute itself. You stay in the driver's seat for every irreversible step.
A hard rule in the system prompt: if uncertain about a threshold or a method's applicability, the assistant says so instead of hallucinating. Refusal is a feature.
Companion sits behind an explicit opt-in with a clear consent surface. Once on, one click also turns it off - no hidden dependencies, no re-onboarding.
100 chat turns and 100 hint calls per user per month during Beta. You always know how many are left. No auto-billing, no surprise cost.
All 7 phases are live. Every step of a PLS-SEM manuscript, from research question to reviewer defense, now has an AI companion built for it.
Answer any open question about your model using four read-only tools. Live since 2026-07-27.
Proactive 1–3 suggestions at the top of every project tab. Live since 2026-07-27.
Turn your research question into a starter model with LVs, modes, paths and blueprint citations. One click writes it into a new model doc with everything pre-drawn. Live since 2026-07-27.
For every LV, generate 4–6 validated items with citations (Davis TAM, Zeithaml SERVQUAL, Cronin & Taylor …). Or start manually and write items yourself. Inline editor, demographic presets, CSV/Markdown/JSON export. Live since 2026-07-27.
Reads the full project (model + data + results) and drafts an ordered advanced-methods pipeline - IPMA before MGA, MICOM before MGA, PLSpredict where out-of-sample matters. Every suggestion cites the threshold it just crossed. Live since 2026-08-03.
Method, Results and Discussion paragraphs generated from your live compute output as LaTeX or Word. Includes reviewer-simulation: what a skeptical reviewer will object to, and the citation that answers it. Live since 2026-08-03.
Every claim is grounded in a vector-indexed library of 50+ PLS-SEM classics (Chin, Henseler, Hair, Sarstedt, Ringle, Shmueli). Citation chips open the source paragraph in a drawer. Live since 2026-08-03.
Free during Beta. Off by default. No credit card. One click in Account settings - or right when you open the robot in the bottom-right corner of the app.