Which metrics track visibility in AI answers?

Follow four measures from exposure to revenue: share of answers mentioning the brand in a fixed question panel, cited pages, assistant-referred sessions, and the sales associated with those visits. A generative engine optimization, or GEO, tool can automate collection. It cannot choose useful questions or see all your sales.

The benchmark

As of the source's October 2026 review, no single data source covered the full journey.

Volume matters. The source reports SparkToro/Gumshoe findings, covered by Search Engine Journal in January 2026: across 2,961 answers from ChatGPT, Claude, and Google's AI to twelve repeated questions, identical brand lists appeared in fewer than 1% of responses. Retain that finding as the cited study's observation, not a universal instability rate.

Google's separate AI report counts impressions; ordinary Web performance remains a different reporting view. Its primary documentation confirms worldwide rollout by August 31, 2026. Bing's June 16 announcement confirms a global preview of citation share and related AI Performance features. Citation share is observational, not traffic or ranking share, and does not identify competitors. Merchant Center's May 27 announcement names the US, Canada, Australia, India, and New Zealand for its forthcoming AI insights; it does not establish worldwide availability. Check the actual account rather than assuming all announced features are live.

Shopify's Agentic page reports assistant-associated sales, orders, and sessions. In one Club women's-apparel example, thirty days showed 174 products accessible in ChatGPT, around 250 AI-agent questions, and €255 in sales. These original euro figures describe one small case.

The rule

Manage the funnel and let revenue inform spending. Three months of growing citations without sessions or sales does not, by itself, justify a larger budget.

Keep the question panel fixed for a quarter. With share of voice p and n collected answers—questions × repetitions × assistants—the source's approximate threshold for a difference between two readings is 2.8 × √(p × (1 − p) ÷ n).

In a fictional example, twenty-four questions repeated three times across three assistants yield 216 answers. At 30% share, the threshold is roughly nine percentage points. Ten questions yield about thirteen points. Real uncertainty may be larger because answers to the same question are correlated. This is a practical noise check, not a formal guarantee of independence.

A paid GEO tool becomes more practical when manual collection exceeds half a day monthly or spans several countries.

The steps

1. Define the panel. Write twenty to thirty purchase questions covering problems, categories, comparisons, and brands. Name three competitors. The brand-citation guide uses around ten for an initial audit; ongoing measurement needs more observations.
2. Repeat consistently. Each month, ask every question three times per assistant in fresh conversations. Record brand presence, links, competitors, and sources. Require the tool to use the target market's language and location, not French-from-France by default. Export raw answers and reject unexplained scores mixing countries and assistants.
3. Cross-check citations. Compare panel sources with Google and Bing cited-page reporting. Update frequently cited pages first, and identify third-party comparisons, media, and forums cited instead of you.
4. Group referrals. Create a GA4 assistant channel group and read rolling four-week trends.
5. Read sales. The supplied ShopifyQL example uses FROM sales SHOW orders, net_sales GROUP BY month, agentic_referring_channel SINCE -90d UNTIL today ORDER BY month ASC. The current official sales schema includes agentic_referring_channel; account access and data coverage still affect what the query returns. Compare it with GA4 AI-channel revenue and a post-purchase “How did you discover us?” option for “an AI assistant.” These produce incomplete lower-bound measures and a rough influence estimate, not additive sales totals.
6. Build the table. Use one indicator per row and one month per column. Annotate measurement incidents in your operating log.

The pitfall

A tracking break can resemble a visibility collapse. Profound’s first-party analysis, checked for this publication, describes its observed ChatGPT citation-link tagging changes from October 8 to October 16, 2026. It saw missing utm_source=chatgpt.com tags and a later recovery in its sampled answers. That is an observation about its sample, not proof of a universal referral-volume decline or an official OpenAI incident confirmation.

What to tell your team

These figures are fictional:

“Across twenty-four purchase questions, we appear in 34% of answers versus 22% three months ago, beyond our approximate noise threshold. Assistants brought $9,800 in recorded sales this month, a lower bound because some referrals lose their identifiers.”

Sources

• Search Engine Journal: SparkToro and Gumshoe research on AI recommendation variability
• Google Search Central: Search Generative AI performance reports
• Practical Ecommerce: Bing AI Performance reporting
• Merchant Center Help: Insights for AI-powered shopping experiences
• Shopify Help: Managing agentic storefronts
• Shopify Developers: ShopifyQL sales schema
• Relevant Audience: Reported ChatGPT tracking-parameter incident
• Search Engine Roundtable: Missing ChatGPT UTM parameters
• Google Search Console Help: Generative AI performance report, English
• Profound: First-party observations of ChatGPT link tagging, October 8–16, 2026
• Microsoft Bing: AI Performance preview features, June 16, 2026
• Google Merchant Center: Planned AI-powered shopping insights