We split 1,235 tracked questions by where they came from
Which questions should you track to measure your brand in ChatGPT? Most guides answer with where to find them: your keywords, Google's People also ask box, Reddit, your own pages. Few check whether the source changes the result.
Proofsource can check, because each question it tracks records its origin. Some are real searches people type into Google, kept word for word: a keyword with its monthly volume, or a People also ask question. The rest are written by a model from the brand's own website, in the words a buyer of the category uses.
We took the brands that had both kinds of question and no brand name in the question. That gave 105 brands in 57 workspaces, 379 real searches and 856 written questions, answered by ChatGPT, Claude, Perplexity and Google AI Overviews between October 16 and November 9, 2026. We dropped answers repeated word for word, which leaves 9,329.
Written questions named the brand in 7.7% of answers, real searches in 5.2%
Written questions named the brand in 457 of 5,970 answers, 7.7% (7.0% to 8.4%). Real searches named it in 174 of 3,359, 5.2% (4.5% to 6.0%). Counted by question, 24.5% of written questions named the brand at least once, against 19.0% of real searches.
That's about half as much again for the questions the brand's own site supplied.
46 of 61 brands did better on their written questions
A pooled rate can come from a few large brands, so we compared each brand with itself. 44 of the 105 were never named on either kind of question. Of the 61 named at least once, 46 did better on their written questions, 14 did better on real searches and 1 came out even.
A real search was 5 words long. A written question was 15
The median real search had 5 words. The median written question had 15. The extra words are where the brand comes in.
One lender's real searches read like "What is the best gold loan?" and "Loan against gold rate of interest?" The questions written from its site named its home state and the features its homepage leads with. Each added detail narrows the field toward the company whose copy supplied it.
That isn't wrong. Some buyers do ask that way. But a written question measures how you do with buyers who already describe your niche in your words, and a real search measures how you do with everyone else.
ChatGPT showed the widest gap
Brand named, real searches against written questions: ChatGPT 7.0% against 12.7%, Google AI Overviews 4.8% against 6.1%, Claude 4.0% against 5.6%, Perplexity 4.0% against 4.4%.
What people type into ChatGPT sits between the two
Neither set is exactly what buyers type into an AI assistant. Semrush's study of US clickstream data from October 2024 to February 2026 put the average ChatGPT prompt in early 2026 at 8.7 words when ChatGPT searched the web and 13.5 words when it didn't. For most of the period, 65% to 85% of prompts matched no keyword in its database of over 27 billion.
So a 5-word Google search is shorter than a typical ChatGPT prompt, and a 15-word written question is longer than most. One flatters you a little and the other undersells you a little. A single figure over a mix of the two hides which one you're looking at.
What we'd do
- Report real searches and written questions as two numbers. If the mix changes, a combined rate moves even when no answer did.
- Read the real-search number as your floor with buyers who haven't found you yet, and the written number as your standing with buyers who already describe your niche.
- Where a written question names you and the plain search on the same topic doesn't, read the plain answer. The names and pages it uses are what you're up against.
- When you turn keywords into long questions, watch which details you add. Each one taken from your own copy lifts your number.
How Proofsource handles it
Each question in Proofsource says where it came from: "Searched on Google" with its monthly volume, "Asked on Google" for People also ask, or "Written from your website". Each topic sums up its mix, so you can see which kind of question a number rests on.
Where this could be wrong
The brands are the ones our users chose to track, many of them small and many in India, and most had only a few questions of each kind. Real searches were kept only after a check that they fit the brand, so they may sit closer to the brand than search as a whole.
The two kinds of question don't cover identical topics, and answers to one question aren't independent of each other, so the intervals above are narrower than they should be. This is under four weeks and four engines.
