Scope label — self-built proof. Every figure shown here comes from a measurement we ran on our own build. We name this scope on the artifact itself. Revenue, closed-deal and ROI figures ship with the evidence behind them — every ROI figure we publish is one that survives a documentation request, and a figure without that evidence does not appear.
50
buyer questions behind a full-grid case study, each one put to 15 engine captures
Every reading keeps its status, not just the wins
17
surfaces read for every prompt, each on its own denominator
The same surfaces the methodology names
7
runs of every prompt, so a figure is a measured rate — not a lucky draw
Rates ship with Wilson 95% intervals
The cost of a soft format. Without a stated denominator and window a result cannot be checked, so the figure you repeat becomes a risk you carry personally the moment somebody asks where it came from. CiteAngle results are reproducible from published conditions, sealed at measurement time, so a rate with its Wilson 95% band holds up, because the conditions behind it stay public.
Why the format is this strict
Case studies in this market are mostly unfalsifiable. A percentage appears, a logo sits next to it, and the reader has no way to ask the only question that matters: compared with what, over what window, measured how? That is fine as decoration and useless as evidence, and it becomes expensive the moment somebody repeats the figure in a board meeting and gets asked where it came from. A result that cannot be wrong cannot be relied on either.
So this format forces every claim to expose the thing that could contradict it. A result line carries its baseline, its window, its denominator and the interval around it, and a delta gets reported as two ranges so a reader can see whether they even separate. Where the evidence stops, the page says so instead of rounding the gap away.
There is a reason we are the ones publishing it. The worked example below is measured on our own build. The first company held to this standard is this one. That comes before any customer is asked to accept a number produced the same way. Everything quoted here is registered with its source and as-of date in the claims registry, and the protocol behind each reading is public on the methodology page. You are welcome to hold any vendor, including us, to exactly this.
1. Use case and industry specificity — name the problem, not the logo
A case study opens with the use case and the industry it belongs to, because that is what a buyer weighs. In AI answers, use-case fit outranks brand recognition when a reader decides whom to shortlist, so a write-up that leads with a logo and a headline lift teaches nothing transferable. Ours leads with the exact question set a buyer in that vertical would ask an AI assistant. It names the surfaces those answers appear on. It names the competitive frame the engine already believes. That is the same starting point our evidence review argues for.
2. Measurement design — the numbers have a method
Every figure in a CiteAngle case study traces to a measurement grid, not an estimate. The full grid runs 50 prompts across 15 engine captures, with every prompt repeated seven times, and read on the 17 surfaces of the US roster. Nothing is sampled. Every reading keeps its status (observed, cited, failed, indeterminate), so the write-up shows the misses next to the wins instead of reporting only the readings that succeeded. Applicable citation rates are measured and reported with Wilson score 95% confidence intervals. None of this is proprietary to the write-up: the grid, the repeat design and the interval math are public in the methodology, and the delivered screens are shown in the sample report. A case study that cannot point to its measurement design is a testimonial, not evidence.
3. Numbers and period — what a result line actually reports
A result is never a single score. Each line carries five things:
| Field | What it states |
|---|---|
| Baseline | The citation rate at the start, with its confidence interval and its denominator |
| Re-measurement | The rate on the same grid at a later point, with its own interval |
| Window | The exact start and end dates the two measurements were taken |
| Denominator | The surfaces, prompts and run count behind each figure |
| Change | A citation-rate delta stated with both intervals — shown only when a re-measurement on the same grid exists |
A change is reported as a citation-rate delta with both intervals attached, so a reader can see whether the two ranges even separate. It is never stated as a bare percentage lift with no denominator, and never as an ROI multiple. If the intervals overlap, we say so.
4. Limitations and scope — where the evidence ends
Every case study ends by stating what it does not prove. AI answers are non-deterministic, so a result is a measured range over a stated window, not a permanent number; a later window can move. The worked example is a self-built proof. We ran the whole grid on our own brand, on the protocol we publish. Every reading it produced is still on the record. A buyer's report goes to that buyer and never onto this page, so the format here has to prove itself on evidence you can open right now. Commercial figures — revenue, pipeline, a closed deal — ship with their evidence attached and a figure without that evidence does not appear, so every line you read is a line you can check.
The format in one checklist — the same four parts, in order, on every case study we publish.
Want this format filled with your own brand's evidence?
Start with a free AI visibility check, then order a Panorama at its listed price — the measurement runs on your live surfaces, and the write-up follows the four parts above.
The person who signs off on every figure here is named on the company page, and the axes we would want a buyer to hold us to are on the comparison page.