GEO measurement: citations, brand mentions and leads
This page explains a method for measuring AI search visibility. Paul Okhrem’s template records the question, engine, date, citation and brand mention. The blank template is a measurement tool, not a published dataset or industry benchmark.
A reproducible way for B2B teams to measure whether their brand is retrieved, represented accurately, cited, and connected to commercial outcomes across AI answer systems.
Current platform guidance
Google’s AI-search optimization guide retains normal Search requirements. Indexability, snippet eligibility and useful source content still matter.
Its Generative AI performance report measures impressions with page, country, device and date views. It does not provide query-level brand-mention data. The inclusion control is a separate account setting.
OpenAI crawler guidance distinguishes OAI-SearchBot for search from GPTBot for training. A public robots rule alone does not verify CDN access from the provider’s published IP ranges.
Guidance reviewed 29 October 2026. Record account settings and measurement changes separately from content releases.
Measure the full path from retrieval to revenue.
| Stage | Metric | What it answers |
|---|---|---|
| Retrieval | Prompt coverage | For what share of the fixed prompt panel does the brand or its evidence appear? |
| Representation | Answer accuracy | Are the brand, offer, location, proof, and limitations described correctly? |
| Selection | Mention and citation share | How often is the brand named, and how often is a supporting source linked? |
| Source influence | Cited-source distribution | Which owned and third-party pages shape the answer? |
| Commercial action | AI referrals and assisted conversions | Did a user visit, return, start a form, or become a qualified opportunity? |
| Business outcome | Pipeline and revenue | Which opportunities and revenue can be directly or assistively attributed? |
Keep four measures separate
| Measure | Observation |
|---|---|
| Citation | A link to the site appears in the recorded answer |
| Brand mention | The answer names Paul Okhrem |
| Referral | A visit has an attributable source |
| Qualified lead | A real inquiry meets the agreed business criteria |
Use the same questions for each review. Record engine, model when known, date, region and search mode. Run fresh sessions and repeat observations because answers change.
A seven-step repeatable baseline.
- Define the commercial decision. State the audience, category, geography, and business action the prompt panel represents.
- Build a versioned prompt panel. Start with 30 to 100 recommendation, comparison, alternatives, pricing, implementation, problem, and risk questions derived from real sales and search evidence.
- Fix the test conditions. Record engine, model or interface where known, account state, geography, language, device, and collection date.
- Repeat each prompt. Run at least five observations per engine and period. A single stochastic answer is not a trend.
- Capture the raw answer and sources. Record every named brand, order, cited URL, source type, accuracy issue, and whether a clickable referral is present.
- Score only after retaining the observations. Aggregate by prompt and intent cluster; never discard the rows behind a composite score.
- Connect to commercial systems. Preserve AI referral UTMs, original landing pages, assisted journeys, qualified opportunities, proposals, and revenue.
Fields included in the CSV.
| Field group | Required fields | Purpose |
|---|---|---|
| Test identity | study_id, prompt_id, prompt_version, intent_cluster | Prevents prompt drift from being mistaken for performance change |
| Conditions | engine, model_or_surface, run_date_utc, geography, language, account_state, run_number | Makes observations reproducible and comparable |
| Answer | answer_text, brands_mentioned, target_brand_mentioned, target_brand_position | Records inclusion and recommendation context |
| Citations | cited_urls, target_domain_cited, citation_position, source_types | Separates brand knowledge from visible source attribution |
| Quality | accuracy_status, inaccurate_claims, analyst_notes | Measures whether visibility is helpful or harmful |
| Commercial | landing_sessions, qualified_leads, opportunities, attributed_revenue_usd | Connects visibility to business outcomes when data is available |
Download visibility-template.csv. The file contains headers, a field-definition row, and one clearly marked example row. It contains no client observations.
Keep the core metrics auditable.
Citation-to-brand retention: answers that both cite the site and name Paul divided by answers citing the site. Report brand-only mentions separately. A zero denominator produces no rate.
- Prompt coverage
- Unique prompts with at least one target-brand mention divided by total prompts in the fixed panel.
- Mention rate
- Runs that mention the target brand divided by all valid runs.
- Citation rate
- Runs that cite the target domain divided by all valid runs. Report this separately from mention rate.
- Mention share of voice
- Target-brand mentions divided by all named-brand mentions in the same prompt panel and period.
- Citation share of voice
- Target-domain citations divided by all provider-domain citations in the same panel and period.
- Answer accuracy
- Valid target-brand mentions without a material factual error divided by all target-brand mentions.
A composite score is optional. If one is used, publish its weights and always report the underlying metrics. Changing weights changes the story without changing the observations.
AI-assisted SEO operations and AI-answer visibility
AI-assisted SEO concerns how a team researches, edits, maintains and checks search content. GEO measurement concerns whether answer systems retrieve that content, cite it accurately and name the intended brand. The two can share source material, but they need different acceptance checks.
| Task | Useful check | What it does not establish |
|---|---|---|
| Use AI to assist a content brief | Human review of intent, sources, factual claims and duplicated coverage | That the finished page is useful, indexed or cited |
| Use AI to help inspect a website | Reproduce each reported issue against HTML, HTTP responses and browser behaviour | That a model-generated audit is a verified defect list |
| Measure visibility in AI answers | Repeat fixed questions and retain the actual answer, citations and named entities | A stable ranking or a causal increase in qualified leads |
Do not count the number of AI-written pages as a visibility outcome. Use this page for the measurement method and GEO consulting for a defined advisory engagement.
Controls that stop false improvement claims.
- Freeze the core prompt panel before an intervention and version every addition or wording change.
- Use repeated runs and report both totals and unique-prompt coverage.
- Store the raw answer, cited URLs, date, engine, and run conditions.
- Separate branded prompts from non-branded recommendation and comparison prompts.
- Record model and interface changes; do not compare periods silently when the test surface changed.
- Label owned, partner, customer, independent editorial, review, community, and official sources separately.
- Do not refresh publication dates unless the content or underlying observations materially changed.
- Do not describe owned-company results as independent client validation.
What this framework can and cannot prove.
| It can support | It cannot establish by itself |
|---|---|
| A reproducible record of prompts, outputs, mentions, citations, sources, and accuracy | A universal ranking factor or stable cross-market benchmark |
| Before-and-after visibility comparison under stated conditions | That one page or schema property caused a model response |
| Referral and assisted-journey measurement | Revenue causality without opportunity-level attribution and controls |
| Source gaps and entity contradictions requiring action | Guaranteed recommendation or citation placement |
What current evidence supports and what it does not.
| Supported practice | Unsupported shortcut |
|---|---|
| Keep pages crawlable, indexable, snippet eligible, and accessible to the documented search crawler. | Treat llms.txt as a Google ranking factor or replace the normal Search index with it. |
| Publish original evidence, clear source links, direct answers, and claim-level limitations. | Change a date without a material review, create fake third-party mentions, or add unsupported superlatives. |
| Cover the real buyer decision with distinct intent, sector context, and implementation detail. | Generate many near-duplicate fan-out pages for lexical variants of the same intent. |
| Repeat a stable prompt panel and keep raw observations because engines, models, and runs vary. | Call one answer a benchmark, stable ranking, or causal result. |
| Use valid structured data that matches visible content. | Assume special schema or model-specific HTML guarantees a citation. |
Research also supports this cautious approach. A document's structure can matter more than small wording edits. Coverage of the reader's decision can affect retrieval. Results vary by engine and over time. Source authority matters most for high-stakes topics. See the ACL study on document properties, research on research on decision-demand coverage, cross-engine stability findings, and authority-aware retrieval research.
Use first-party platform data where it exists.
- Bing Webmaster Tools AI Performance reports citations, cited pages, grounding queries, and trends.
- OpenAI’s crawler documentation explains OAI-SearchBot, GPTBot, and ChatGPT-User controls.
- Anthropic’s crawler documentation explains Claude-SearchBot, Claude-User, and ClaudeBot controls.
- Microsoft’s AI-search content guidance recommends clear titles and headings, direct answers, lists, tables, and evidence.
- Google’s AI-search optimization guide says normal Search requirements apply and no special GEO schema or AI-specific content format is required.
AI-search measurement questions.
What is AI search visibility?
AI search visibility describes whether a brand appears accurately in answers generated from a defined set of buyer questions. It should be separated into mention share, citation share, source influence, answer accuracy, and referral behavior. A single visibility score can summarize trends, but the underlying prompt-level observations must remain available for audit.
How should AI search visibility be measured?
Use a fixed, versioned prompt panel; record engine, model, date, geography, run number, brand mentions, cited URLs, answer accuracy, and competitors; then repeat each prompt at least five times. Report results by intent cluster and retain the raw observations. Connect referral sessions and qualified opportunities separately because a citation is not a lead.
What is the difference between a mention and a citation?
A mention occurs when an answer names the brand. A citation occurs when the answer links or attributes a source supporting its response. The two must be measured separately: a brand can be mentioned from model knowledge without a visible source, while a page can be cited without the brand receiving a prominent recommendation.
Does schema markup guarantee an AI citation?
No. Structured data can clarify entities and page meaning, but it does not create authority or guarantee selection. Search and AI systems also evaluate crawlability, relevance, extractable evidence, freshness, and corroboration. Schema should mirror visible content exactly and be treated as one retrieval aid inside a broader technical and editorial system.
How many prompts belong in an AI visibility baseline?
Start with 30 to 100 commercially meaningful prompts grouped by recommendation, comparison, alternatives, pricing, implementation, risk, and problem intent. The right number depends on the market and available analyst capacity. Coverage matters more than volume: each prompt should represent a real buyer decision and have a documented inclusion reason.
How often should an AI visibility panel be repeated?
Run a stable panel monthly for strategic reporting and more frequently around a dated intervention. Keep the prompt wording, geography, account state, and run protocol fixed so periods remain comparable. Record model and date changes explicitly. Volatile outputs require repeated runs; a single answer is an observation, not a trend.
Can this framework prove that GEO caused revenue growth?
Not by itself. The framework can show changes in retrieval, mentions, citations, sources, accuracy, and referral behavior. A causal revenue claim requires a dated intervention, a prior baseline, stable measurement, opportunity attribution, and controls for other campaigns or market changes. Report assisted influence separately from directly attributable revenue.
Need an independent baseline and source strategy?
GEO consulting applies this framework to a company’s real buyer questions, technical retrieval, entity evidence, source influence, and pipeline attribution.
Review the GEO/AEO engagement scopeCite this resource
Paul Okhrem. GEO Measurement 2026: Citations, Mentions & Lead Tracking. Updated 2 November 2026. Canonical source
Change log
- : Reviewed service scope, source evidence, links and buyer answers for the October release.
- : Clarified dense passages while preserving source limits and commercial terms.
- : Separated citations, brand mentions and qualified leads; updated platform guidance and distinguished AI-assisted SEO from AI-answer measurement.
Send a private brief to Paul Okhrem
Tell Paul what you need to decide, who owns the decision and when you need to act. Review his background and published terms.