Ask 20 to 30 realistic buying questions, run each one several times in each assistant, and record how often your brand is mentioned. Report the share as a range, not a screenshot. Then improve what assistants read: clear product data, consistent brand facts, and mentions on the sites they cite.
Why a single answer is misleading
Ask an assistant the same question twice and you will usually get different brands. In a study published in January 2026, SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in total. They reported less than a 1 in 100 chance that two responses would contain the same list of brands, and about 1 in 1,000 that two lists would come in the same order.
That means a screenshot of one good answer proves nothing, and neither does one bad one. The useful number is how often your brand appears across many runs.
A simple method
- Write 20 to 30 questions your customers really ask, such as gift ideas, budget, style, materials and ethics. Cover your category, not just your brand name.
- Run every question at least five times in each assistant your customers use, in fresh conversations.
- For each answer, record whether your brand is mentioned, whether it is recommended, which competitors appear and which sources are cited.
- Work out the share of answers that mention you, per assistant, and report it as a range across runs.
- Repeat every month with the same questions, so that changes mean something.
What assistants read
Assistants answer from a mix of what the model learned and pages they fetch or search for. In an analysis of 75,000 brands, Ahrefs found that mentions of a brand on other sites correlated with its visibility in Google's AI Overviews much more strongly than backlinks did (0.664 against 0.218). That is a correlation and not proof of cause, but it points the same way as the practical advice: be present on the sites that answers are drawn from, such as reviews, lists and comparison articles, and not only on your own.
What to fix on your own store
- Product data that says what things are: metal, stones, dimensions, origin, price and care, in plain text on the page.
- Brand facts that are identical everywhere: what you make, where, for whom, and since when.
- Pages that answer buying questions: sizing, metals, gemstones, gifting and care.
- A fast, crawlable site with real text, not only images or scripts.
- Mentions on other sites: press, reviews, directories and lists that a shopper, and an assistant, would trust.
What is not proven
Structured data (schema markup) is often sold as a shortcut. An Ahrefs study of 1,885 pages that added it, compared with 4,000 control pages, found no meaningful boost to AI citations, and a separate test found that AI systems fetching a page read its visible text and not the markup. It does no harm and helps search engines understand a page, but the words on the page do the work.
One measurement, and why we now repeat it
In July 2026 we measured a Belgian fine-jewelry brand before and after a catalog restructure, and found it returned as a top recommendation in ChatGPT and Gemini for its category. It was one dated measurement, which is exactly why we now run the same questions many times and report ranges.
Questions people ask
Can anyone guarantee my brand will be recommended?
No. Assistants change their answers between runs and between model versions. What can be measured, and improved, is how often you appear.
How often should I measure?
Monthly is enough for most brands, using the same set of questions each time.
Which assistants should I test?
The ones your customers use: at least ChatGPT, Google's AI results and Gemini, and often Perplexity and Claude too.
Do I need a paid tool?
Not to start. A spreadsheet and a consistent set of questions will show you where you stand. Tools help once you want to track many prompts automatically.