Methodology
You can't evaluate an AI-search measurement product without auditing the math. Here's exactly what Stareo does — query construction, brand detection, citation matching, experiments, forecasts.
Share of Model is the percentage of (prompt × persona × engine) combinations where your brand surfaces in the answer. Computed weekly, per engine, then averaged for the headline number. We query each engine separately because their answers diverge; the average tells you 'across the AI-search surface, how often do you show up.'
Every workspace has a prompt set (10–100+ prompts, owner-curated or mined from real search-volume data via DataForSEO Labs). Each prompt is paired with each persona — short personality cues we prepend as system context (e.g. 'Australian SMB founder', 'finance director, 50–250 employee firm'). For every (prompt × persona × engine), we fire one live query.
A Haiku 4.5 response parser reads each engine response with structured output and emits brand_mentioned (boolean), brand_mention_position (1–N if mentioned), brand_sentiment, cited_sources, competitor_mentions. Parser prompts are versioned in our LLM router so we can re-score historical responses if the parser changes.
When a brand is mentioned, we run cosine similarity between the response sentences and the embeddings of every passage we've extracted from your site. The closest match (≥0.78 cosine, verified by Haiku 4.5) increments that passage's total_ai_citations counter. Over weeks, this tells you which passages are doing the work — and which ones, if rewritten, would compound the effect.
Stareo doesn't claim 'we made a change and Share of Model went up — must be us.' Every targeted change runs as an experiment: hypothesis, control (your existing passage), treatment (the rewrite), apply, measure at 5 minutes / 24 hours / 7 days. At 7 days we run a binomial test on citation-rate delta and require the 95% confidence interval lower bound to exceed zero before declaring a win. If it doesn't, we revert.
AI engines have non-determinism, and Google's classical SERP has its own churn — Semrush Sensor and AccuRanker Grump publish daily volatility signals. When volatility crosses a threshold for your vertical, Stareo pauses publish-class operations (deployments + new content) until the engines stabilise. Measurement continues; we just don't ship into a noisy week.
The 30-day Share of Model forecast on your dashboard is a Prophet-style time-series fit with p10 / p50 / p90 confidence intervals. We never quote a point estimate without the interval — for workspaces with <4 weeks of history we fall back to vertical-cohort priors, and we mark that on the chart so you know the forecast is borrowed strength, not your own data.
Every Writer-generated draft is scored on an Information Gain (IG) axis against the live top-10 SERP for the target query. We measure how much of the draft is novel — entities, statistics, claims — versus what's already in the top-10. Drafts below IG 0.6 are rejected before ComplianceOfficer ever sees them. The product is for people who refuse to ship undifferentiated content.
We publish: methodology, code, retention windows. We don't publish: customer-specific data, individual prompt sets, the exact rule weights in our Strategist (those are tuned on cohort outcomes and would be gamed if we published them). If you're evaluating Stareo against a competitor and a number on this page doesn't match what they claim, ask both vendors to publish their math.
Questions? Read what Stareo deliberately doesn't do, then try a 14-day trial.