How to Measure GEO and AI-Search Visibility Without Guessing

To measure AI-search visibility, stop collecting screenshots and run a fixed, dated prompt panel. Test the same prompts on the same engines on a regular schedule, record whether your business is mentioned or cited, which page and what claim, check the claim's accuracy, and combine the results with first-party search and website data.

A screenshot showing your brand in one AI answer is not a GEO report. Generated answers vary by engine, prompt wording, location, date and sometimes from one run to the next. A single good result tells you that it happened once.

Build a stable prompt set#

Write prompts across four intent groups. Using Pivra as the worked example:

• Category: "AI chatbot for an Australian plumbing business."
• Problem: "How can a tradie capture website enquiries after hours?"
• Comparison: "Pivra versus Chatbase for an Australian service business."
• Implementation: "How do I qualify service area in website chat?"

Use about ten prompts per priority market. Keep two sets:

• A fixed set, worded identically every time, for measuring trends.
• A rotating set for discovery, so you notice new questions and new competitors.

Test the same engines on roughly the same day each month, signed out where possible, and note whether web search was enabled. Record the country you tested from, because answers differ by location.

Record six fields per result#

The claim and accuracy fields are the ones most teams skip, and they matter most. A favourable mention with the wrong price is a brand problem, not a win. A citation to an outdated article is a maintenance problem: update or retire that page.

Treat visibility as a distribution, not a rank#

There is no single "AI ranking". A recent critical survey of GEO research stresses that retrieval, citation, prominence, absorption into the answer and resulting traffic are distinct things, and results vary by engine and prompt (critical survey of GEO research). Earlier academic work reports that source citations, statistics and clearer structure can improve visibility in generated answers in experimental settings (GEO: Generative Engine Optimization; structural feature engineering for GEO). These are experimental findings, not a reliable formula for any individual site.

So report rates across the panel ("mentioned in this many of the fixed prompts this month") rather than a position, and compare months, not individual runs.

Use first-party data alongside the panel#

The prompt panel shows what engines say. Your own data shows what happened:

• Search Console shows impressions, clicks and queries from Google Search. Google has also announced controls and reporting for its generative AI features as they roll out (search generative AI control; new controls and insights for website owners).
• Website analytics show referrals from AI products where the referrer is passed.
• Lead records show whether those visits turned into enquiries.

Do not infer visits from screenshots. If an engine cites you but your analytics show no referrals, report exactly that.

A one-page monthly output#

Keep the report short enough that someone reads it:

• Prompts tested, engines, country and date.
• Mention rate and citation rate on the fixed set.
• Accurate-claim rate.
• Top cited pages.
• Incorrect or outdated claims found, with the page that needs fixing.
• Visits and enquiries attributable where measurable.
• Next content fixes, each with an owner.

The "next fixes" line is the point of the exercise. Measurement that does not change a page is overhead.

Common mistakes#

• Changing the prompts every month, which makes trends meaningless.
• Testing only branded prompts. Of course the engine mentions you when you name yourself.
• Celebrating a mention without checking the claim.
• Mixing engines and countries in one number.
• Treating llms.txt files or manufactured mentions as shortcuts. Google advises prioritising helpful, original content over tactics like these (optimising for generative AI features).

Where your own pages fit#

The fastest way to improve the accuracy field is to keep your own pages current, because they are what engines read. Clear answer pages, accurate pricing on the pricing page and honest comparison pages give an engine correct facts to work from. If you also run a website chatbot trained on those pages, fixing a page fixes the chatbot's answer too.

For the broader picture of what AI search does and does not change, see AI SEO for small business.

Next step: replace random checking with a dated, repeatable 30-prompt panel, and run the first round this month.

Sources#

• Optimizing visibility in generative engines: a critical survey (arXiv)
• GEO: Generative Engine Optimization (arXiv)
• Structural feature engineering for generative engine optimisation (arXiv)
• Search generative AI control, Search Console Help
• New opportunities, control and insights for website owners, Google
• Optimising your website for generative AI features on Google Search