Ahrefs finds only 14.4% of 66.6M ChatGPT replies triggered a web search

What the study found

Ahrefs analyzed 66.6 million responses that ChatGPT produced in free, logged-out sessions during August and September 2026, and found that just 14.4% of them involved a web search. The remaining 85.6% were answered purely from the model's training data. The analysis was published on October 12, 2026, and written up on dev.to.

The dataset covered the GPT-5.6 Luna model across all countries, along with the 15 million follow-up search queries those responses generated. Ahrefs supplemented it with citation data from its Brand Radar product, which tracked the domains ChatGPT cited in US answers from March through September 2026, at a volume of 2.4 to 3.8 million answers per month.

What ChatGPT does when it does search

According to Ahrefs, searching on ChatGPT is narrow rather than exhaustive. Responses that triggered a search averaged about 1.6 fan-out queries each, which is closer to a targeted lookup than a crawl of the open web.

Only 6.8% of those fan-out queries used the site: operator to restrict results to a single domain. Because that is 6.8% of the 14.4% of responses that searched at all, roughly 1% of all responses in the dataset involved a search confined to one website. Of those single-site queries, 75.9% targeted brand-owned domains — the official apple.com or ikea.com — rather than retailers, forums or review sites.

Queries are shifting toward primary sources

The wording of ChatGPT's search queries changed quickly between August and September 2026. The share of queries containing "2026" rose from 17.6% to 22.9%, "official" grew from 10.1% to 13.1%, "price" climbed from 6.0% to 8.9%, and "india" went from 3.0% to 6.2%. The word "best" barely moved, inching from 5.0% to 5.3%.

The dev.to write-up reads this as a shift away from listicle-style prompts toward primary-source lookups, such as a brand's own pricing page for the current year.

Reddit citations collapsed on ChatGPT

Ahrefs' citation data shows ChatGPT's references to Reddit fell 88% between March and September 2026. That undercuts a staple recommendation in the generative-engine-optimization community: get mentioned on Reddit because LLMs lean on it.

Perplexity tells a different story. Per data from Evertune cited in the write-up, Reddit still accounts for about one in five Perplexity citations, roughly 20%. The two engines are diverging, and a tactic that succeeds on one can be invisible on the other.

The limits of the data

The figures apply only to free, logged-out usage. The write-up cautions that logged-in users on paid plans may trigger searches more often, so 14.4% is best read as a floor rather than a ceiling. Since the free tier is where most users sit, it is still the number that matters for visibility.

The author also recommends auditing whether AI crawlers can read a site at all — robots.txt access, and whether content sits in the HTML or loads behind JavaScript — while noting that no audit can confirm whether content made it into a model's training data; only querying the model itself offers a hint.

Why it matters

If six out of seven ChatGPT answers come from parametric memory rather than live retrieval, the content that counts most is content that was already in the training set. A site that was crawlable and well-linked before the training cutoff is baked into the model; a site launched later, or blocked during training, competes only for the sliver of responses that trigger a search.

Three consequences follow. First, having been crawlable during training matters more than being crawlable now. Second, primary-source pages — documentation, pricing, specifications — match the queries ChatGPT actually issues far better than summary blog posts. Third, no single tactic spans engines: ChatGPT leans on training data and brand sites, Perplexity on live retrieval and Reddit, and Google AI Overviews on their own index and YouTube, according to the write-up. Visibility work now has to be done per engine, and the Ahrefs numbers are a reminder of how small the live-retrieval window on ChatGPT currently is.