CitadexDocs

How the Numbers Are Measured: Ranges, Data Sources, Platform Changes and Evidence

The 95% ranges, each answer’s data source and model, platform-wide change and model-update markers, repeated sampling, Frozen comparison, the AI searches tab, the three evidence tabs and corrections, and each engine’s position in the evidence.

AEO Fundamentals6 min read

Why this matters

Ask AI the same question twice and the answers can differ, and AI platforms change models and citation habits on their own. These features show how many answers each number rests on, where they were captured, and whether a change came from your brand or from the platform.

95% ranges

Rates show the answer count and a 95% range. On the Dashboard card From mention to recommendation, “95% range 41–63% · 23/45” means 23 of 45 answers matched; with that many answers, the true rate lies between 41% and 63% with 95% probability. More answers give a narrower range.

  • Under 10 answers only “x/n” is shown, without a percentage.
  • Retests use the ranges too: a result counts as Improved only when the before and after ranges don’t overlap.

Data source and model

Open a prompt in Prompt Tracking. The answer tab says how that answer was captured:

  • “Source: captured from the AI app itself (what users see)”
  • “Source: model API (a benchmark; may differ from the app)”

followed by the Model name. Answers scanned before this was recorded say so, and get it from the next scan. The From mention to recommendation card sums up the sources of all answers in its bottom line.

Platform-wide changes and model updates

Citadex sees data across all projects. When an AI engine changes for every project on the same day, the Dashboard shows a Platform-wide changes card (hidden when there is none), for example:

  • “Cited links per answer across all projects down 31% (4.2 → 2.9)”
  • “Brands named per answer across all projects up 18% (…)”
  • “New model xxx in use (62% of that day’s answers)”, the model-update marker.

How to use it: if your numbers dropped on the same day on that engine, the engine is the likely cause, not your brand. The same events appear as Platform-wide change markers on the Coverage over time and actions chart (Achievements → View evidence), and in This week’s tasks such drops are labelled as such.

Repeated sampling

In a prompt’s panel, tick Ask 3× per scan. On every scan each engine answers the prompt 3 times, and the result shows as “Latest sampling, 3 answers: mentioned x/3 · recommended y/3”, so you see how often your brand comes up.

  • Available on Business / Agency, for a limited number of prompts.
  • It uses about 3× the credits for that prompt on each scan. Ticking it opens a confirmation (“Turn on "3 runs per scan"?”); nothing changes until you click Turn on. You can turn it off at any time.
  • The credit check before each scan includes the extra runs; if the balance doesn't cover them, the scan doesn't run.

Frozen comparison

Adding or removing prompts moves the average coverage by itself. Frozen comparison locks a group of prompts and compares only those, week by week.

  1. 1

    Freeze

    At the top of Prompt Tracking, click Frozen comparison → Freeze current prompts, or Freeze this market’s prompts.

  2. 2

    Follow the weeks

    Each set shows its size and weekly coverage, e.g. “Same 100 prompts: week of 9/1 32% → week of 9/29 41%”. Each week uses every prompt’s last check that week; brackets show how many prompts were checked.

  3. 3

    Delete

    Delete removes only the set. Scan data is not affected.

Freezing also saves a baseline: each prompt’s last coverage in the 30 days before freezing. The set then shows “Against the baseline, the same N prompts: x% → y%”, comparing only the prompts checked again since. When you freeze one market’s prompts, a prompt later moved to another market is marked as moved and not counted.

Method versions

Results reports, Frozen comparison sets and re-test verdicts keep the numbers they were made with, and record the classification date and the method versions, e.g. “Classified as of 2026-11-17 (brand_class v1 · stage v1), computed with cohort_coverage v1”.

  • When a method is updated, existing reports, frozen baselines and re-test verdicts are not recomputed; a frozen set’s current values are still computed with the method it was frozen with, and the current method version is shown with them.
  • Reports, sets and verdicts made before method versions were recorded show “Method version not recorded”. Those sets have no baseline, and their trend uses the current method.
  • Live pages use the current classification.

AI searches

Many AI engines search the web before answering and cite pages from what they find.

  • Per answer: Searches the AI ran before answering in the answer tab lists them; engines that don’t expose their searches say so.
  • In articles: when you tick prompts and click Generate strategy article, the dialog shows Include AI searches (N): a few searches from those prompts' recent answers that never cited you, which the article also answers; untick it to leave them out. Weekly topics also favour these prompts and list the searches under each topic. The separate AI searches tab was retired. This costs nothing extra.

The evidence panel: answer, Sources, Earlier answers

Click a prompt in Tracked to see its results per engine. The panel has three tabs:

TabWhat it holds
answerRecommendation label and reason, data source and model, AI searches, the full answer (translated or original), and Judged wrong? below
SourcesEvery source this answer cited, with the count on the tab
Earlier answersOlder answers for this prompt on this engine, newest first, each openable; corrected ones are tagged Corrected

Every raw answer is archived. When a judgement is wrong, use Judged wrong?; the correction is kept beside the answer and the answer itself is never changed.

Each engine’s position is in the evidence

Each row of the Tracked list shows coverage (how many engines mention you); click Evidence for each engine’s result. The evidence has one tab per engine:

  • #n on a tab is your position in that engine’s list.
  • No number: you were mentioned, but not in a numbered list.
  • Engines that did not mention you show as not mentioned.
  • 📎: your site was cited without your brand being named.
  • ⚠: this check returned no answer.

Sort the list with the Order button, which shows the current order (e.g. Order: Latest checked): "Priority, high to low", Custom order, Newest added, Oldest added, Latest checked, Coverage (high to low), Coverage (low to high). With Priority, high to low, each row shows its commercial opportunity grade (High, Medium, Low); hover or click it for Commercial opportunity: the scenario, confidence, main reasons, the basis of each factor and the conditions, with "Unknown" for anything without data, never a guess.

To see whether AI still names you when the wording changes, use Phrasing test in the row's ⋯ menu: the same question asked 3 other ways plus the original, each once on ChatGPT, Gemini and Perplexity, with a table of which engines name you for each phrasing. A one-off check: nothing is tracked unless you add it. It uses credits.

FAQ

Two scans gave different results. Is that normal? Yes. AI answers vary from run to run. Look at the trend over several scans, or turn on repeated sampling for key prompts.

My coverage dropped and I didn’t change anything. Check Platform-wide changes on the Dashboard first. A change on the same engine on the same day points to the platform.

How is Frozen comparison different from the normal trend? The normal trend follows every prompt you track now; a frozen set counts only the prompts frozen together, so when its number moves, those prompts really moved.

Was this article helpful?