The short answer

Capture the full ChatGPT Search answer, the question and its linked sources under recorded conditions. Count a citation to your domain separately from a mention of your name or a recommendation to buy from you. Keep non-search conversations, controlled API tests and consumer Search observations in distinct reporting groups.

A worked record from one test session

On November 3, 2026, Arrow ran four questions in logged-out ChatGPT, with web search requested. The table keeps each outcome, including the failed capture, rather than reporting a single score. Arrow’s domain appeared only in the branded control. That answer shows ChatGPT can reach the site for a question naming it; it says nothing about selection for buyer questions.

Four questions in one session cannot support a rate or a trend. Their use is as a baseline to repeat with the same wording, and as a reminder that an empty source list and a failed capture are different results.

A worked record from one test session
QuestionStatusLinked domains recorded
Which AI visibility platforms help a small B2B marketing team inspect citations and turn the findings into website improvements?Failed capture: answer incomplete when savedNot measured
Which GEO platforms fit a B2B team that has limited engineering capacity and needs changes implemented?Validsearchatlas.com, athenahq.ai, tryxlr8.ai, scrunch.com
Do I need an llms.txt file for AI search visibility?ValidNone found in the page links
What is Arrow AI (arrow-ai.us) and what does it do? (branded control)Validarrow-ai.us

Capture enough information to review the result

Keep the exact prompt, full answer, source URLs, language, intended market and relevant session context. A screenshot of a logo alone may not show which question was asked or what the source supported. Preserve a readable response record as well as the image when your collection method permits it.

If a source link redirects, record the visible citation and the final destination. Check the passage before concluding that the assistant accurately represented your page.

Use a citation-tracking worksheet with explicit outcomes

Create one row per observation. Keep repeat observations as separate rows and disclose how many were collected for each question. Review ambiguous source or brand matches before counting them.

Use a citation-tracking worksheet with explicit outcomes
FieldWhat to save
Observation ID and UTC timestampStable reference for the full saved response.
Question and conditionsExact prompt, language, market, product surface and relevant session context.
Search and source evidenceWhether Search was used; visible citation URL and final destination.
Brand and support checkCorrect entity; mention, citation and recommendation as separate flags.
Collection statusValid observation, failure or unmeasured; retain the reason.

Separate search access from observed selection

OpenAI documents different crawler purposes, including search discovery and potential model-training use. Review access for the crawler relevant to your objective. Passing that check does not turn it into a citation result.

Use the observed citation gaps to review your source pages: a missing condition, unsupported claim or outdated product description may be actionable. Compare repeat observations after the change without claiming that every subsequent difference was caused by that edit.

Calculate citation rate without hiding collection gaps

For a fixed panel, citation rate equals valid answers citing your domain divided by all valid answers in that panel. Count an answer once for this domain-level rate even when it links to several pages on the domain. Keep page-level citation counts as a separate measure.

Illustrative example: 60 checks were planned, 50 yielded reviewable answers and 10 did not. If 8 valid answers cited the domain, the observed citation rate is 8/50 = 16%, with 83.3% collection coverage. Neither 8/60 nor 10/60 describes the valid-answer citation rate.

A GA4 referral visit shows that someone followed a trackable link. It does not reveal all citations, and a citation without a click is not a visit. Report citations, visits and conversions in separate columns.

What to check

  1. Record whether Search was used, along with the visible model or product mode and observation date.
  2. Open the cited page and check whether it supports the statement attributed to it.
  3. Report valid observations, non-mentions and collection failures before calculating rates.

A practical example

An answer can cite an Arrow measurement guide while recommending a different product. That is an observed source citation for Arrow, not an Arrow product recommendation. A report should preserve both facts rather than collapsing them into one positive score.

Sources and further reading

These steps are Arrow AI's implementation guidance. The example is illustrative; it is not a measured customer result. Publication and source access do not guarantee an AI recommendation.

Apply this to your website

Identify the source gaps and decide what to improve with Arrow's free audit. Review what Arrow GEO measures before choosing a reporting scope.