Measure GEO with citations, recommendations and qualified opportunities

CH.01

Traffic is one part of the outcome

Traffic remains useful when it is tied to a business objective. It cannot, by itself, tell you how often an AI answer cites your work, whether a recommendation is positive, or whether a lead meets your sales criteria.

Our Zero-Click Whitepaper compares studies with different samples, periods and click definitions. They help explain why visibility and website visits can move differently; they do not prove that traffic must keep falling or that a fall is evidence of successful GEO.

Pew Research studied browsing from 900 U.S. adults in March 2025 and collected the corresponding search results in April. A link in the AI summary was clicked on 1% of visits to Google search pages with a summary. That denominator is search-page visits, not citation links, all AI users or your website visitors. Use your own conversion counts and definitions when judging commercial value.

CH.02

Tier one: citation rate

Citation rate measures how often a captured answer links to your target pages. Track unlinked brand mentions separately. Neither metric replaces search rankings; each describes a different surface.

For this framework, citation rate is the number of eligible captured answers containing at least one target URL divided by all eligible captured answers, reported per engine. Count an answer once even if it cites several target URLs. Define eligibility before collection, and report requested runs, captured answers, missing answers and errors alongside the rate. Record when an AI Overview does not appear and state how that observation enters your denominator.

How to measure it

• →Prompt set: start with the buying questions that matter to your business. A 100–300 prompt set can be a planning range, not a statistical requirement. Label educational, recommendation, comparison and validation questions; version the set when the business changes.
• →Sampling: repeated runs reveal variation. Three runs per prompt can be an initial cost estimate, not proof of reliability. Retain each answer and report the numerator, denominator and variation; do not silently replace missing results with zero or hide them through majority voting.
• →Collection: record engine, product surface, model or mode when available, date, locale and account context. An API response is not automatically equivalent to the consumer interface. Use manual checks or permitted tooling that matches the surface you intend to measure.

Common misreadings

• →A single screenshot shows one answer, not a stable citation rate. Keep the prompt and collection context so the observation can be repeated.
• →A linked citation and an unlinked brand mention are different fields. A citation also does not prove endorsement, factual correctness or purchase intent; inspect its context.
• →Report engines separately and show sample counts. A combined rate can change merely because the share of observations from each engine changed.

CH.03

Tier two: share of answer

A brand can be cited as a source without being recommended to a buyer. This tier records recommendation presence and sentiment in a defined set of buying questions.

Here, share of answer means the share of eligible recommendation or comparison answers in which the brand belongs to a predefined recommendation group. Specify whether that group is an explicit top pick or the first 1–3 listed options; those are different rules. This is a reporting convention, not a universal industry standard or measured buyer shortlist.

How to measure it

• →Use the recommendation and comparison subset of the versioned prompt set. Show its own answer count; its denominator differs from the full citation-rate set.
• →Define the coding rule before reviewing answers. Record first-group recommendation, other mention or absence, plus positive, neutral or negative context. Preserve the supporting answer for ambiguous cases.
• →Aggregate at a cadence suited to sample volume and decisions. Monthly review is a planning recommendation, not proof that recommendation presence changes more slowly. Human review can check ambiguous automated labels.

Common misreadings

• →High citation rate on educational questions does not imply recommendation presence on buying questions. Compare the subsets before deciding what information is missing.
• →Share of answer is an observation in your chosen sample, not market share or a demonstrated correlation with sales.
• →Negative mentions need their own label. More appearances can mean more criticism, so a raw count is not a success measure.

CH.04

Tier three: pipeline contribution

The third tier connects observable interactions to qualified opportunities. Pipeline value, closed revenue and causal impact are different measures; a report should name which one it can support.

Define the qualification stage, opportunity value, observation window and credit rule with sales before calculating contribution. Keep an observed referrer, a buyer-reported influence and model-assigned credit as separate evidence fields. Match permitted records by an appropriate identifier and avoid counting the same opportunity more than once. Missing source information remains unknown.

Common misreadings

• →Small samples need counts, qualification rules and uncertainty. Do not assume AI referrals convert better because of a channel label or another company’s reported multiple.
• →A credit model assigns contribution under its rules; it does not prove causation. Show the selected model and lookback window. First-user acquisition and event-attribution reports answer different questions.
• →Use a window that fits the actual sales cycle and data availability. Monthly reporting and quarterly decisions can be useful cadences, but they do not establish a universal 6–18 month cycle or guarantee that pipeline will arrive.

CH.05

A reporting field specification you can adapt

The table below is Tenten’s proposed field specification. Use the fields needed for the decision, document their definitions and assign an owner; it is not a record of measured client performance.

Keep raw observations and interpreted metrics distinguishable. Publish numerator, denominator, sample coverage and update date with each rate. When a definition changes, keep its previous version and mark the break in comparability rather than freezing a known error forever.

CH.06

Connect each review to a decision

Choose the review cadence from the decision and available evidence. Weekly, monthly and quarterly meetings are options, not proof that the three metrics form a causal sequence.

Start the report with the question to be decided, the relevant observations and the unresolved limits. A change in one tier can suggest a follow-up question in another; it cannot substitute for that tier’s evidence.