ChatGPT answers 6 in 7 prompts without searching and shifts citations to primary sources, Ahrefs finds

ChatGPT's free, logged-out model answered about six in seven prompts without searching the web, and when it did search it increasingly went to the original source of a fact rather than to sites that summarise it, according to an Ahrefs study published on 12 October 2026. Ahrefs analysed 66.6 million GPT-5.6 Luna responses and their 15 million fan-out queries from August and September 2026, plus separate citation data for US ChatGPT answers from March to September. Only 14.4% of responses ran a web search, and between March and September PubMed climbed from 64th to 4th among ChatGPT's most-cited domains while WebMD fell from 18th to 79th.

The findings matter for anyone whose AI visibility depends on explainer content, aggregators or the big community sites. This article sets out what Ahrefs measured, the numbers behind each claim, how the study was built, what its limits are, and what has changed since Relevant Audience last covered ChatGPT citations.

What Ahrefs studied

The Ahrefs study, written by Ahrefs content marketer Louise Linehan, uses two separate datasets, and it is worth keeping them apart because they answer different questions.

The first is a fan-out dataset. "Fan-out" queries are the background searches ChatGPT runs before it writes an answer. Ahrefs collected 66.6 million responses from free, logged-out sessions on the GPT-5.6 Luna model, across all countries, from August and September 2026, along with the 15 million fan-out queries those responses produced. Ahrefs chose logged-out free sessions on purpose: logging out strips chat history and memory, so the data shows ChatGPT's default behaviour rather than answers shaped by one person's past chats.

The second is a citation dataset from Ahrefs' own Brand Radar product: the domains ChatGPT cited in US answers from March to September 2026, covering 2.4 million to 3.8 million answers a month. This dataset is the basis for the claims about which sites gained and lost citations.

Most prompts never trigger a search

Only 14.4% of the 66.6 million responses involved a web search. When ChatGPT did search, it ran about 1.6 fan-out queries per searching response on average. Search Engine Journal, which reported the study on 12 October, noted that OpenAI's own help page says ChatGPT may search the web on its own when a question would benefit from current information.

Ahrefs' reading is that each web search costs OpenAI money, so the free model skips it when it can answer from what it already knows. The practical consequence, in Ahrefs' words, is that for many prompts a brand's chance of appearing depends on what ChatGPT already knows about it, not on what it can find at that moment.

When ChatGPT searches, it goes to brand sites

Of the 15 million fan-out queries, 6.8% used the site: operator to search a single website, spread across about 116,000 domains. Among site: searches aimed at the roughly 160 most-searched domains, 75.9% targeted a brand site, which Ahrefs defines as a company's own website such as apple.com or ikea.com, as opposed to a retailer, government site, forum or review site. Retailers such as Walmart, Costco and Amazon accounted for 11.9%, and Reddit and Facebook together for 1.9%.

The most site: searched domains, based on around one million site: queries, were mostly consumer brands that sell products. The top five are below.

Further down the list, reddit.com was ninth with 2,586 site: mentions, just ahead of gov.br. Government domains appear throughout the top 150, including gov.br, gov.uk, France's service-public.fr, the NHTSA, the FDA and the SEC.

Ahrefs also ran a separate test of 20,000 US prompts to see how question type changes the sources. For factual questions about pricing, product features or company details, 84.4% of answers cited at least one official source and 63.4% of all citations went to official websites. For comparison questions such as "A vs B" or "which is cheaper", those figures dropped to 68.3% and 46.4%. In other words, ChatGPT leaned on official sources for facts and on other sites when it had to weigh options.

Fan-out wording shifted toward prices, reviews and "official"

Between August and September 2026, the words inside ChatGPT's fan-out queries moved toward prices, specifications and reviews. The word "official" appeared in 13.1% of English fan-out queries in September, up from 10.1% in August. "Price" rose from 6.04% to 8.88% of fan-outs, a relative rise of about 47% in one month. "Reviews" grew by about a third, and "specifications" entered the top 50 terms for the first time. "Meaning" and "definition" each fell by about a fifth, though Ahrefs says that from such small shares this is more likely a dip than a trend.

Ahrefs added a caution of its own through an example. When it filtered its data to prompts about running shoes, 1,474 US ChatGPT responses had fan-out queries, and "best" appeared 206 times, "review" 117 times, "vs" 105 times and "official" only 65 times. For that category ChatGPT mostly looked for recent third-party reviews and head-to-head comparisons, which is a different picture from the overall data. Category matters.

Citations moved from explainers to primary sources

The citation data is where the headline shift shows up. Comparing ChatGPT's 200 most-cited domains in US answers in March (2.4 million answers) and September (3.8 million answers), Ahrefs found that government health and research sites, regulators, brands and hands-on testers rose, while explainer and aggregator sites fell.

Wikipedia went from about 1 in 7 answers to 1 in 35 over the same period, and YouTube from about 1 in 370 to 1 in 500. Search Engine Journal gave the percentages as 19.1% to 2.3% for Reddit and 14.1% to 2.9% for Wikipedia. Ahrefs says all three are still among ChatGPT's most-cited domains, but their share has shrunk fast.

The climbers Ahrefs names include government health and research bodies (PubMed, FDA, CDC, MedlinePlus, DailyMed, NIH, NHS), regulators (SEC, EPA, FTC, CFPB, NHTSA, CPSC), brand and support pages (Apple, Google Support, Microsoft Support, Toyota, Lowe's) and hands-on testers (RTINGS, CarComplaints, RepairPal). The sites that lost ground include health and finance explainers (WebMD, Healthline, Verywell Health, Investopedia, Livestrong) and encyclopedias such as Wikipedia and Britannica. Ahrefs' own summary is that ChatGPT "seems to be prioritizing original sources over sites that summarize them."

A shift like this could be an artefact of new prompts being added to Brand Radar over time. To check, Ahrefs compared about 25,000 prompts that were among its 150,000 highest-volume prompts in both March and September, and says every site in its chart moved in the same direction on those identical prompts.

OpenAI's publishing partners

One pattern runs the other way. Of the 17 publishers and platforms that Ahrefs says have licensing deals with OpenAI, Reddit is the only one in the top 150 most site: searched domains, so ChatGPT rarely looks them up directly. Yet in Brand Radar, partner titles among ChatGPT's 200 most-cited domains more than doubled their combined share of citations, from 7.5% in March to 15.8% in July, and still held 12.5% in late September. Ahrefs offers one possible reason, that partner content may already be available to ChatGPT without a search, but presents it as a possibility, not a finding.

What is new since RA's earlier Reddit coverage

Relevant Audience covered one part of this thread earlier, under the headline "Reddit's ChatGPT citation share fell 86 percent and nobody can explain the timing". That piece was about a single platform's citation drop over a short window. The Ahrefs study confirms the direction with a much larger dataset, putting Reddit's decline at 88% over six months, but the bigger news is everything around it: how rarely the free model searches at all, how often it goes straight to brand domains with site: searches, how its query wording moved toward prices and reviews in a single month, and that the citations Reddit and Wikipedia lost went largely to primary sources rather than to other community sites.

The study's method and limits

Ahrefs is open about several limits, and they shape how far the numbers travel.

• It is one model at one time. Ahrefs describes the fan-out figures as a snapshot of GPT-5.6 Luna between August and September 2026, and says OpenAI changes how ChatGPT retrieves information constantly.
• Free, logged-out sessions search less. Ahrefs notes that paid and reasoning models search more, so its 14.4% search rate and 1.6 fan-outs per searching response are not a description of every ChatGPT user.
• The prompts are broad. They come from search demand across every topic and country, while many other studies used commercial "best X" prompts that naturally trigger brand lookups. Ahrefs cites another study of about 4,000 prompts run through the API on GPT-5.6 Sol that found 7.61 fan-outs per prompt, and Search Engine Journal points out that the two are not directly comparable because models, prompts and collection methods differ.
• The citation data is US only, and it measures citation rank, not traffic. Search Engine Journal makes this point: a domain's rank in Brand Radar does not show how many clicks it received.
• The two datasets are separate. The fan-out data is worldwide and covers August to September; the citation data is US and covers March to September. They are not the same sample.

Ahrefs also sells Brand Radar, the product that supplied the citation data, and the article includes instructions for running the same analysis in it. That does not make the numbers wrong, but readers should know the source has a commercial stake in AI visibility tracking.

What to check on your own site

Ahrefs' practical advice is that if ChatGPT is going to run a site: search on your domain for prices, specs and support details, those facts need to be on your site, in plain text, on a page it can find. The study points to a few checks that follow directly from the data.

• Pricing, specifications, policies and support answers should exist as readable text on your own domain, not only inside images, PDFs or marketplace listings.
• If your content mostly summarises research that someone else published first, the data suggests ChatGPT may cite the original source instead. Ahrefs' advice is to invest in pages only you can publish, such as original research, hands-on testing and your own product facts.
• Comparison questions still pull in third-party sites, so independent reviews and comparison coverage matter where buyers weigh options.
• Check which sources ChatGPT cites for your core topics now, and again after each model update, since the pattern could change.

These are the levers that generative engine optimisation work focuses on, and they apply alongside, not instead of, ordinary search visibility.

What this means for Thai marketers

The study does not break out Thailand. The fan-out data covers all countries combined and the citation data is US only, so none of the figures can be read as Thai numbers, and Ahrefs says nothing about Thai-language prompts. The finding that does carry over is mechanical rather than national: when the free model searches, it often runs a site: search on a brand's own domain to find prices and specifications. A Thai brand whose prices, product details or service terms live only in images, chat apps or marketplace listings gives that search nothing to read. Putting those facts in plain text on the brand's own site, in Thai and English where the audience needs both, is the step the data supports. For teams planning this work, ChatGPT SEO starts with exactly that kind of on-site fact coverage.

FAQ

How often does ChatGPT search the web?

In Ahrefs' data, only 14.4% of 66.6 million free, logged-out GPT-5.6 Luna responses from August and September 2026 ran a web search. Searching responses averaged about 1.6 fan-out queries, and Ahrefs notes that paid and reasoning models search more.

What is a fan-out query?

A fan-out query is a background search ChatGPT runs before writing an answer. Ahrefs analysed 15 million of them and found 6.8% used the site: operator to search a single website.

Which sites gained ChatGPT citations?

Primary sources gained, according to Ahrefs' US citation data from March to September 2026. PubMed rose from 64th to 4th, the FDA entered at 8th, and regulators, brand support pages and hands-on testers such as RTINGS also climbed.

Is Reddit still cited by ChatGPT?

Yes, but far less often: Reddit appeared in about 1 in 5 US ChatGPT answers in March 2026 and about 1 in 43 by late September, per Ahrefs. It was still the ninth most site: searched domain in the fan-out data.

Does this apply to Thailand?

Not directly, because the study has no Thailand breakdown. The fan-out data is worldwide in aggregate and the citation data is US only, so Thai brands should treat it as a signal about how the model behaves, not as local figures.

The bottom line

The Ahrefs data points one way: ChatGPT's free model answers most questions from memory, and when it does look something up it increasingly goes to whoever owns the fact. Pages that only restate other people's work lost ground between March and September 2026, and brand domains became a regular target of site: searches. It is one snapshot of one model, and the next update could move things again. If you want to see how your brand shows up in ChatGPT answers and what to fix on your own site, the Relevant Audience GEO team can help.