> ## Documentation Index > Fetch the complete documentation index at: https://docs.getcited.in/llms.txt > Use this file to discover all available pages before exploring further. # Reading your data with confidence > How to interpret your Cited dashboard — when a number is signal, when it's noise, and how long to wait before reacting. AI visibility data is noisier than traditional web analytics. The same prompt asked to the same AI engine twice can produce different brand lists; AI platforms quietly update their models; live retrieval pulls different web pages across runs. None of that means the data is unreliable — it just means a single day's reading isn't the whole story. This page explains how to read your dashboard with the right level of confidence so you act on real movement and ignore noise. ## What a single day really tells you Each day, every prompt in your library is sent to every tracked AI platform and the responses are parsed. Your dashboard then aggregates the results into the day's metrics — [mention rate](/concepts/metrics/mention-rate), [share of voice](/concepts/metrics/share-of-voice), [average position](/concepts/metrics/average-position), [sentiment](/concepts/metrics/sentiment), [citation rate](/concepts/metrics/citation-rate). A single day's number is an aggregate of many responses, not one. That aggregation is what makes daily metrics stable enough to put in front of you. But because AI responses are [non-deterministic](/concepts/foundations/non-determinism), day-over-day shifts of a few points are normal background variance, not evidence of real change in your visibility. ## When to trust a movement Not every change in a metric deserves a reaction. Use this rough rule: * **Single-day change:** treat as noise unless it's dramatic (≥ 15 percentage points). Most one-day swings are just LLM variance. * **3–5 day directional move:** worth investigating, not worth acting on. If mention rate has slipped four days in a row, start looking — but don't change content or strategy yet. * **2-week trend:** reliable. If a metric has moved consistently for two weeks, the change is real. This is the minimum window we recommend for content or strategy decisions. * **Month-over-month:** the strongest comparison window. Long enough to wash out non-determinism, short enough to be actionable. If you find yourself reading the dashboard daily, watch the 14-day rolling view rather than the single-day spot value — it'll save you reacting to noise. ## The five things that produce noise that isn't a signal Some movement in your metrics has nothing to do with what your brand or your competitors did. The most common sources: 1. **LLM non-determinism.** The same prompt run again produces a slightly different brand list. 2. **Model updates from AI providers.** OpenAI, Google, Anthropic, Perplexity, and xAI update their models on their own schedules, sometimes without public notice. 3. **Live retrieval variance.** Search-enabled platforms like [Perplexity](/concepts/platforms/how-perplexity-ranks-sources) and [ChatGPT](/concepts/platforms/how-chatgpt-search-works) pull different web pages across runs. 4. **News and seasonal spikes.** A brand mentioned in a press cycle gets a temporary lift that fades as the news cycle moves on. 5. **Platform-side outages.** If a tracked platform was rate-limited or down when a prompt ran, that prompt contributed no data for the day. When you see an unexpected movement, the diagnostic question is whether competitors in your category moved with you (category-wide shift) or whether your brand moved alone (brand-specific signal). Brand-alone moves that hold for two weeks are the ones worth acting on. ## Reading the dashboard the way Cited does A few habits make the data more useful in practice: * Anchor on the **14-day rolling view**, not the single-day number, for any decision that costs time or money. * Compare against **your own past trend**, not against absolute thresholds. A 6% mention rate may be excellent for your category and weak for another. * Use **share of voice** alongside mention rate. Mention rate tells you whether you're visible at all; share of voice tells you whether you're visible *relative to competitors* — which is what actually matters when prompts surface a list of brands. * When the [AI Narrative](/methodology/refresh-cadence) refreshes, it incorporates the latest aggregated data — the narrative reflects the trend, not yesterday's noise. ## Related concepts * [Non-determinism](/concepts/foundations/non-determinism) — why even good measurement produces some variance * [Mention rate](/concepts/metrics/mention-rate) — the headline visibility metric * [Refresh cadence](/methodology/refresh-cadence) — when each dashboard surface updates ## Frequently asked questions Reliable to within a few percentage points day-over-day at Pro and Scale tiers, where each daily run produces several hundred prompt-platform data points. At Starter tier, variance is larger. In all cases, the 14-day rolling view is the most reliable read. More prompts do increase statistical confidence, but each prompt has a measurement cost. The plan-tier prompt sizes (25 / 75 / 125+) balance reliability against cost. Deep-dive audits with 200+ prompts are available for point-in-time snapshots that need higher confidence. Check three things: did the drop persist for 3+ consecutive days; did competitors in your category move with you or did you drop alone; was there a known model update from the relevant AI platform in that window. Persistent + brand-alone = worth investigating. Anything shorter than three days = wait.