A widely shared statistic appeared to answer a basic GEO question: how often does ChatGPT search the web? Its publisher has since corrected what the number measures.
Ahrefs reported that 14.4% of 66.6 million free, logged-out GPT-5.6 Luna responses contained fan-out queries. The sample covered August and September 2026. An October correction explains that the original description, the proportion that searched the web, was too broad. The recorded event was a response containing fan-out queries. Those responses averaged about 1.6 such queries. [1]
The correction changes how the number can be used. The remaining 85.6% cannot be classified as answers drawn entirely from model knowledge on the strength of this statistic. Absence of a recorded fan-out is insufficient evidence of absence of retrieval.
Four events that need separate columns
A useful measurement system distinguishes whether a search happened, which queries it generated, which pages it retrieved, and which sources it cited. A dashboard that merges those events can report a precise number while answering the wrong question.
| Observation | What it tells you | What remains unknown |
|---|---|---|
| Search invocation | The system called a search tool | Whether it found or used your content |
| Recorded fan-out query | A query expansion was observed | Whether all retrieval activity was captured |
| Retrieved URL | A page entered the observed retrieval set | Whether the final answer relied on it |
| Cited URL | The answer attached a source link | Whether the citation supports the claim, or anyone clicked |
OpenAI says ChatGPT may search automatically when current information would help, and users can initiate search themselves. Its documentation also describes rewriting a prompt into targeted queries and sometimes issuing more searches after reviewing initial results. That describes the mechanism; it does not provide a platform-wide search-frequency percentage. [2]
Why 14.4% and 89.6% are not opposing answers
AirOps examined 15,000 original prompts and 548,534 retrieved pages. Its report says two or more fan-out queries occurred on 89.6% of searches. The prompt set was divided evenly between eight informational and commercial intent types. It was a designed research sample, not a representative census of everyday ChatGPT conversations. [3]
Putting that percentage next to Ahrefs' figure does not establish a disagreement. The measured event, sample construction, and documented conditions differ. Nor does it establish that the difference is entirely explained by subscription tier or reasoning mode. That would require a controlled comparison.
There is evidence that mode deserves its own reporting dimension. Kevin Indig tested 100 prompts under minimal and high reasoning with GPT-5.2, using Semrush's tooling. The share of responses with at least one external citation rose from 50% to 68%, while recorded fan-out queries increased 4.6-fold. That is a result for those prompts and settings, not the percentage of all ChatGPT sessions that search. [4]
Before adopting any benchmark, write down its denominator in a sentence. Responses with a recorded query expansion divided by all responses in this panel is a usable definition. How often ChatGPT searches still needs a population, an observable event, and a collection method.
Measure the decisions your customers actually make
For a brand, a platform-wide average would still be an incomplete planning input. Someone asking for help rewriting an email is a different commercial opportunity from someone comparing vendors. The relevant question is how the system handles the tasks that could introduce, evaluate, or eliminate your business.
Start with a small, fixed set drawn from sales conversations, site searches, customer questions, and comparison requests. Preserve the wording. Add a few realistic variations as separate prompts rather than silently replacing the originals each week. Record the intended market and language.
Run an automatic-search condition separately from an explicitly requested-search condition. The first explores what happens without intervention. The second explores your visibility when retrieval is requested. Combining them would hide a major difference in the test design.
A practical first batch might use 30 prompts across three intent groups, repeated three times in each of two conditions: 180 observations. This is a proposed diagnostic workload, not a power calculation or a claim that 180 responses represent your market. Keep repeated responses grouped by prompt when assessing uncertainty; they are not independent customers.
A minimum record for each response
- Conditions: date, interface, reported model, account state, search setting, language, and relevant location settings.
- Input: exact prompt and whether the conversation began fresh.
- Retrieval: observed search call, exposed queries, and retrieved URLs where available.
- Answer: full response, brand mentions, cited URLs, and the claims attached to those citations.
- Collection quality: complete, partial, failed, or unavailable telemetry.
Use unknown when the interface does not expose the event you need. A missing trace should not become a zero. Keep failures outside the successful-response denominator and report their count so a broken collector cannot masquerade as declining visibility.
Then calculate three distinct rates: observed search use among responses where search status is known; brand mention rate among completed responses; and citation rate among completed responses. Add a conditional citation rate among observed searching responses if the instrumentation supports it. Publish the counts beneath each percentage.
What changes in the content plan
If relevant prompts search but consistently overlook your pages, inspect the retrieved sources and the questions they answer. A missing product detail, unclear comparison, inaccessible page, or weak supporting evidence gives you a concrete editing task.
If your sources are retrieved but seldom cited, compare what the selected pages contribute. Retrieval alone is not a completed visibility outcome. If search status is unknown, fix the measurement before diagnosing an absence from retrieval.
If citations improve, keep checking whether they describe the business accurately and whether qualified visits or enquiries follow. A larger citation count can be useful without being equivalent to additional revenue.
Before applying a search-frequency benchmark to your content budget, check the counted event, the denominator, and whether the sample includes the questions your customers ask.