Measurement protocol
A measurement you can defend in front of your leadership team.
What AI engines say about your brand varies from one answer to the next. Our protocol separates what actually moved from what is just chance.
Was the measurement actually completed?
Every question, on every engine, as many times as planned. Otherwise, we do not publish.
Do the numbers reconcile?
Every total must add back up to its components, every rate to its denominator.
What can a single measurement tell you?
What rises above the noise, and nothing more. The limits are written down in plain terms.
A deliverable that does not answer these three questions is not an audit. It is a dashboard.
Step 1 of 5 · Measure
Inside "Measure": six steps, none starts until the previous one is complete.
Measure, Diagnose, Decide, Produce, Re-measure: the re-measurement reruns exactly this chain.
- 1Buying questions50 questions validated with you
- 2Collection across 5 engines5 runs per question
- 3Coverage check100% required before analysis
- 4AnalysisNamed, source only or absent
- 5Cross-checks19 checks before delivery
- 6DeliverableFindings, appendix and data
50 questions × 4 assistants × 5 runs = 1,000 assistant answers, and 50 questions × 5 = 250 Google AI Overviews searches, counted separately (Core tier). An incomplete measurement is rerun, never extrapolated.
Why repeat
The same question, asked 5 times, does not get the same answer.
A single answer measures nothing: it is a coin toss. That is why every question is asked 5 times on every engine.
Our noise floor, ±15%, was measured by running the full protocol twice on the same night, on August 9, 2026, before any comparison. A smaller variation is not a result.
“Which ITSM tool should a 200-person company choose?”
- Answer 1Named · 2nd position
- Answer 2Absent
- Answer 3Named · 4th position
- Answer 4Source only
- Answer 5Named · 3rd position
Result: named 3 times out of 5. Neither “always” nor “never”.
Non-negotiable rules
What we never allow ourselves, so your numbers hold up.
Every rate is reported against the full protocol, never against a subset picked after the fact.
Being named in the answer and appearing only as a source are two distinct outcomes, counted separately.
A question that already contains your name is never mixed with the others: it would inflate the result.
A percentage without a margin of error says nothing. Every published rate has one.
The competitor list is reviewed with you by an expert, not inferred by an algorithm.
What the measurement does not allow you to conclude is in the report, at the same level as the results.
Comparison
A tracking tool shows you numbers. An audit tells you which ones to trust.
| A tracking tool | A ZivRank audit | |
|---|---|---|
| Coverage | Shown whenever data is available | 100% required before any analysis |
| Runs | Often one answer per question | 5 runs per question and per engine |
| Variations | Every variation is shown | Noise floor of ±15% set before comparison |
| Competitors | Detected automatically | Validated with you by an expert |
| Data | A dashboard | Timestamped raw answers handed over |
Before delivery, 19 cross-checks verify that every number in the report reconciles with the raw data.
What you receive
Everything you need to verify, and to rerun the measurement.
Findings report
Where you are named, where you are not, and who is named instead of you.
Methods and data appendix
Scope, denominators, margins of error and limits.
Technical audit
What prevents AI engines from reading and citing your site.
Replayable measurement file
The same questions, replayed identically at every re-measurement.
Timestamped raw answers
Every answer from every engine, with its sources. You can check everything.
Honesty
What the measurement does not say.
Every report has a limits section. Here it is, in short.
- One period is not a trend
A measurement is a snapshot. Trends are read across several measurements under the same protocol.
- Answer accuracy is not scored
We measure whether and how you are named, not whether everything the AI says is correct.
- No causality claimed
We document what was published and what was observed afterward. Nothing more.
- Google AI Overviews measured separately
The metric is different: it is never added to the assistants.
After the measurement
Decisions, not opinions.
- Every action addresses a measured gap. A question where you are absent, a source that cites your competitors: no gap, no action.
- Every goal sits above the noise. A goal below ±15% would not be measurable: we do not set one.
- An observed pickup time of 3 to 8 weeks. That is how long AI engines take to pick up a published asset. We re-measure after, not before.
- What we don't do. No fake reviews, no mass-produced content, no ranking promises.
FAQ
Questions about the protocol
Credits and prices: pricing. Agencies: the agency offer. See also the baseline audit.
How do you measure results?
We rerun exactly the same protocol: same questions, same engines, same models. We compare before and after, engine by engine and question by question, and we publish what moved as well as what did not. Raw answers, sources and timestamps are kept and handed over to you.
Which AI engines do you measure?
ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews, measured separately because its metric is not the same.
Are Google AI Overviews results added to the assistants?
No. The 4 assistants give 1,000 answers per measurement; Google AI Overviews is measured separately, on 250 searches, because its metric is not the same. The two figures are never added together.
How is the noise floor set?
The full protocol was run twice on the same night, on August 9, 2026, before any comparison: the gap observed, ±15%, is our noise floor. It is neither a universal margin of error nor a confidence interval.
How long before you see results?
In the published case, 44 days separate the first measurement from the second, with assets live for only 8 to 13 days: the re-measurement came before the usual pickup window ended, and citations of the brand had already increased. The full effect is therefore probably understated. Each engine picks up new assets at its own pace; ChatGPT is the slowest. This is a measured case, not a promise.
Start by seeing where you stand.
Up to 5 of your buying questions, expert video walkthrough within 48 hours. Free.