A number you cannot check is just a rumour.
Almost nothing an AI-visibility tool reports can be independently verified by the person reading it. That makes the reporting habits more important than the feature list.
The problem
Any tool can print a citation rate. Without the answers behind it, an honest account of what failed to run, and a range instead of a false decimal, that number is a claim you have no way to audit.
Why you can check it, explained in about a minute
Crawler identity verdicts, as they appear in the product.
Address resolves to the operator's published records
Claimed GPTBot, resolved to an unrelated host
Operator publishes no records to check against
Unverifiable is its own bucket on purpose. It is never counted as a confirmed visit, and never dropped either.
How it works
What it measures, what comes out, and why the number can be trusted.
- 01
Evidence first, score second
Runs store the answer text and the cited URLs. Crawls store the page structure. Scores are derived from that afterwards, so you can always get back to what produced a number.
- 02
A check that could not run says so
There is a real difference between “we looked and found nothing” and “we could not look”. Collapsing them into a zero is the most common quiet lie in this category, and RankSage keeps them apart everywhere.
- 03
Ranges where the thing is variable
Answer engines are non-deterministic. Reporting a single decimal implies a precision that does not exist, so measurements that vary are reported as ranges.
- 04
Identity is verified, not accepted
A crawler hit claiming to be GPTBot is checked against the operator's published records before it counts. Rejected claims are shown in their own bucket rather than deleted or quietly included.
- 05
Nothing degrades silently
If a provider is down, an integration is missing, or a source could not be reached, that appears where the affected number appears. A quiet fallback is how you end up trusting a broken pipeline for months.
What it does not do
- None of this makes an estimate a measurement. Where something is modelled, it is labelled as modelled.
- It cannot verify what a provider does not expose — no tool can attribute a visit to a specific AI conversation.
- Being honest about uncertainty means some numbers look less impressive than a competitor's. That is the trade.
Works with
Verified AI Crawlers
A hit that says ClaudeBot might not be. Every claimed identity is checked against the operator's own DNS records before it counts.
AI Share of Voice
See how often ten answer engines — from ChatGPT to Grok — name you, reported with a confidence range, not one flattering number.
AI Traffic & ROI
Count the visits AI actually sends you, including the ones your analytics files as direct because the referrer was stripped.
Questions
How do I know an AI visibility tool is telling the truth?
Ask three things: does it keep the raw answers behind its scores, does it distinguish a failed check from a negative result, and does it report variable measurements as ranges. A tool that does none of those is asking for trust it has not earned.
Why do some numbers show a range instead of one figure?
Because the underlying thing varies between runs. A single decimal would look more precise than reality and you would end up reacting to noise.
What happens if an integration breaks?
It surfaces where the affected numbers appear, rather than the numbers quietly shrinking. Silent degradation is the failure mode this product is built to avoid.
AI is already answering for you.
Sign in to Google Analytics and Search Console — read-only, about five minutes, nothing to install on your site.