← Docs

Honest limits.

We wrote this page before we wrote the marketing copy. It's the one doc we want every prospect to read before they pay us.

Out of scope (v1)

These are real platforms / products real customers ask about. We chose not to build them in v1 so we could ship the GEO + auto-deploy loop properly. Some land in v2; some never will.

  • Webflow, Shopify, Wix, Squarespace, Framer adapters — too many template variants to safely auto-deploy without a human in the loop. v2.
  • Custom non-Git Next.js sites (e.g. internally-deployed without a public Git repo) — no path for us to PR changes. v2.
  • Google Business Profile / Local SEO automation — gated by Google API approval. v2.
  • Backlink monitoring + digital PR drafting — adjacent to our wedge, distracts from the GEO loop. v2.
  • Wikipedia article creation — Wikipedia community policy is unambiguous: paid editing of one's own organization is prohibited. We push schema/sameAs out, never in. Forever.
  • Image generation, video / YouTube optimization — different model stack. v2 / v3.
  • Mobile app — desktop dashboard suffices for the workflow. Never.

Where we keep a human in the loop

Stareo is autonomous within bounds. These bounds exist because failures here are expensive — brand, legal, or career-damaging. By default, none of these auto-publish.

  • YMYL content (Your Money Your Life: finance, health, legal, safety) — Stareo can draft and brief, but ComplianceOfficer routes the result to your review queue, never to deploy.
  • Wikidata writes — even when we identify a missing claim, we never push to Wikidata without explicit owner approval. Inflation of false claims is a real risk.
  • Author entity creation — Person schema requires verifiable identity. Stareo flags candidates; you confirm before we publish a sameAs network.
  • Velocity-cap overrides — defaults are 5 pages/day and 2 articles/day. Raising them requires explicit owner approval. We don't unlock fire-hose mode automatically.

What's still experimental in v1

Honest framing: these features ship in v1 but our confidence in their long-term accuracy is lower. We measure them, refine them, and surface the confidence interval where applicable.

  • 30-day Share of Model forecasts — driven by linear regression on weekly history. Wide confidence intervals when you have <12 weeks of data; we surface p10/p50/p90 explicitly so you can see the uncertainty.
  • Counterfactual citation experiments — engines have stochasticity; same prompt + persona doesn't yield identical responses. We require ≥20 samples per arm for a confident verdict and label sub-threshold runs as 'further_experiment' rather than fudging significance.
  • PassageMatcher cosine matching — false positives possible when the engine paraphrases. We verify with Haiku 4.5 at confidence ≥0.78 cosine, but this is an active area for improvement.
  • Cohort priors from fix_pattern_outcomes — only published once we have ≥10 workspaces in a vertical × geography × size cohort. Until then, your workspace falls back to severity-base scoring.

What we'll never do

Strong opinions, weakly held — except these.

  • Auto-publish without ComplianceOfficer approval. The compliance log is append-only at the database level (Postgres trigger), not just by convention.
  • Strip C2PA provenance. Every artifact Stareo produces carries a signed manifest. EU AI Act + California SB 942 require it, and even where they don't, we still do.
  • Use customer data to train models. Your prompts, your content, your competitor data — we don't ship any of it to third-party fine-tuning. Anthropic's standard data-handling guarantees apply to inference; nothing leaves that loop.
  • Hide the receipts. Every change has reasoning + expected outcome + measurement plan stored against it. You can audit any deployment back to the LLM call that proposed it.