How Do I Normalize Visibility Scores Across ChatGPT, Perplexity, Gemini, and Claude?

Why can't you compare raw scores across assistants?

Because each assistant answers differently, and those differences move the numbers even when your standing is the same:

• List length. One assistant may name three brands per answer and another seven. A brand appears more often in longer lists without being any more favoured.
• Naming versus citing. A study of AI answers by Semrush and Growth Memo found that when a brand's domain appeared in a Gemini answer, Gemini named the brand 83.7% of the time but linked its site only 21.4% of the time. ChatGPT cited them in 87% of appearances but named them in only 20.7%. A citation-based score and a mention-based score will rank those two assistants in opposite order.
• How often they search. Assistants that search the web more often lean on current pages; others lean on training data. The mix changes over time.
• Answer format. Some answers are tables or shopping cards, others are prose with a single recommendation.
• Audience size. Pew's 2026 survey found 44% of US adults use ChatGPT and 24% use Gemini. Being named in each is worth different amounts.

So a 40% appearance rate on one assistant and 30% on another does not mean you are stronger on the first.

Step 1: How do you make the measurements comparable?

• Same prompts on every assistant, from a fixed set that reflects real buyer questions.
• Same runs per prompt, enough to keep noise manageable. How many is enough is covered in how many times to test a prompt to verify visibility.
• Same definitions of a mention, a recommendation and a citation, applied by the same labelling method.
• Same window of time, location and language.
• Modes kept apart. Separate search from non-search answers, and shopping results from text, where the assistant has them.
• The same competitor set, with all their name variants, tracked on every assistant.

Step 2: Which normalized measure should you use?

Share of voice is the most widely useful. Pair it with a baseline index for trends. The leader index is the clearest for executives, since "we are at 44% of the leader's visibility on Gemini" needs no explaining.

What does a worked example look like?

Consider an illustrative software company tracking 60 buyer questions, five runs each, 300 answers per assistant, in a category of 20 tracked brands.

The raw rates say Perplexity is the company's best assistant and Claude its worst. Share of voice says the opposite about Perplexity. With a complete brand list, lift is share of voice multiplied by the number of brands, so it gives the same ranking; its value is that "1.14 times chance" is easier to read. Its long lists include the company often, but no more than chance would suggest, while ChatGPT's short lists include it well above chance. The leader index shows the company trails the leader by a similar margin on three assistants, and further on Claude. Each measure answers a different question, which is why a report should show more than one.

Step 3: How do you combine assistants into one number?

Only with explicit weights. Options:

• Audience weights, from usage surveys or traffic data for your market.
• Buyer weights, from asking customers which assistants they used before buying. This is often the best choice for B2B.
• Revenue weights, from AI referral conversions per assistant, where tracking is good enough.
• Equal weights, only if you have no better information, and say so.

In the example, weighting share of voice 50% ChatGPT, 20% Gemini, 15% Perplexity and 15% Claude, based on a customer survey, gives a composite of 6.8%. Equal weights give 6.4%. The difference is small here, but it can be large when a brand is strong on a little-used assistant. Write the weights into the report, and keep them fixed for a year unless the evidence changes.

Step 4: How do you show uncertainty?

Each assistant's figure has its own sampling noise, and the composite inherits all of it. Report the number of answers behind each figure. To estimate a range for the composite, resample prompts: draw your prompt set with replacement many times, recompute each assistant's share of voice and the weighted composite, and take the middle 95% of results. Measuring run-to-run noise directly is covered in how to measure the noise floor for ChatGPT citation tracking.

What mistakes should you avoid?

• Averaging raw rates across assistants as if they were on the same scale.
• Mixing mentions and citations in one score, given how differently assistants handle them.
• Using different prompt sets for different assistants, often because a tool tracks them differently.
• Trusting a vendor score without knowing its formula. Ask what goes into it and whether it is normalized by list length. Why tool scores often leave you without a next step is covered in why GEO tools show visibility scores but not what to fix.
• Ignoring mode mix, when one assistant searched far more often this month than last.

What do people ask about comparing AI visibility across assistants?

Should I include an assistant few of my customers use?

Track it if it is growing or relevant to a segment, but give it a small weight in any composite. Report it separately so its movements do not distort the headline.

Can Google AI Overviews be normalized the same way?

Yes, with one extra step. AI Overviews appear only on some searches, so report how often an Overview appears for your questions, then your share of voice within the Overviews that appear.

How often should the weights be updated?

Once a year, or when usage shifts sharply. Changing weights often makes trends impossible to read.