AI visibility methodology
The measurement standard behind every NimbleSEO visibility score: reproducible captures, effective sample sizes, descriptive uncertainty, and inspectable citation evidence.
A visibility estimate summarizes what retained answers show for a defined set of prompts, engines, locales, and dates. It is always read with its effective sample, capture method, and uncertainty state. If there is not enough defensible evidence, NimbleSEO shows no number.
Lag-1 autocorrelation can reduce raw runs to an estimated effective sample size. This can widen intervals, but it does not establish independence or correct for every shared source, engine or time effect.
Independence correctionPrior-window intervals are compared with the next observed rate. That rate is itself noisy; this describes compatibility and does not validate nominal coverage or a forecast.
Descriptive compatibility checkPrompts that name the tracked brand are measured in their own lane and never inflate headline visibility or share of voice.
Branded-lane exclusionWhen a provider response shape changes, affected captures keep a degraded reason and are excluded from metrics instead of silently becoming a zero.
Schema-drift canaryClassic search and AI-answer visibility share one workspace, evidence store, action queue, and plan contract. Prompt checks, crawl pages, seats, and managed credits are itemized on the pricing page.
We report observed citation rates with descriptive Wilson intervals. Their nominal 95% level assumes independent observations; estimated effective sample sizes do not prove independence or calibration. Small samples remain inconclusive. These ranges describe the selected captures, not consumer visibility or future performance.
Representative questions with provenance
Measurement starts with the questions a real audience asks, not an opaque list of synthetic phrases.
- Prompts can come from connected search data, autocomplete, community questions, CRM personas, support language, imports, or disclosed expansion.
- Source, topic, locale, region, intent, funnel stage, and demand provenance remain attached to every tracked prompt.
- Discovery candidates never enter recurring reports until they are explicitly tracked.
Retained evidence, not a screenshot claim
Every reported result traces back to a timestamped answer and its extraction context.
- Evidence records the engine, prompt, answer excerpt, capture method, timestamp, locale, and answer hash when available.
- Citations retain URL, normalized domain, order, snippet, ownership class, and answer context.
- Bare API-model, search-grounded, browser-real, imported, and manual lanes are labeled separately.
Measurements you can inspect
Scores exist to prioritize work; they are not presented as exact market-share accounting.
- Core measurements include mention rate, citation share, answer position, source diversity, sentiment, engine coverage, and competitor share.
- Weighted share of voice uses tracked non-branded prompts and keeps the run count, engine mix, date range, and confidence state visible.
- Actions rank evidence quality, source authority, competitor gap, closeability, and the available workflow path.
Variance changes what the product recommends
Generative answers vary, so repeated captures and uncertainty change what the product recommends—they are not a footnote.
- Repeat runs produce trustworthy, noisy, unreliable, or inconclusive states.
- Trend and alert language is downgraded when the effective sample is weak or run-to-run overlap is low.
- Samples below the minimum support are suppressed or shown as inconclusive rather than forced into a score.
Mention rate, answer position, citation presence, engine coverage, and competitor context.
Effective sample size, repeat captures, citation overlap, brand-cite consistency, and descriptive variation.
Evidence quality, source authority, competitor gap, closeability, owner path, and expected workflow impact.
- First-party and licensed demand signals are labeled separately from directional discovery proxies.
- Engine, interface, model, region, locale, account mode, capture method, and time are retained where available so reruns remain comparable.
- Browser-real and API-model evidence stay in distinct lanes unless a report explicitly discloses the comparison.
- Citation presence, referral traffic, conversion, revenue, and ranking movement are distinct results; attribution appears only when connected evidence supports it.
- Engine and algorithm changes can alter answers without a customer-side change, so volatility is reported rather than hidden.
- No ranking, mention, citation, traffic, or revenue result is guaranteed.