HOW TO MEASURE YOUR AI SEARCH VISIBILITY (SHARE OF MODEL, STEP BY STEP)

To measure your AI search visibility, write 10 to 20 golden prompts that mirror the questions your buyers actually ask, run each prompt 3 times per engine across at least 3 AI engines (ChatGPT, Perplexity, Gemini, Claude) with live web search enabled, record every brand mentioned in every answer, then compute Share of Model: your brand citations divided by total category citations, times 100. A Share of Model of 35% or higher signals category leadership; 0% means AI engines do not surface you for any tested buyer intent. Freeze the prompt set once it is written so month-over-month comparisons stay valid, and re-measure monthly, because AI citations decay in roughly 3 months. Track each engine separately: citation overlap between platforms is only about 11%, so a blended number hides the per-engine variance that tells you what to fix. The full process, with prompt design rules and scoring bands, follows below.

WHAT IS SHARE OF MODEL?

Share of Model (SoM) is the AI-era version of share of voice: the percentage of AI-generated answers in your category that mention your brand. It matters because the traffic behind it is disproportionately valuable. AI-referred visitors convert at 14.2% on average versus 2.8% for Google organic, a 5x premium, and studies across verticals put the range at 4.4x to 23x. Share of Model: The Metric That Replaces SEO Rankings covers why the old dashboard metrics cannot see this. This post covers how to actually produce the number.

STEP 1: WRITE GOLDEN PROMPTS FROM REAL BUYER QUESTIONS

Golden prompts are full natural-language questions, not keywords. "Industrial gasket manufacturer" is a keyword. "Who makes the most reliable gaskets for aerospace fuel systems?" is a golden prompt. Source them from sales call recordings, support tickets, the questions on your quote forms, and what your last five customers say they asked before finding you.

Design rules that hold up in practice:

• 10 to 20 prompts for a baseline; 20 to 50 for ongoing competitive tracking.
• Mostly non-branded. 80% of the set should not contain your name. "Best pain management clinic near Snellville GA" tells you whether you get recommended. "Is [your brand] good?" only tells you whether you exist.
• Include a geographic qualifier where buying is regional, which covers most B2B and virtually all healthcare.
• Cover the journey: problem questions ("how do I fix X"), comparison questions ("X vs Y"), and decision questions ("best X for Y").

STEP 2: RUN EACH PROMPT 3X PER ENGINE, WITH LIVE WEB SEARCH ON

AI answers vary between runs, so a single run is an anecdote, not a measurement. Run every prompt 3 times on every engine and aggregate. Test at least 3 engines; we run 4: ChatGPT, Perplexity, Gemini (the closest proxy for Google AI Overviews behavior), and Claude.

Live web search matters more than most guides admit. With search off, you are measuring what the model memorized during training, which can be a year stale. With search on, you are measuring retrieval-time visibility: what the engine finds and selects today. They are different numbers, and only the second one responds to your optimization work within weeks. A 15-prompt set run 3x across 4 engines is 180 data points, enough to be statistically meaningful.

STEP 3: COUNT BRANDED AND NON-BRANDED MENTIONS

For every response, record: was your brand or domain mentioned (yes/no), at what position (first recommendation or fifth), in what context (recommended, merely listed, or cited as a source), and which competitors appeared. Count competitor mentions with the same care as your own, because they are the denominator of the formula. Expect brand-citation behavior to differ wildly by engine: ChatGPT cites brands in only about 0.59% of responses while Perplexity cites brands in roughly 13%, a 46x gap. Perplexity founder Aravind Srinivas has been explicit about that design:

"Every answer in Perplexity comes with sources from the web."

So a low ChatGPT count is normal; a low Perplexity count is a red flag.

STEP 4: COMPUTE YOUR SHARE OF MODEL AND READ IT HONESTLY

Divide your citations by total category citations and multiply by 100, overall and per engine. Then read the number against these bands, which we calibrated across our own client measurements:

• 0%: invisible. AI engines do not surface you for any tested intent. Most companies we test for the first time are here.
• 5-15%: you surface occasionally, usually for branded queries only.
• 15-30%: meaningful presence, not leadership. Content depth and third-party citations are the usual gap.
• 30-35%: approaching leadership.
• 35%+: category leadership. The engine default-recommends you. Above 40%, your job flips to defense.

STEP 5: FREEZE THE PROMPT SET AND RE-MEASURE MONTHLY

The single most common measurement mistake is editing prompts between runs. Change the questions and you have changed the test; the trend line becomes fiction. Freeze the set after the baseline. If the business genuinely changes (new product line, new market), version the set: keep the original prompts running for continuity and add a clearly labeled v2 cohort.

Re-measure monthly because the underlying data moves fast: AI citations decay once content is about 3 months old, new content can enter citation pools within 3 to 5 business days, and roughly 50% of content cited in AI answers is less than 13 weeks old. The decay mechanics are a post of their own. Monthly cadence catches both your gains and your erosion while there is still time to respond.

WHY PER-ENGINE VARIANCE IS SIGNAL, NOT NOISE

A company at 30% SoM on Claude and 5% on Perplexity does not have "an average of 17.5%." It has two different problems. Claude favors authoritative, technically precise sources; Perplexity is Reddit-heavy; ChatGPT leans on Wikipedia for 26 to 48% of its top citations; Google AI Overviews runs query fan-out and rewards topic-cluster coverage. The fix for a Perplexity gap (community presence, third-party lists) is different from the fix for a Gemini gap (sub-question coverage). We broke down each engine's selection mechanics here. Report per engine, prescribe per engine.

TOOLS VS. DOING IT MANUALLY

You can run this entire process by hand in a spreadsheet: 15 prompts x 3 runs x 4 engines is an afternoon of work, monthly. Tracking platforms automate the runs from $49 per month at the entry level to enterprise contracts. We built our own pipeline that derives golden prompts from a company's site, runs live tests across all 4 engines, and scores the results; it is documented on the methodology page and included in every free scorecard. Examples of what the output looks like are in our work, and ongoing monthly measurement is part of every retainer. Whatever you use, the rules above still apply: frozen prompts, 3x runs, live search, per-engine reporting.