Public · Methodology
AI visibility methodology
What a visibility number means here: Wilson intervals over effective sample sizes, a calibration backtest on the intervals themselves, a separate branded lane, and a drift canary — all shipped, all disclosed.
What a number means here
We report visibility as an observed citation frequency with a 95% confidence interval (Wilson score over the actual run count), not a single number. The band is the honest part: a tight ±5pp needs ~385 runs, so on real capture counts it stays wide — and when there are too few runs we say inconclusive rather than show a confident %. The major incumbents (Ahrefs Brand Radar, Profound, Semrush) still report a single number, not a range — this is the “weather report from one thermometer” problem, and we solve it.
The incumbent comparison above was last verified against the named vendors’ own public pages on 2026-07-10. When your prompts run, each estimate carries its run count and interval — while no runs exist for your brand, no number is shown at all.
Effective sample size (n_eff), not raw run counts
Engines re-serve persisted answers between captures, so raw run counts overstate independence. Intervals are widened to an effective independent-observation count (n_eff) derived from lag-1 autocorrelation — never the raw count.
shipped: src/lib/v3/av-stability-model.ts
The error bars are themselves backtested
A CI-calibration backtest continuously checks whether published intervals actually cover next-window outcomes — build the interval from prior weeks, test it against the next week, report realized vs nominal coverage. If the bands run tight, the page says so.
shipped: src/lib/v3/ci-calibration.ts
Branded prompts are excluded from headline visibility
Asking an engine "what is <brand>?" and counting the echo is not visibility. Branded prompts are measured in a separate, clearly-labeled lane and never inflate the headline number.
shipped: supabase/migrations/20260630000012_geo_prompt_branded_lane.sql
A schema-drift canary guards the pipeline
When an engine’s response shape drifts, affected runs are flagged and excluded from metrics rather than silently miscounted. Degraded captures carry the reason on the row.
shipped: supabase/migrations/20260630000015_geo_runs_drift_flag.sql
One engine, two lanes, one bill
Classic search and AI-answer visibility run in the same workspace on the same subscription: the find→fix loop, rank tracking, and AI-visibility sampling share one prompt/keyword cohort, one evidence store, and one plan contract — prompt checks, crawl pages, and credits are all itemized on the pricing page. No second tool, no second invoice, and no re-explaining your site to a separate AI product.
Measurement model
1. Prompt universe
How prompts enter measurement
Prompts should represent the questions buyers, customers, and searchers actually ask, not only synthetic phrases.
- Accepted sources include GSC queries, autocomplete, PAA, community questions, CRM personas, support snippets, manual imports, and fanout expansion.
- Each candidate keeps source, source query, topic, prompt group, locale, region, intent, funnel stage, demand proxy, difficulty proxy, and status.
- Tracked prompts are the only prompts used for recurring visibility reporting; parked candidates are discovery inputs, not report claims.
2. Capture
What counts as evidence
AI visibility should be based on retained answer evidence, not screenshots copied into reports after the fact.
- Evidence rows store engine, prompt, answer excerpt, raw answer trail when available, capture method, account mode, viewport, answer hash, and timestamp.
- Citation rows store URL, domain, order, snippet, sentiment, support confidence, and position context.
- Every engine’s capture mode (bare API model, search-grounded API, or manually seeded) is disclosed per engine rather than hidden behind one platform count.
3. Metrics
How scores are read
Scores are operating signals. They should help teams decide what to fix, not pretend to be exact market-share accounting.
- Core metrics: brand mention rate, citation share, cited pages, source diversity, sentiment, answer position, competitor share, and action backlog.
- AI Visibility Score is directional and should be interpreted with sample size, date range, engine mix, and confidence band.
- Source actions rank by owned/competitor citations, engine spread, proof quality, authority, closeability, and expected workflow impact.
4. Confidence
How uncertainty is handled
Generative answer engines are non-deterministic. Repeated captures, sample size, and variance language are required for defensible reporting.
- Repeated prompt-engine captures produce trust labels such as trustworthy, noisy, unreliable, or not enough runs.
- Confidence bands appear next to trend and alert numbers when the sample allows it; below the minimum run count the label is inconclusive.
- Low-confidence drops and competitor surges are downgraded or marked directional rather than triggering high-severity claims.
Formula disclosure
Visibility
Mention rate, answer position, citation presence, engine coverage, and competitor context.
Confidence
Sample size, repeat captures, citation overlap, brand-cite consistency, and variance band.
Action priority
Proof quality, source authority, competitor gap, closeability, owner path, and expected impact.
Known limitations
- The product does not yet have Ahrefs- or Semrush-scale proprietary global prompt demand.
- Prompt volume and difficulty proxies are directional unless backed by licensed or first-party demand data.
- AI answers vary by account state, geography, model version, interface, prompt wording, personalization, and time.
- Browser-real capture is stronger evidence than API-only capture, but it still needs consistent account, viewport, region, and parser QA.
- Citation presence does not guarantee traffic, conversion, revenue, or ranking movement without attribution evidence.
- Benchmark rows are suppressed when the sample is too small and must be quoted with the included disclosure sentence.
Competitor standard
Descriptions of other products are dated claims: last verified 2026-07-10 against the linked sources below, and rechecked in the monthly currency review. Tell us if something changed.
Ahrefs
Real demand plus semantic coverage
Ahrefs Brand Radar publicly frames its methodology around search-backed prompts, People Also Ask expansion, fanout, and recurring high-demand topics rather than arbitrary prompt lists.
- Implication: prompt discovery needs demand provenance, topic coverage, and expansion traceability.
- Our current edge: first-party GSC, support, CRM, autocomplete, community, and fanout sources retain prompt provenance.
- Gap: no Ahrefs-scale global keyword corpus or market-wide modeled prompt index.
Semrush
Data-source and score transparency
Semrush documents AI Visibility Toolkit data sources, benchmark-style visibility scores, cited pages, citations, monthly audience, prompt research, and exportable reports.
- Implication: buyers expect every score to say where it came from and what it means.
- Our current edge: raw evidence, citation URLs, confidence language, and action queues live beside the score.
- Gap: public score formulas, engine coverage, and data freshness need to be visible without login.
Surfer
Explain the score users act on
Surfer explains that Content Score is improved through content structure and term optimization, while warning against over-optimization.
- Implication: a score is only useful when users understand which levers move it and which actions are unsafe.
- Our current edge: source actions and recovery runs keep recommendation factors attached to tasks.
- Gap: AI visibility charts still need formula and confidence language everywhere.
Profound and Peec
Evidence, citations, sentiment, and actions
Focused AI visibility tools sell the operational loop: prompts, answers, citations, sentiment, share of voice, source opportunities, alerts, and action paths.
- Implication: methodology must explain the evidence chain from prompt to answer to source to action.
- Our current edge: evidence browser, source-trigger map, action queue, and recovery handoff are already connected.
- Gap: scheduled browser capture and UI-vs-API comparison are not yet complete.
Sources reviewed
Links checked 2026-07-10.