Understanding the numbers.
The tools return the same measured numbers as the dashboard — with the same honesty about what they do and don't mean. Read this before you quote an AI-visibility figure to anyone.
The MCP tools return exactly what the dashboard computes — no extra precision, no rounding away the uncertainty. But an assistant can make any number sound confident, so ScoutRival ships the caveats with the data: each tool's payload carries the range and the honesty note attached to its figure. This page explains what those caveats mean, so that when you or your assistant repeat a number, you keep the part that stops it from lying.
Two engines, not four
AI-visibility measures two engines, and they measure different things:
- ChatGPT — web-grounded. These are the answers that carry citations, so this is your citation signal: whether AI, when it searches the web to answer, actually names and links you.
- Gemini — an ungrounded "model memory" probe. It runs without web search, so it tells you what the model already believes about you from training, with no citations. It is a genuinely useful different signal, not a weaker version of the same one.
Never read the figures as "3 engines" or "4 engines", and don't attribute a run to Claude — Claude is not one of the engines we call. A Gemini-only run is a memory signal, not a citation signal, and the payload labels it that way. If you quote a number, quote which engine it came from.
Every rate carries a range
Every rate — mention rate, share of voice, citation share — is measured from a finite sample of answers, so it comes with a ± confidence range. That range is not decoration; it tells you how much the number is allowed to wobble on its own.
- A change smaller than its range is not a real move. If you were at 42% ± 12 and you're now at 47%, that's flat — treat it as flat, not a trend. The tools grey out a delta smaller than its interval rather than draw a green arrow.
- Too few mentions means no percentage. When a figure is computed from very few data points — sentiment below the minimum sample, for instance — we show raw counts instead of a percentage. "30% positive" off three mentions is noise with a decimal point, and we refuse to print it.
Position is an average, not a rank
Where we report a position, it is an average across answers, not a fixed rank. AI answers do not have a stable leaderboard: ask the same prompt again and the brands can come back in a different order. So a position of 3 means "on average you land around third when you're named", not "you rank #3 in ChatGPT". Never present it as a rank, and never promise a customer a stable AI ranking — there isn't one to promise.
Answers are non-deterministic
AI answers are non-deterministic. Ask the same question twice and the set of brands named can differ, sometimes by a lot. That is exactly why we sample across many prompts and runs and report rates with ranges rather than reading a single answer as the truth.
It also means a single run is a snapshot, not a verdict. Read one run for direction, not for a final score. And nothing runs on a schedule — a check happens when you ask for one, and you can re-check any day. If an assistant tells you ScoutRival "checks your AI visibility daily", that is wrong: it is on demand, every time.
Traffic lags a few days
Search Console traffic — Google or Bing — lags by roughly three days. The most recent days will always look low, simply because that data hasn't landed yet, not because your traffic fell. Don't read the last two days as a drop. A few more things to hold in mind when you quote traffic figures:
- Position in search is an average too, the same as in AI answers — not a fixed rank for the query.
- Bing click-through is derived, not reported directly — it's computed from clicks and impressions.
- Bing rows are weekly buckets, so don't compare a Bing week against a Google day as though they were the same grain.
"Didn't run" is not "said no"
An engine that never answered is not the same as one that answered and didn't mention you. The tools keep these apart: an engine that never returned — an outage, a rate limit, a skipped run — is marked as not having run, and is not scored against you as a zero. A failed call is not evidence of absence.
- Percentages come only from the answers that actually returned. If some calls failed, they are reported, not quietly folded into the denominator to make the sample look bigger.
- Failures are surfaced, not hidden. When a run reports "3 failed", you can see why, so a transient rate-limit doesn't get mistaken for a broken integration.
- We don't publish AI query-volume numbers. Nobody has AI query logs, so any "your prompt gets N searches a month" figure would be a guess dressed as a measurement — we don't show one.
An assistant will happily turn "42% ± 12" into "you're at 42%". When you share a ScoutRival number, keep its range and its caveat — the honesty is the product, and a figure stripped of its uncertainty is a more confident claim than the data supports.