How we measure AI visibility
Engines and data sources
We ask each engine the exact prompts you track, in the market and language you choose, and store every answer in full together with its citations. Live answers come from official APIs, either through DataForSEO's AI Optimization and SERP APIs or directly from the providers.
| Engine | Live source |
|---|---|
| ChatGPT OpenAI | DataForSEO LLM Scraper: queued scheduled checks; live on demand. Provider UI model is observed, not force-selected. |
| Perplexity Perplexity | DataForSEO LLM Responses; or Perplexity Sonar API |
| Gemini | DataForSEO LLM Scraper: queued scheduled checks; live on demand. Provider UI model is observed. |
| Claude Anthropic | DataForSEO LLM Responses (web search on); or Anthropic Messages API with the web search tool |
| AI Overviews Google Search | DataForSEO SERP API — the AI Overview block on Google's results page |
| AI Mode Google Search | DataForSEO SERP API — Google AI Mode endpoint |
API answers with web search are a close proxy for what consumers see in each app, but not an identical copy: consumer apps may personalise answers, use different model versions or show different sources. We label engines accordingly and recommend reading trends rather than single answers.
Sandbox mode
Accounts without live data credentials run in sandbox mode. Answers are generated by a deterministic simulator, clearly labelled Sandbox data throughout the product, and never mixed with live data. When live sources are connected, tracking starts a fresh baseline.
Metric definitions
Metrics use all distinct stored observations in the rolling 28-day window, not just the latest answer. Sample ranges describe observed variability, not a population guarantee. We do not invent missing demand or benchmarks.
Share of AI Demand (estimated) is the AI search-volume-weighted mean mention rate. Each prompt contributes its mean outcome once, so more repeats cannot multiply demand. Volumes are estimated by DataForSEO, refreshed monthly in batches for editable head keywords. Unknown volumes are excluded and flagged; measured zero remains zero. With no positive known volume the metric is unavailable.
- mentioned — 1 if any of the brand's names, aliases or domains appears in the answer (word-boundary, case-insensitive; text inside URLs does not count).
- position — the 1-based order of the brand's first mention among all tracked brands mentioned in that answer; empty when not mentioned.
- position weight — 1 → 1.0, 2 → 0.85, 3 → 0.7, 4–5 → 0.55, 6+ → 0.4.
- cited — 1 if any citation's registrable domain belongs to the brand.
AnswerScore = mentioned × 100 × (0.7 × position weight + 0.3 × cited)
| Metric | Definition |
|---|---|
| AI Visibility Score | Importance-weighted mean of AnswerScore over every prompt × engine in the window (importance 1–3). Also computed per engine and per topic. |
| Mention rate | Answers that mention the brand ÷ all answers. |
| Share of voice | The brand's mentioned-answer count ÷ all detected brand appearances, including untracked brands. Detection coverage is disclosed; old answers lacking entities cannot establish a complete denominator. |
| Average position | Mean position over answers that mention the brand. Lower is better. |
| Citation share | Answers citing the brand's domain ÷ all answers over 28 days. Multiple links in one answer count once. The confidence panel hides this metric below 20 answers. |
| Net sentiment | % positive mentions − % negative mentions, hidden below 20 labelled mentions; n is shown. Lexicon estimates are not a factual-accuracy judge. |
Dashboards show rolling 28-day windows computed daily. Diagnostic chart deltas are not alerts. Alerts compare disjoint 28-day periods at two consecutive actual scans and require same-direction difference intervals excluding zero. Regime/composition changes and insufficient observations remain Within noise.
Sentiment and recommendation stance
We use a transparent lexicon with negation handling, marked estimated. Paid judge and writer calls without exact billed-USD accounting are disabled; deterministic templates remain available without provider spending.
Sampling and limitations
- Ordinary scans take one observation per prompt and engine. Confidence comes primarily from rolling history: Wilson 90% intervals per cell, deterministic seeded prompt-cluster bootstrap for totals and weighted demand. Constant/single clusters use a conservative Wilson fallback; delta endpoint envelopes are conservative, not independently calibrated 90% intervals.
- Repeated samples are not bought just because a result is near 50%. Confirmation spending is reserved for candidate changes and capped at 10% of the plan provider budget. Sparse data can remain inconclusive indefinitely.
- All paid features share a monthly provider-spend budget. Extra confirmations stop first, daily refresh degrades to weekly, then pitches pause. Actual billing can exceed a per-call reservation because upstream token APIs do not offer a USD ceiling; overruns are recorded and stop later admission.
- Results depend on the market and language configured for each workspace.
- Google AI Overviews do not appear for every query. When none is shown, the brand counts as not visible for that prompt.
- Brand detection relies on names, aliases and domains you configure. Add aliases for abbreviations and product names to avoid misses.
The AEO audit
The audit fetches the homepage, robots.txt, the XML sitemap, llms.txt and a sample of internal pages, then runs 48 checks. Each check passes, warns or fails; category scores are weighted means of their checks, and the overall score weights categories as follows:
| Category | Checks | Weight |
|---|---|---|
| AI crawler access | 10 | 25% |
| Structured data | 11 | 20% |
| Answerability | 9 | 25% |
| Authority & trust | 7 | 15% |
| Technical health | 11 | 15% |
Failing and warning checks become recommendations, ordered by impact, then severity, then weight.