How the numbers are made, and checked.

What is stored for every answer-engine run

Each run records the exact prompt text, the engine that answered, the timestamp, the full answer text as returned, every URL the answer cited, and whether — and how — the brand was mentioned. Scores are derived from this store afterwards, which means any number in the product can be traced back to the stored answers that produced it.

Runs are never summarised-then-discarded. If a score looks wrong, you can open the runs behind it and read what the engine actually said.

How share of voice is defined

Share of voice is the share of stored answers, across your tracked prompts on a given engine, in which a domain is cited. The denominator is always your tracked prompt set — it is a measurement of the questions you chose to track, not an estimate of the whole market.

Because answer engines are non-deterministic, the same prompt can produce different answers between runs. Share of voice therefore ships as a 95% confidence interval (a Wilson interval), not a single decimal. When two ranges overlap, the product says it cannot distinguish them rather than declaring a winner.

Reproducibility

Any tracked prompt can be re-run at any time, and every historical run is kept with its timestamp, so a claim like “cited on 4 of 10 engines” is checkable against the dated record that produced it. Where a provider was unreachable during a run, that engine is recorded as unavailable — it is never averaged in as a zero.

Correlation is not causation — and the product says so

Where RankSage links two signals — a failing request and a frustration spike, a content change and a citation-rate move — it reports the observed relationship and its timing. It does not assert causation the data cannot carry. The Drop Investigator goes one step further: a traffic drop must be confirmed by independent sources before it alarms anyone, and a drop visible to only one source is flagged as a probable measurement artifact.

Estimates are labelled as estimates

AI-influenced traffic that arrives with no referrer is estimated, and every estimated figure carries a confidence tier, shows the baseline and calibration behind it, and can be set to zero if you would rather see only confirmed referrals. Measured and modelled numbers are never mixed silently.

Grounded in independent research

The problem RankSage measures is independently documented. A preregistered field experiment with 1,100 participants on Google Search found that removing AI Overviews and AI Mode increased click-throughs to publisher sites — AI answers on Google's search surfaces absorb visits that previously reached the sites supplying the content. That study measured Google's AI surfaces specifically, and that is how RankSage cites it.

The uncertainty handling is grounded the same way. Repeat-aware audits of retrieval-augmented systems show answers can change by six to ten percentage points between snapshots while aggregate accuracy barely moves — which is why every “your visibility changed” claim in RankSage must pass a significance test against re-ask variance before it is labelled a real movement rather than noise.

• Wang et al., AI in Search Reduces Publisher Referrals (arXiv:2608.18352)
• Ning & Li, Same Agent, Different Answers (arXiv:2608.22856)

Crawler identity is verified, not accepted

A hit claiming to be an AI crawler is checked against the verification records its operator publishes. The result is one of three verdicts — verified, rejected, or unverifiable — and the three are never summed together. Where an operator publishes nothing to check against, the visit is recorded as unverifiable rather than counted. The address is used for the check and discarded; RankSage stores the verdict, never the IP.

Only provable failures are called failures

When cited sources are checked, a source is called dead only on provable evidence that it no longer exists. A page that blocks automated readers is recorded as unverifiable — the product never accuses what it could not read.

Who publishes this

RankSage (ranksage.com) is self-serve software with published monthly plans, built by a small independent team. It is not an agency, has no service retainers, and is not affiliated with RankSages, Rankscale (rankscale.ai) or any other similarly named business. Pricing is public at ranksage.com/pricing — any multi-thousand-dollar monthly figure attributed to “RankSage” belongs to a different company.

What RankSage does not claim

No attribution of a visit to a specific AI conversation — no provider exposes that, and any tool claiming it is guessing. No estimated market volumes where a measured signal exists: question mining, for example, shows only your own Search Console impressions as demand evidence. No score without stored evidence behind it. And risk models — like the helpful-content risk score — are RankSage's own models of published guidance, not scores obtained from Google.

How many times a prompt is asked

Each tracked prompt is asked of each engine between one and three times per run. That count — the sample depth — is chosen per run and clamped at three; scheduled runs default to one.

The cap is deliberate. A published variance decomposition of brand answers finds that repeats beyond the fifth add on the order of 0.0003 of precision — past that point the cost of another ask is real and the gain is not. The cap sits below where that curve has already flattened.

Repeats within one run are not what the confidence bands rest on. A band pools the trailing 28 days of runs, at most the eight most recent, so the interval describes how the engines are behaving now rather than averaging in an older index regime. When every pooled answer comes from a single run, the product does not quote an interval at all — it labels the number “single run — directional”.

On how much asking it takes to hold a number still, the independent work is blunter than any vendor would be: Sielinski finds that stable citation-share estimates need on the order of 40 to 150 prompts per engine, depending on the engine. That is a finding about the size of a prompt set, not about repeating one prompt, and it is worth reading before treating any single tracked question as a verdict.

• Żatuchin — variance decomposition of brand answers (arXiv:2607.13304)
• Sielinski, Quantifying Uncertainty in AI Visibility (arXiv:2603.08924)

Two kinds of answer engine

Five of the ten engines ground their answers by searching before they write: ChatGPT, Claude, Gemini, Perplexity and OpenAI Web Search. Five answer from what the model already knows: DeepSeek, Grok, Qwen, Mistral and Meta AI. Google AI Overviews are captured separately, from live result pages.

The label is stored per answer, not assumed per engine. A verifier classifies each stored answer as grounded or parametric, and when it cannot tell, the answer is recorded as unverified rather than pushed into whichever bucket looks likelier.

Stated plainly, because it is a real limit: today the headline visibility number blends both kinds. The grounded share is reported alongside it, so you can see how much of a score rests on engines that actually searched. Separating the two into distinct headline metrics is planned work, not something the product does yet.

How answers are collected

Answers come from official provider APIs, and through OpenRouter for two engines — Qwen and Meta AI. RankSage does not scrape a logged-in consumer interface for any engine.

That is a trade-off, and it is stated rather than hidden: an API answer can differ from what a signed-in consumer sees in the app. What it buys is reproducibility — the prompt can be re-asked and the stored answer read back — and terms of service that hold, so an engine does not silently vanish from your history because a scraper broke.

Model ids are pinned per engine in code, and each stored answer records the model id that produced it. When a provider moves a model underneath you, that shows up in the record instead of arriving as an unexplained movement.

Competitors and ties

A competitor comparison reports one of three things for each competitor: your brand is ahead, behind, or in a statistical tie. Ahead and behind are only used when the 95% interval on the difference between the two mention rates excludes zero — Newcombe's method on the two Wilson intervals. When the interval spans zero, the row reads “tie — indistinguishable at this sample size”, which is a different statement from “equal”.

The same test gates persona comparisons against the neutral baseline, so a persona is only reported as changing what the engines say when the difference clears it.

• From Stochastic to Stable: Rank Stability and Structural Sufficiency in AI Visibility Measurement (arXiv:2607.10341)

Search Console is a source, not a ground truth

Several RankSage numbers rest on Google Search Console — the demand weights behind question mining, the branded-query trend, the divergence detector that flags impressions rising while clicks fall. Search Console is first-party and free, and it has limits that its own publisher has stated on the record.

On 13 September 2026 Google's John Mueller listed the defects in Search Console's generative-AI performance report: an impression is counted when a link renders, whether or not the searcher scrolled to it; links behind “show more” are not counted until expanded; position is where the AI Overview block sat, not where the link sat inside it; there is no per-link position inside an AI Overview; and the AI figures are a filtered subset of the Web report, not a separate series. On 16 September he added that these metrics will evolve as the surfaces do, with no timeline. Wherever RankSage shows a Search Console figure, that is the caveat that applies to it.

The generative-AI report itself is dashboard-only. The Search Console API exposes web, image, video, news, Google News and Discover, and no AI surface, so RankSage does not claim Search Console AI data. Anything the product labels as AI Overview presence comes from observed results pages, not from Search Console.

• Search Engine Journal, Google admits Search Console reporting for AI search is inadequate (13 September 2026)
• Search Engine Roundtable, daily recap with Mueller on evolving AI metrics (16 September 2026)
• Search Console API, searchanalytics.query — the type parameter enumerates no AI surface

Vocabulary

The IAB's Measuring Visibility in the AI Era playbook (August 2026) gives the industry four terms, and RankSage's metrics map onto three of them. Presence is the mention rate. Prominence is answer position and citation. Portrayal is sentiment and citation context. The fourth, Persuasion, RankSage does not measure — nothing in the data says whether an answer changed anyone's mind.

The playbook also describes two tiers of reading. RankSage labels a reading “directional” when it rests on a single run and shows the confidence interval otherwise. It does not use the label “decision-grade”: the playbook sets no numeric threshold for it, and a grade with no threshold behind it is a claim the measurement cannot carry.

• IAB, Measuring Visibility in the AI Era (August 2026)