GEO KPIs: AIsee's Framework for Measuring AI Visibility

Measure the answer, the evidence, and the outcome

A GEO measurement model needs three layers. Report diagnostics organize a site's AI presence, competitive visibility, and website readiness into investigation priorities. Answer observations record whether a brand or owned content appears, how it is framed, and what evidence accompanies it. Business outcomes record referrals, assisted conversions, sales feedback, and other observable effects.

Keeping the layers separate prevents false conclusions. A brand mention is not automatically positive. An owned citation may sit next to a claim the page does not support. Referral traffic can be small even when the brand appears frequently because many interactions end without a click.

Define the measurement universe before calculating a rate. Specify the prompt set, markets, languages, engines, account or personalization state, test dates, and number of runs. A minimum observation record contains prompt_id, prompt text, intent, engine, run date, full answer, cited URLs, mentioned brands, evaluator, and quality labels.

Core GEO KPIs and formulas

Always show the numerator, denominator, and sampling window with a percentage. "40% mention rate" is incomplete; "20 mentions across 50 eligible responses in the July core prompt set" is auditable.

Do not infer a visit or non-visit for each generated response unless the product exposes a defensible response-to-session link. Treat unlinked mentions as an answer-level measure and referral sessions as a separate analytics outcome.

Example: a team tests 30 unbranded prompts twice across two surfaces, producing 120 eligible responses. ExampleCo appears in 24, and 15 include an owned citation. The mention rate is 24 / 120 = 20%; the owned citation rate is 15 / 120 = 12.5%. If nine mentions contain a material error, the accurate mention rate is only 15 / 120 = 12.5%. This is an illustrative formula, not AIsee product or customer data.

Score citation quality, accuracy, and sentiment separately

Binary presence metrics are easy to automate, but human review is needed for meaning. Use short, documented rubrics and retain the supporting excerpt.

Citation quality can use a three-point scale:

• 2 — direct support: the cited page explicitly supports the nearby claim at the relevant scope and date.
• 1 — partial support: the page is relevant but requires inference, omits a material qualification, or supports only part of the claim.
• 0 — unsupported: the page does not substantiate the claim, is outdated for the claim, or points to the wrong entity.

Answer accuracy should distinguish minor wording issues from material errors. A material error could change a buyer’s decision, misstate a capability, confuse two products, or present an expired fact as current. Report both material-error rate and fully accurate rate so cosmetic differences do not hide genuine risk.

Sentiment should be coded as positive, neutral, negative, or mixed according to a written rubric. Do not treat every limitation as negative; an accurate trade-off can be neutral and useful. Pair sentiment with accuracy because a flattering but false description is not a good outcome.

Use double review for a small calibration sample. If reviewers disagree often, refine the rubric before scaling. The AI citation tracking guide provides a practical record structure for citations and nearby claims.

Build a prompt sampling protocol that supports trend analysis

Prompt selection determines what the dashboard means. Build the core set from real customer research, search queries, sales questions, support tickets, category language, and product evaluation tasks. Group prompts by journey stage: problem discovery, category education, solution comparison, validation, and purchase.

1. Version the prompt set. Give every prompt a stable ID. Do not silently rewrite prompts between reporting periods.
2. Separate branded and unbranded questions. A brand-name prompt measures answer control; an unbranded prompt measures competitive discovery.
3. Stratify by market. Keep language, location, audience type, and product segment explicit.
4. Repeat observations. Generated responses vary, so run enough repeated samples to see a distribution rather than a single result.
5. Control what you can. Document login state, personalization, conversation history, surface, and date. Use clean sessions when the research question requires them.
6. Keep an exploratory set. Test emerging questions without contaminating the stable trend series.

Do not claim statistical confidence unless the sampling design supports it. For operational monitoring, show sample sizes, changes in prompt coverage, and uncertainty notes. A five-point swing from ten responses is different from the same swing across several hundred controlled observations.

Annotate major content changes, launches, outages, and platform changes on the timeline. Correlation is not attribution, but annotations help teams form testable explanations and decide what to investigate next.

Design a dashboard from diagnostic to business outcomes

A useful GEO dashboard should answer "what changed, where, and what do we do?" Organize it in four rows:

• AIsee report diagnostics: Overall, AI Presence, Competitor Landscape, Strategy Review, provider rows, and accepted Action Plan tasks. Preserve the report date and do not relabel these values as public-interface percentages.
• Answer presence and quality: mention rate, owned citation rate, AI share of voice, direct-support citation share, material-error rate, and sentiment.
• Source performance: cited domains, cited owned URLs, uncited priority pages, and citation gains or losses after updates.
• Business outcomes: LLM referral sessions, engaged sessions, conversions, assisted pipeline when defensibly matched, and qualitative sales signals.

For referral reporting, use both referrer data and documented campaign parameters. OpenAI states that ChatGPT adds utm_source=chatgpt.com to referral URLs, which gives publishers a specific tracking signal. Maintain a transparent source classification table because browser privacy, apps, redirects, and platform changes can cause undercounting.

Google introduced a Generative AI performance report for selected Search Console properties in June 2026. The current report combines AI Overviews and AI Mode and exposes aggregate impression visibility by page, country, device, and date. It does not provide an AI Overview versus AI Mode split, queries, clicks, or sessions. Use it as a visibility input, not as isolated AI Overview traffic.

AIsee's role in this stack is to structure the audit and turn findings into a reviewable backlog. Controlled prompt observations and analytics remain separate evidence streams. The AI visibility audit page documents that boundary with the dated self-analysis.

Turn the scorecard into a repeatable operating cadence

Review high-risk accuracy issues continuously, the core prompt set monthly, and the broader strategy quarterly. The appropriate cadence depends on product change and market volatility; consistency matters more than a fashionable interval.

At each review, classify changes into an action queue:

1. Fix: inaccurate or unsupported answers tied to ambiguous, inconsistent, or outdated owned content.
2. Strengthen: cited pages that lack a direct passage, primary source, comparison criterion, or update date.
3. Expand: recurring prompts where the brand has relevant expertise but no authoritative page.
4. Consolidate: overlapping pages that split signals or contradict one another.
5. Monitor: movements with insufficient evidence for a content change.

Measure interventions with predeclared success criteria. Example: “For the 25 evaluation prompts, increase accurate owned-citation rate from the documented baseline over three monthly samples, without increasing material-error rate.” Record which URLs changed and when. Rerun the same prompt protocol, then review both the aggregate and the answer excerpts.

GEO measurement is valuable when it changes decisions. A smaller, carefully defined prompt set with preserved evidence is more useful than a vast automated score nobody can audit. Start with the AI search visibility baseline, use the GEO audit checklist to find page-level causes, and keep classic search and conversion reporting beside the GEO scorecard.

Frequently asked questions