Know what the evidence can support

Before changing a message or funding an initiative, inspect what was observed, how it was counted and what remains unknown. This public method covers AI-answer observations: one source for a growth decision, without a claim of market-wide or revenue causation.

1 of 8 · Variable answers

AI answers vary, so one answer proves little

AI responses can vary between runs. ChoiceOrigin records what was observed under the stated conditions rather than treating AI output as deterministic.

2 of 8 · Repetition

One run per provider is an observation, not a measured share

Below 30 answers we show counts, not percentages.

The detail: One run per provider is an observation, not a measured share

A Snapshot asks each of its 8 buyer questions once on each supported provider: at most 24 answers. Repeating the same questions over several runs is what turns these counts into rates; the Choice Diagnostic observes repeatedly.

How many answers a figure rests on, and how it is shown
AnswersShown as
3, one buyer question1 of 3, count only
8, one provider3 of 8, count only
24, whole Snapshot7 of 24, count only
Sample data, fictional brands

3 of 8 · Decision panels

Buyer-question panels

In a Snapshot, buyer questions are written by an AI model from your website, in the words a buyer would use. A Choice Diagnostic uses buyer questions and name variants you approve.

  1. 1

    Draft

    Written from your website, in the words a buyer would use

  2. 2

    Approve

    In a Choice Diagnostic, you approve every question

  3. 3

    Pin

    Every re-observation asks exactly the same panel. A change to any of it is reported, never compared silently.

4 of 8 · Provenance

Provider and model provenance

Every answer is stored with the provider, the model identifier exactly as returned, and the time. If a provider changes its model between runs, the comparison says so.

The detail: Provider and model provenance

Each question is asked through each provider's API in a clean session, with no history.

provider
gemini
model
as returned by the provider, stored verbatim
requested_at
2026-09-30T14:02:11Z
question
buyer question 1 of 8
answer
stored in full, as returned
sources
reviewhub.example/crm
counted_as
Northwind Pipeline and Brightline named, Acme CRM absent
One evidence record. Sample data, fictional brands

5 of 8 · Gaps

Unavailable providers and partial results

Providers fail, rate-limit or refuse. What fails is marked unavailable and left out of totals. We never estimate missing answers.

  • Perplexity8 of 8
  • OpenAI8 of 8
  • Gemini3 of 8, partial
Partial run. Totals use 19 answers, not 24. Sample data, fictional brands

6 of 8 · Before and after

Before and after, tested as a change

A re-run asks the same buyer questions on the same providers. The change is read with an exact test. If the test cannot tell a change from variation, the reading says so in words, next to the counts, not with a coloured arrow.

Runs observed with different method versions are not compared.

  • 2 Sep 20265 of 24
  • 30 Sep 20268 of 24
  • ReadingNo change detected
This sample could only detect a rise of 7 or more surfaced answers. Sample data, fictional brands

7 of 8 · Limits

What we cannot tell you

These limits are part of every report, not small print.

  • Why a model chose. We see co-occurrence in answers, not reasons.
  • What an individual buyer saw. Signed-in history, location and memory change answers; we use clean API sessions.
  • Revenue impact. We measure who AI answers name, not pipeline. Linking the two is your data, with our help.
  • Providers we do not query. Today: Gemini, OpenAI and Perplexity only.
  • Causation from a before and after alone. Without a comparison, movement is observed, not attributed.

8 of 8 · No fake precision

Why there is no visibility score

A single 0 to 100 index hides the count, the sample and the uncertainty behind it. Two scores that look different can be the same result. We show what was counted instead.

What we do not publish

73/100

No count, no sample, no way to check.

What we publish

7 of 24

Answers naming Acme CRM, out of 24 answers to 8 buyer questions on 3 providers, with the method version and every answer as returned.

Sample data, fictional brands

How each number is counted

Every figure in a report is a count of recorded answers. This table gives the exact numerator and denominator behind each one, as the report computes them.

Question
One buyer question asked in an audit. A Snapshot plans up to 8.
Run
One ask of one question to one provider. An audit can ask the same question more than once on one provider; every ask is a separate run with its own answer.
Answer
What a provider returned for one run. If a call was retried, only the latest attempt is kept.
Counted answer
An answer the provider returned and that was scored. Failed, timed-out and unavailable runs are not counted.
The numerator and denominator behind each number a report shows
Number shownNumeratorDenominator
A provider surfaced your company in K of N answersCounted answers from that provider whose text names your company (a case-insensitive whole-word match on the company name).Counted answers from that provider.
A question's result for one provider: surfaced, absent or unavailableSurfaced when at least one counted answer for that question names your company. Absent when counted answers exist and none does.Counted answers for that question and provider. With none, the result is unavailable and carries no count.
Questions where AI surfaced youQuestions where at least one counted answer, from any provider, names your company.Questions with at least one counted answer.
N answers recorded of M requestedCounted answers, across all providers.All runs planned for the audit, counted or not.
K of N answers cited at least one sourceCounted answers that list at least one source.Counted answers.
A provider cited your site in K of N answersCounted answers from that provider whose cited sources include your own domain.Counted answers from that provider.
X of Y answers checked named a specific company, firm or personChecked answers that name your company, or put forward another company, firm, product or person as an option.Counted answers that went through the name check.
A company or site appeared in K of N answersCounted answers from that provider that name that company, or cite that site, counted once per answer.Counted answers from that provider.
Before and after: K of N answers, then K of N answersCounted answers that name your company, in the first audit and in the second.Counted answers in the question and provider pairs that have a counted answer in both audits. A pair with no counted answer in either audit is left out and reported as left out, never as absent.

A run that failed, timed out or was unavailable is not counted: it leaves both the numerator and the denominator and the report marks it unavailable. It is treated as unavailable, never as not named. It still counts among the runs requested, so the report shows how many were left out.

  • Counts, not measured shares. Every run's answer counts in N, so when a question is asked several times on one provider, N grows with the runs, and those answers are repeats of the same question, not independent buyers. We show counts with no confidence range, at one run or several.
  • Answers are collected through each provider's API, in a clean session. They can differ from what the same provider shows in its consumer apps.

Loading the recorded report...

Explore ChoiceOrigin for your company.

Inspect a recorded report, or request a demo for your company.