April 2026 study: named when asked, missing when not
What seven companies’ AI answers showed in April 2026, and the limits of what a small study on older models can tell you.
What we found
Named when asked about, missing when not
In April 2026 we asked two models a set of prompts about each of seven companies’ categories, from prompts that described a need without naming anyone to prompts about the company by name.
The pattern was the same for all seven. When a prompt named the company, the answers described it. When a prompt described the category or the buyer’s need without naming anyone, none of the seven was named.
In other words, the models knew who these companies were, but didn’t think of them when someone asked for a recommendation. That second kind of question is the one a buyer asks before they have a shortlist.
Known by name is not the same as recommended
Counts of companies, not rates. Each company’s results rest on a few prompts of each kind.
Limits
What this study can and can’t tell you
It shows a pattern worth checking for your own brand. It isn’t a benchmark to measure yourself against.
- Seven companies, chosen by usThey weren’t a random sample, and all sold to businesses. The same pattern may or may not hold in other categories, including retail.
- A few prompts of each kindWith 26 to 30 prompts spread over six kinds, each kind rests on four or five prompts per company. Zero out of five has a likely range of 0% to 43%, so “none named” means “rarely, if at all”, not a measured 0%.
- Two models, now supersededGPT-5.4 Mini and Sonnet 4.6 have since been replaced. Current assistants, and assistants with live web search, may answer differently.
- No ranges or repeat countsThe study didn’t report likely ranges or how many times each prompt was asked, so treat its results as indicative.
How we measure now
This study isn’t comparable with Persipica’s results
Persipica measures differently, so its results shouldn’t be read against this study’s.
Persipica sends 27 prompts by default to five current assistants (ChatGPT, Claude, Gemini, Perplexity and Grok) and to Google’s AI answers (AI Overviews and AI Mode), three times each on the chat assistants, with live web search on. It counts the mention rate only on prompts that don’t name the brand, gives every rate a likely range across prompts, reports how many answers arrived, and labels matches it hasn’t verified.
How Persipica measures sets out the method, and the sample report shows the result for a made-up brand.
Keep reading
Next in the library, and two more that go with this one.
How AI assistants choose which brands to recommend: a guide to GEO
What GEO is, how ChatGPT, Claude, Gemini, Perplexity, Grok and Google’s AI decide which brands and products to name, and how to measure it honestly.
13 min read GuideHow to track AI brand mentions
Which prompts to ask, how many times to ask them and how to read the result, so a change in how often AI names you is a real change and not noise.
4 min read ToolWhat AI recommendations could be worth: a calculator
Put in your own numbers from your analytics and see what a change in how often AI names you could be worth each month, with the assumptions spelled out.
3 min readSee where AI names you, and where it doesn't.
Persipica asks AI assistants the questions people ask about you, and keeps every answer and source behind each number.