What changed
On October 12, 2026 Ahrefs published a proprietary study of 66.6 million free logged-out GPT-5.6 Luna responses and 15 million fan-out queries collected in August–September. It found 14.4% of responses used web search, averaging 1.6 queries among responses that searched. The site: operator appeared in 6.8% of fan-out queries, with official brand domains prevalent among the most frequently searched sites. A separate US citation sample showed large declines in Reddit citations and greater representation of original information sources.
Why it matters
Publishers and SaaS operators should not assume every AI answer fetches live pages, or that generic content-optimisation tricks drive citations. Accessible first-party pricing, specifications, original research and product documentation are more defensible investments, but this one free logged-out model sample cannot be generalized to all paid, logged-in or thinking workflows.
Most responses in this sample never searched
Ahrefs measured web search in 14.4% of 66.6 million responses. The 1.6 average applies only to responses that searched. The 15 million fan-outs are queries, not people or referrals; this is a proprietary sample of a single model and mode.
The queries increasingly seek first-party facts
Of all fan-outs, 6.8% contained site: operators. Among the most frequently searched domains, brand sites predominated. The term price rose from 6.04% of August queries to 8.88% in September. That is observed query behaviour, not proof that every first-party page will rank.
A separate citation dataset finds large shifts
Ahrefs' US Brand Radar citations across 2.4–3.8 million monthly answers suggest Reddit citations fell 88% between March and September, while research bodies, regulators, brands and hands-on testers gained. A roughly 25,000-prompt overlapping subset preserved the directional change, but causation is not established.
Independent trackers observe different averages
Promptwatch independently tracks fan-out and reported around one to two queries per response in parts of 2026, but other studies using commercial or reasoning prompts find many more. Sampling, model version and logged-in state matter; cross-study figures are not interchangeable.
What builders should measure
Keep first-party facts, current pricing, documentation and original evidence discoverable. Track representative commercial prompts and real referrals, not citation counts alone. Treat changes in AI-search behaviour as time-sensitive and model-specific.