How to measure GEO visibility

There is no single GEO ranking

Generative search surfaces provide answers, summaries and source links. The response can vary with phrasing, language, location, time, product mode and the web sources currently available. It does not produce a stable list comparable to a classic position from one to ten.

GEO visibility should therefore not begin with one number. It begins with a named surface and an observable question:

• Can the relevant crawler reach the public site?
• Is a specific page cited as a source?
• Does a traceable referral visit occur?
• Does that visit lead to meaningful use or a business goal?

These levels need different sources and should not be blended into a universal score.

Four levels of measurement

1. Technical access

The first level checks whether relevant content is public, crawlable and indexable. It includes:

• successful HTTP retrieval
• no unintended robots block
• no unintended noindex
• a clear canonical URL
• important information available as readable text
• crawlable internal links
• consistent structured data where it fits the content

Google says that AI Overviews and AI Mode use the same foundational SEO requirements. A page must be indexed and eligible to appear with a snippet. According to Google, no additional AI markup or special file is required to qualify, and none guarantees inclusion.

OpenAI advises publishers not to block OAI-SearchBot when they want content to be discovered, summarized and linked in ChatGPT search. Accessible content still has no entitlement to a citation. Crawl permission is a prerequisite, not an outcome.

2. Citation or source reference

A citation is the visible use of a URL as a source in a generated answer. Record it with the surface, period and URL.

Bing introduced AI Performance in Bing Webmaster Tools as a public preview in February 2026. Microsoft says its reporting includes total citations, average cited pages, grounding queries and URL level activity across supported AI surfaces. Microsoft also states that a citation count does not indicate placement, authority or the role of a page within an individual answer.

Google Search Central says appearances in AI Overviews and AI Mode are included in the general web search type in Search Console performance reporting. There is no separate general GEO position in that report. A rise in web impressions therefore cannot automatically be attributed to an AI surface.

3. Referral traffic

When an answer contains a clickable link and someone visits, analytics may identify the referral. OpenAI explicitly names analytics as a way for publishers to observe ChatGPT referral traffic.

Check:

• referrer or source and medium definition
• landing page
• market and language
• consent and tracking coverage
• target event after landing

Not every citation receives a click. Not every click sends an unambiguous referrer. Referral traffic is its own measurement, not a complete representation of citations.

4. Use and business impact

The most valuable visit is not necessarily the most common one. A small referral stream can lead people to useful documentation, contact or product use. High visibility with a poor landing page may produce little value.

After arrival, measure the same target events used for other channels. Define them clearly. An enquiry is not automatically qualified, and a page view is not a conversion.

Keep a surface specific measurement record

Use a separate record for each product surface.

Field	Example documentation
Surface	Bing Copilot, ChatGPT Search or Google AI Mode
Market and language	United Kingdom, English
Measurement source	Bing AI Performance, analytics or manual observation
Topic	specific question or intent group
Cited URL	complete canonical path
Observation time	date and time zone
Signal	citation, referral, impression or target event
Limit	sample, aggregated data or missing referrer

This makes the supported claim clear. “Bing AI Performance reported citations for URL X” is reviewable. “Our brand leads in AI” is not a technical statement without a defined market and suitable comparative data.

Treat manual prompt panels as samples

When a platform offers no publisher report, teams often repeat a set of test questions. Such prompt panels can provide qualitative clues, but they are not neutral ranking monitors.

Define in advance:

1. real user questions rather than artificial keyword variations
2. language and market
3. exact surface and product mode
4. signed in or signed out state
5. observation time
6. entities, sources and response properties to review
7. unchanged repetition across several points in time

Document the result as a sample. Personalization, model updates and response variation limit comparability. Automated queries must also comply with the relevant platform terms.

One possible measure is the share of panel questions that contain at least one supported citation to the site. It applies only to that defined panel, not to every user question or the whole market.

What makes a useful baseline

A GEO baseline does not combine arbitrary tool scores. It records the starting state at each level:

• crawl and index access for relevant URLs
• existing Bing AI Performance citations where the report is available
• observed referral sources in analytics
• a defined manual prompt sample where needed
• target events on the affected landing pages
• current content and publication date

Capture the baseline before the change. New data sources, tracking rules or prompt panels are method changes and should be marked as such.

Which content changes are technically justified

Google recommends the same foundations for AI features as for search: accessible pages, internal links, good user experience, important information in text and structured data that matches visible content. This is not a GEO trick. It is sound publishing.

Useful qualities for citable subject matter include:

• a clearly answered user question
• precise terms and consistent entities
• visible authorship and date
• primary sources for external facts
• explicit limits and uncertainty
• a durable URL representing the topic
• updates when the product or source changes

Large sets of near duplicate pages, invented studies and text written only for crawlers do not improve subject quality. They make maintenance and trust harder.

Observe after a change

Keep execution and impact separate again:

1. Is the new content live and technically accessible?
2. Was it recrawled by the relevant search systems?
3. Do citation or referral signals change in the defined surface?
4. Do visitors use the page meaningfully?
5. Are there enough observations for a decision?

Generative systems and search indexes change independently of your site. A citation observed after an update is initially a temporal observation. An impact claim requires the scope, alternative explanations and repeatability to be reviewed.

What an honest GEO report contains

A good report names the surface, source, method and limit. It separates:

• technical prerequisites
• directly reported citations
• manual samples
• referral visits
• website goals

It does not invent a universal visibility number when the platforms provide no shared population. That restraint does not weaken the report. It shows which decision the available data can genuinely support.

MarketingOS is intended to preserve that provenance. Evidence remains connected to source and time, an action receives its own approval, and later observation is not inferred from activity.