Ahrefs Finds Web Search in About 1 in 7 Free ChatGPT Answers

Ahrefs found that web search ran in 14.4% of 66.6 million free, logged-out ChatGPT responses, according to research released October 12. A SaaS team using citation scores to assign content work needs that boundary. Each score describes a particular set of answers, with a particular opportunity to consult the web.

The study captured GPT-5.6 Luna responses during August and September, using broad prompts across topics and countries. About 1 in 7 answers triggered search. The finding doesn’t establish the search rate for your buyers’ product questions, paid accounts, or personalized conversations.

Louise Linehan, a content marketer at Ahrefs, gives a practical recommendation its own subheading in the study.

“Make sure your brand site can answer the factual questions”

That advice gives the website owner a concrete job. Publish the information the company controls, and make it available when an assistant checks. The search rate gives the reporting owner a separate job: disclose which answers the dashboard measured.

An answer, a search, and a citation count different things

OpenAI’s documentation describes an assistant that can search automatically or use search when a person selects it. The service rewrites questions into targeted queries, and it can issue further searches after reviewing results. These background searches are commonly called query fan-out.

Three events sit along that process. The assistant generates an answer. It sometimes retrieves information. It sometimes credits a source in the finished response. A brand mention belongs to the answer; a fetched page belongs to retrieval; a link supporting a claim belongs to citation.

One total can’t explain all three.

Imagine an internal reporting template with columns for whether the brand appeared, whether search ran, and whether a company page was cited. Those are proposed fields, not results from an experiment. They let a marketer distinguish absence from an answer, absence from the retrieved pages, and absence from the final sources. Each diagnosis points to a different investigation.

Our earlier reporting on AI visibility and referral traffic covered another boundary: appearing in an answer doesn’t establish that a customer visited. This new study adds a step earlier in the process. Before interpreting the citation count, establish how the answer was collected.

The prompt set changes the result

An earlier practitioner investigation makes the sampling problem tangible. Suganthan Mohanadasan, co-founder of Snippet Digital, published account-level research in August, including citation observations captured July 24-25. The work used a logged-in Plus account in Dubai and leaned toward software and AI tools. The investigation predates Monday’s Ahrefs release.

Mohanadasan writes in the caveats:

“A different query mix moves the numbers.”

Their captures showed brand names entering search queries before pages were fetched, while the final answer still depended on subsequent retrieval. The revised article explicitly acknowledges that search can change the outcome. This supports inspecting the steps in your own buyer questions; it doesn’t supply a population-wide rule about which brands will win.

The researcher also sells related tools and agency services. Ahrefs promotes Brand Radar in its study. Both disclosures matter when the practical conclusion points toward measurement work those businesses provide.

A separate Nectiv analysis used roughly 4,000 prompts and GPT-5.6 Sol through an API, the developer interface. It recorded 7.61 background searches per prompt. Comparing that directly with a logged-out consumer sample would mix model, collection method, and question set. Their disagreement gives a marketer a reason to request the setup, rather than average the numbers into a benchmark.

Keep the buyer question attached to the score

For a lean marketing team, the useful next step is a narrower reporting brief. Choose questions connected to the sale, preserve their wording, and record the model, access tier, location, and search setting alongside each capture. Repeat the same set so a change in the questions doesn’t masquerade as a change in visibility.

Separate category discovery from product verification. A question about which software to consider tests whether the brand enters an answer. A question about an integration or current price tests whether the assistant can find and use the company’s facts. Neither result, by itself, proves pipeline impact.

Site work still earns a place. OpenAI lists crawl access as an eligibility requirement, while warning that placement isn’t guaranteed. Our coverage of agents struggling with B2B pricing examined why that information needs attention. Keep current facts available, check the answers customers receive, and investigate errors before adding more pages.

The Ahrefs sample doesn’t establish a content budget split or a universal ranking tactic. It strengthens the case for making the sample visible in the AI content and brand brief. Require that disclosure before a citation score starts deciding where the next week of marketing work goes.