AI-citation reporting · Methodology

How we measure AI citations.

Most “AI visibility” reporting fails one test: can the client check it? This page documents our full measurement protocol (the prompts, the engines, the logging, and the places where the method has limits) so every number in your report is checkable.

Fixed prompt set · logged model versions · movement, not attribution

01 · Why we publish this

A number you can’t check is a claim, not a measurement.

“AI visibility” reporting has earned its reputation for vagueness: screenshots without prompts, scores without model versions, dashboards without timestamps. If you can’t see how a number was produced, you can’t tell measurement apart from marketing.

So we publish the protocol. Every figure in a Point Visible report traces back to a logged run: the exact prompt, the engine, the model version, the timestamp. You can audit any data point in your report against that log, and you can audit our method against this page.

02 · How measurement works

The measurement protocol.

  1. 1

    A fixed prompt set, frozen at baseline

    For your category we build a set of roughly 20–25 buyer-intent questions, the questions a real buyer would put to an AI assistant before a purchase decision. The set is frozen at your baseline run and stays identical every quarter. A question list that changes between runs can show whatever a vendor wants it to show; a frozen one can’t.

  2. 2

    Three engines, measured separately

    Every prompt runs against ChatGPT, Perplexity and Google AI Overviews. The three cite differently, so results are reported per engine. Collapsing them into one “visibility score” hides more than it shows.

  3. 3

    Model versions and timestamps logged on every run

    AI engines ship updates constantly. Without version logs you can’t tell whether movement came from your coverage or from a model change. Every run in your report records the model version and the timestamp, so you can see exactly which runs straddle an update.

  4. 4

    Repeated runs, because answers are non-deterministic

    The same prompt on the same engine can return different answers minutes apart. We run each prompt repeatedly per engine and score across the full set of runs. One lucky answer does not count as presence.

  5. 5

    Quarterly cadence, aligned to placement dates

    Measurement re-runs each quarter, and the results are plotted against the dates your links and mentions went live. Movement and work always appear side by side, on the same timeline.

Parameter Specification
Prompt set ~20–25 buyer-intent questions per client category · frozen at baseline
Engines ChatGPT · Perplexity · Google AI Overviews
Logged per run Model version · timestamp
Sampling Repeated runs per prompt, per engine · scored across all runs
Cadence Quarterly · aligned to placement dates
First run Free baseline · human-reviewed · ready in 24 hours

03 · What we report

Four things, every quarter.

Citation presence

For each prompt on each engine: is your brand cited or recommended in the answer. Scored across the repeated runs, not a single sample.

Who is cited instead

The competitors and publications occupying the answers you’re absent from. This is where the next quarter’s placement targets come from.

Movement vs. baseline

Change per engine since your baseline run: up, down, or flat. Flat is reported as flat.

Placement dates, side by side

Every movement chart is annotated with the dates placements went live, so you can judge the correlation yourself.

The pledge

We report movement, not attribution. AI answers shift for reasons beyond any single link: model updates, competitors’ coverage, community discussion. We show you the movement and the work, side by side, and let the correlation speak. We will show you flat results when they are flat.

04 · Known limitations

Where the method has limits.

No AI-answer measurement is free of these. We list them so you can read your report with the same caveats we apply when we write it.

Limitation 01

Non-determinism

The same prompt, same engine, same day can return different answers. Repeated runs narrow the variance; they don’t eliminate it. Treat single-prompt differences as noise and consistent movement across the set as signal.

Limitation 02

Model updates reset baselines

A major model update can shift answers across the board overnight. When that happens, quarter-over-quarter comparison weakens, and the report says so. The logged model versions show exactly where the break sits.

Limitation 03

Personalization variance

What an engine shows you personally can differ from what it shows our measurement runs: answers vary with location, history and session. We measure under consistent conditions, but no setup reproduces every individual user’s context.

05 · How the free baseline relates

The baseline is this exact methodology, first run free.

The free AI-citation baseline is not a lite version. Same prompt-set construction, same three engines, same version and timestamp logging. Human-reviewed before it reaches you, ready in 24 hours. All it takes is two fields: your domain and a work email.

If you become a client, the baseline is re-run quarterly, and every future report measures movement against it. If you don’t, you still keep the baseline: it’s a real measurement either way.

Run your baseline. Then check it against this page.

The same protocol documented above, run once for your domain, free, and human-reviewed before it goes out.

No call required · ready in 24 hours · yours either way