Evidence, with context.
Every live metric starts with a stored answer you can inspect. Our results are controlled samples, not a view into every AI conversation.
Read a result in three steps
- Check the questions, providers, dates and sample count.
- Open supporting answers to see what was actually said or cited.
- Investigate a repeated gap, make a considered change, then track comparable questions again.
No percentage is universally good or bad. Changing questions, competitors or provider mix can change the result without a change in your brand.
A repeatable sample
Each tracking run records your prompt, provider, model, timestamp, returned text and available sources. We preserve historical answers when you edit a prompt or run tracking again.
Answers can vary between runs. Models and search indexes change. Geography may affect answers, and a market named in a prompt does not guarantee geographic targeting.
What a result represents
API samples may differ from consumer ChatGPT, Gemini or Perplexity experiences. SignalLens cannot observe private AI conversations or represent every answer every user receives.
Failures are recorded independently and excluded from metric denominators. Missing citations mean no citation evidence was returned, not that a source has never influenced a model.
Metric dictionary
- AI Visibility
- Brand-present valid responses ÷ all valid selected responses × 100. Each answer counts once, regardless of repeated name mentions. This is your prompt sample, not all AI activity.
- Mentions
- Count valid selected responses with at least one detected brand name, configured alias or domain reference. A mention is not necessarily a recommendation.
- Share of Voice
- Your response-level mentions ÷ summed response-level mentions of your brand and selected competitors × 100. An answer can mention several brands. Changing competitors changes this denominator; this is not market share.
- Citation Rate
- Valid responses with a provider citation to the brand’s recorded domain or subdomain ÷ valid selected responses from citation-capable providers × 100. Each response counts once. Non-grounded Gemini is excluded; its citation rate is unavailable. A prose URL alone is not a citation.
- Prompt Coverage
- Selected tracked prompts with a brand-present valid response ÷ all selected tracked prompts × 100. Questions without an answer stay in the denominator. Live history can include archived questions. This measures your question set, not market demand.
- Competitor Delta
- Read the direction in the view: the live competitor matrix and sampled-day reports use comparison brand minus your brand; legacy prompt comparisons and demo summaries use your brand minus the comparison brand where shown. Positive means the first named brand has higher visibility. A difference is not statistical significance.
- Average Position
- Sum of detected brand positions ÷ responses with a meaningful ordered position. Unranked prose is excluded. This is position inside the sampled answer, not a search ranking.
- Valid Response
- Only responses marked valid by the capture/normalization rules enter presence metrics. Valid does not mean fact-checked, correct or representative. Failures remain separate and do not count as brand absence.
- Tracking Run
- A run records questions, provider/model, time and available answer evidence. Some tasks may fail while others succeed. Repeat runs can differ; configuration and provider changes affect comparability.
- Source Gap
- In live Sources, a citation gap requires competitor-present citation evidence and a target-present comparison with the same question revision, provider, model and locale, with no target-present citation to that source in the selection. Missing comparison evidence is not a confirmed gap. Demo gaps are simpler illustrative co-occurrence examples.
- Brand Perception
- Separate analysis with supporting excerpts and model-reported confidence. It does not change deterministic visibility or citation metrics. Confidence is not measured accuracy and a few answers do not establish market sentiment. Demo labels are authored examples.
Presence, sources and interpretation
Gemini visibility is measured from controlled non-grounded Gemini responses. Google Search grounding is disabled for manual and scheduled tracking. Ordinary answers support deterministic brand/competitor detection, visibility, share of voice, prompt coverage and trends; separate model-assisted perception remains available. Gemini citations and source intelligence are unavailable, not zero. Previous Google-grounded Gemini captures are excluded and removed from the active analytics database.
Brand detection uses exact names, configured aliases and domain references with conservative word boundaries. A mention does not imply endorsement. Citation association means a cited domain matches a tracked brand; third-party endorsement is not inferred from a co-mention.
Comparisons use the same selected response set. No response means no observation; we do not substitute zeroes or fabricate previous-period changes. Inspect the captured answer before acting on any opportunity.
Demo workspaces contain authored fictional examples and are always labeled Demo data. They are structurally separate from live projects.
Detailed intelligence rules · intelligence-v1
- Comparable evidence
- Every comparison uses the same selected date, provider, prompt, current topic/category and recorded locale filters. Failures are excluded. Tracked prompts with no answer remain in the coverage denominator; archived prompts are included when selected history contains their evidence.
- Prompt classification
- Zero valid responses: Unmeasured. One or two: Insufficient data. With at least three: Strong means target present in every response; Mixed means some; Competitor-led means target absent but a tracked competitor present; Gap means no tracked brand present. Three is a conservative display guard, not a statistical significance threshold.
- Won, lost and shared
- A response is Won when the target is detected and the selected competitor is not; Lost is the reverse; Shared means both; No presence means neither. A question may have several outcomes across samples. These labels do not imply quality, endorsement or ranking.
- Competitor delta
- The intelligence matrix shows competitor visibility minus target visibility in percentage points. All brands share the selected valid-response denominator. Mention rate equals visibility; citation count is unique responses citing that brand’s recorded own domain, not repeated citation markers.
- Topics and locale
- Business topics use current prompt groups. Historical topic labels were not captured, so editing a group changes its analytical grouping without rewriting the underlying answer. Recorded market/language come from the original run snapshot. Unknown values stay unknown; market context is not guaranteed provider geolocation.
- Citation gaps and source overlap
- Source frequency is unique valid citing responses. Target-only, competitor-only and shared describe brand co-occurrence in those responses, separately from own-domain association. A gap requires no target-present citation to the source and at least one target-present comparison with the same question revision, provider, model and locale in the selected window. Without that comparison we report insufficient evidence, not a gap.
- URL identity
- Only returned provider citations enter source statistics. We remove fragments, www and known tracking parameters; sort remaining query parameters; and preserve meaningful query values, HTTP/HTTPS and path trailing-slash differences. Without a verified canonical or redirect relation we do not assume distinct resources are identical. No source crawling occurs.
- Trends and changes
- Daily observations use only captured valid samples; missing dates are not zeroes. Change detection compares the latest two sampled UTC dates, matching exact question text, prompt ID, provider, model and locale. The latest response in each cell/date is used. At least three matched cells are required for a headline visibility delta. Before/after evidence remains inspectable below that guard. Changes are observations, not causal findings or statistical proof.
- Opportunity order
- Investigations originate from prompt/topic/provider gaps, own-domain citation absence, comparable citation gaps, matched visibility decline and single-domain citation concentration. Minimum three supporting valid responses. Sort is transparent: response count, provider breadth, prompt breadth, recency, stable identifier. No ROI forecast, weighted AI score or guaranteed visibility improvement.
- Model-assisted interpretation
- Only persisted, separately requested analysis supplies attribute themes. Themes retain model and exact supporting excerpts. They never alter deterministic metrics and opening intelligence pages does not call a model. Sparse interpretation samples cannot establish broad brand sentiment.
- Reports and exports
- Reports inherit the selected evidence scope and identify their method version. CSV uses explicit analytical fields and spreadsheet-formula neutralization; raw credentials and server configuration are not exported. Browser printing supports saving a concise report as PDF. Original evidence remains available through response drill-down.