Clickstream data puts live retrieval on about a third of prompts, which changes where your visibility work pays off.
Retrieval is a decision the system makes per prompt
ChatGPT answers some questions from the model’s own weights and others by fetching live pages. That difference decides whether your site can be cited at all.
When retrieval does not fire, the answer draws on training data. You get described by whatever the model absorbed about your company months or years ago.
The share of prompts that trigger a search
Semrush analysts, working from over a billion lines of US clickstream data, measured web search firing on 34.5% of ChatGPT queries in February 2026, down from 46% in late 2024, according to the company’s ChatGPT search insights.
Across the full period they studied, the rate moved between 15% and 66.3%. OpenAI adjusts the trigger, so treat any single figure as a snapshot rather than a constant.
Most prompts are not keywords
The same researchers compared ChatGPT prompts against a 27 billion keyword database. Between 65% and 85% of prompts matched nothing in it.
People type things into an assistant that do not appear in a search box. Your keyword report tells you little about the prompts where your category comes up.
Prompts are drifting toward search behavior
Average length of search-enabled prompts rose from 4.7 words to 8.7 words between the opening months of 2025 and the same months of 2026.
Over a shorter window, the share of prompts using traditional search language climbed from 18.9% in October 2025 to 34.9% in February 2026.
Users are learning to treat the assistant as a research tool. Longer prompts also give the system more to retrieve against.
More referrals, fewer domains sharing them
ChatGPT outbound referral traffic grew 206% year over year through January 2026.
The count of unique domains receiving any referral peaked near 260,000 in October 2025, then fell to about 170,000 by February 2026. The top 10 domains took a little over 30% of referrals in February 2026.
Total ChatGPT traffic plateaued around 1 billion monthly visits from November 2025 into early 2026. The referral growth came from heavier linking inside a flat user base, concentrated on a shrinking set of sites.
Reading this from an agency vantage point
Subscribe PR works on this problem for clients from Phoenix, and the split between the two answer modes is the thing worth watching.
A prompt that triggers retrieval rewards pages that are current, specific and easy to quote. A prompt answered from the model’s memory rewards everything that shaped the training corpus: press coverage, directory records, forum threads, documentation other people wrote about you.
Those two surfaces respond to different work, and most teams fund only the first.
Two checks worth running this week
Ask the assistant your three highest-intent buyer questions and note which answers come back with live links attached. Those are the prompts where a page of yours can compete.
Then search your server logs for OAI-SearchBot, which OpenAI documents as the crawler that surfaces sites in ChatGPT search. If it has not visited, retrieval has nothing recent of yours to pull.
Where the leverage sits
The retrieval rate will keep moving, and you have no say in it. You do have a say in whether your pages are worth retrieving, and in whether the rest of the web describes you accurately on the prompts where retrieval sits out.
A free AI visibility audit will show which of your buyer questions currently return your name in live assistant answers.
Subscribe PR is an answer engine optimization and digital PR agency in Phoenix.
