The search engine optimization business has spent the previous couple of months taking ChatGPT aside to see the way it searches and finds its sources.
We needed to see for ourselves how ChatGPT retrieves data—and what meaning for AI visibility—so we analyzed 66.6 million free, logged-out GPT-5.6 Luna responses and their 15 million fan-out queries from August and September.
Logging out additionally strips out chat historical past and reminiscence, so we see ChatGPT’s default conduct moderately than solutions formed by one particular person’s previous chats.
Key findings
- Solely 14.4% of the 66.6 million responses required an internet search.
- Once they did, ChatGPT ran about 1.6 fan-out queries on common.
- 6.8% of fan-out queries (i.e. 6.8% of 14.4%) used a
web site:operator to look a single web site, unfold throughout ~116K totally different domains. - 75.9% of the commonest
web site:searches pointed at a model web site, versus retailers, authorities websites, boards, or evaluation websites. - The phrase “official” appeared in 13.1% of English fan-out queries in September, up from 10.1% in August.
- “Value” and “evaluation” fan-outs rose as much as 47% in a month.
- In separate quotation information, ChatGPT’s Reddit citations dropped 88% in six months.
- ChatGPT appears to be shifting from websites that summarize data to the sources that produce it.
ChatGPT solutions 6 in 7 prompts with out looking out the online
Solely 14.4% of the 66.6 million ChatGPT responses we tracked triggered an internet search.
When ChatGPT did search, it ran a median of about 1.6 fan-out queries for every response that searched.
Each internet search prices OpenAI cash, so it is sensible for ChatGPT to skip one when it may possibly reply properly sufficient from what it already is aware of.
This matches with WordLift’s research which discovered that free-tier type fashions could lean extra on coaching information to maintain prices down.
Which means, for a lot of prompts, your probabilities of showing rely on what ChatGPT already is aware of about your model, moderately than what it may possibly discover.
76% of ChatGPT’s ‘web site:’ searches goal a model web site
Of the 15 million fan-outs we analyzed, 6.8% used the web site: operator to look a single web site.
And among the many ~160 most-searched websites, three in 4 of these searches focused a model web site.
By “model web site,” we imply an organization web site like apple.com or ikea.com, versus a retailer, authorities web site, discussion board, or evaluation web site.
Retailers like Walmart, Costco and Amazon made up 11.9%, and Reddit and Fb simply 1.9% between them.
The phrases contained in the fan-outs present the identical give attention to official, first-hand data, with “official” and “worth” the commonest after “2026”, and rising by a number of proportion factors month-over-month.
None of this can be a massive shock. In order for you a product’s worth or specs, a model web site is the plain place to test.
When a query has no single proper reply, that’s when ChatGPT is extra more likely to look past official sources.
As an example, in a separate take a look at of 20,000 US prompts, 84.4% of factual questions on pricing, product options or firm particulars cited a minimum of one official supply.
For comparability questions like “A vs B” or “which is cheaper”, that dropped to 68.3%.
The combo of sources modified too.
For factual questions, 63.4% of citations have been to official web sites.
For comparability questions, it was 46.4%, so most citations went to different websites.
So ChatGPT depends most on official sources for information, and leans extra on different websites when it has to weigh one possibility in opposition to one other.
We’ll dive deeper into fan-out question language just a little later.
Make sure that your model web site can reply the factual questions
If ChatGPT goes to run web site:youbrand.com searches in your costs, specs and assist particulars, these information have to be in your web site, in plain textual content, on a web page it may possibly discover.
Client manufacturers make up 8 of ChatGPT’s 10 most ‘web site:’ searched domains
The domains ChatGPT targets most with web site: are massive client manufacturers.
These rankings are, in fact, based mostly on our personal dataset and prompts, however with round 1 million web site: queries behind them, they provide a powerful sense of which domains ChatGPT leans on most.
These are principally manufacturers that promote merchandise, so a model web site is the pure supply for particulars like costs and specs.
Authorities domains additionally function closely.
Gov.br, gov.uk, France’s service-public.fr, the NHTSA, the FDA and the SEC all sit within the high 150, which matches the observations of Lily Ray, Founding father of Algorythmic, that ChatGPT favors official sources for YMYL questions.
Be aware additionally that Reddit continues to be the ninth most web site: searched area.
ChatGPT hardly ever ‘web site:’ searches OpenAI’s companions, however their quotation share has doubled
Of the 17 publishers and platforms that presently have licensing deals with OpenAI, Reddit is the one one within the high 150 most web site: searched domains.
One attainable purpose is that ChatGPT already has its accomplice content material pre-loaded and able to go.
- David McSweeney, founding father of QueryBurst, discovered that some high-authority publications, together with Time, Enterprise Insider and Forbes, have pre-synthesized, page-level summaries that look like dropped straight into ChatGPT’s context.
- Suganthan Mohanadasan noticed a separate retrieval supply in ChatGPT’s visitors known as “labrador”. He describes it as a “licensed tier”, and factors out that a number of of its publishers have signed content material offers with OpenAI.
These web site: searches present what ChatGPT appears up throughout question fan-out, however not what it finally ends up citing.
The cited domains inform a unique story. Accomplice publications are in all places.
In Model Radar, accomplice titles amongst ChatGPT’s 200 most-cited domains greater than doubled their mixed share of citations, from 7.5% in March to fifteen.8% in July—and nonetheless held 12.5% in late September.
“Value” and “evaluation” fan-outs rose as much as 47% in a month
The phrases ChatGPT provides to its fan-out queries trace on the web page sorts it needs to retrieve.
Between August and September, extra of ChatGPT’s fan-out queries have been about costs, product specs, and opinions, and barely fewer have been about definitions.
A achieve of 1 or 2 proportion factors can sound small, however these phrases all began from a low base, and their relative rise exhibits starker development.
As an example, “Value” went from 6.04% to eight.88% of fan-outs, a relative rise of practically half (47%) in a single month.
Including to this, “Critiques” grew by a 3rd, and “specs” entered the highest 50 for the primary time.
On the flip aspect, “Which means” and “definition” every fell by a couple of fifth, although from small shares, in order that’s seemingly a dip moderately than a pattern.
Verify which of your pages match the rising question sorts
You possibly can assess which fan-out question sorts are most distinguished in your business utilizing Model Radar.
Simply head to the AI responses report, filter to ChatGPT and your market, then filter the “Fan-out queries” column for phrases like “worth”, “specs” or “opinions” to see how typically ChatGPT searches for every in your subject.
ChatGPT’s Reddit citations dropped 88% in six months
Alongside the fan-out information, we checked out which internet sites ChatGPT cites.
This attracts on a separate dataset: Model Radar’s ChatGPT citations within the US from March to September 2026, masking 2.4 to three.8 million solutions a month.
Snippet Digital discovered that Reddit citations in ChatGPT dropped 86% in August, and information shared by Aleyda Solís, founding father of Orainti, confirmed social, reference, and information websites dropping essentially the most quotation share between July and August.
Our Model Radar information exhibits the identical drop throughout Reddit, Wikipedia, and YouTube.
In March, about 1 in each 5 ChatGPT solutions cited Reddit, however by late September it was about 1 in 43.
Equally, Wikipedia dropped from 1 in 7 solutions to 1 in 35 over the identical interval, and YouTube from about 1 in 370 solutions to 1 in 500.
These websites are nonetheless amongst ChatGPT’s most-cited domains, however their share has shrunk quick, so if a variety of your AI visibility comes by them, keep watch over how typically ChatGPT cites them in your matters.
ChatGPT is shifting from websites that summarize data to the sources that produce it
The citations Reddit, Wikipedia and YouTube misplaced went to analysis our bodies, regulators, manufacturers and hands-on testers, whereas explainer and aggregator websites fell.
The chart under exhibits among the greatest rank modifications amongst ChatGPT’s 200 most-cited domains, throughout 2.4 million solutions in March and three.8 million in September.
PubMed climbed from sixty fourth to 4th amongst ChatGPT’s most-cited domains and the FDA entered at eighth, whereas well being websites that summarize that analysis, like WebMD and Verywell Well being, fell.
To ensure this wasn’t only a results of new prompts being added to Model Radar, we in contrast ~25,000 prompts that appeared among the many 150,000 highest-volume prompts in each March and September.
Each web site within the chart moved in the identical course on these an identical prompts.
The prompts in any set will have an effect on which internet sites present up, however the sample is constant: ChatGPT appears to be prioritizing authentic sources over websites that summarize them.
In case your content material primarily summarizes data that another person publishes first, ChatGPT could also be extra more likely to cite the unique supply as an alternative.
The pages holding up finest are those solely you’ll be able to publish, like authentic analysis, hands-on testing, and your individual product information.
Tips on how to analyze fan-out queries in your business
Analyzing fan-out queries in bulk exhibits you what ChatGPT checks earlier than it recommends a model in your business, so you realize what content material to create and the place to earn protection.
Pull the fan-out queries in your complete business without delay
A single fan-out question tells you little or no, so it helps to see many aspect by aspect to acknowledge a sample.
In Model Radar, you’ll be able to examine fan-out queries throughout ChatGPT and Perplexity.
Simply head to the AI responses report, add a filter, and enter your subject as immediate or response textual content.
We’ve chosen “trainers”.
Then filter the platform to ChatGPT and the situation to your market, and take a look at the “Fan-out queries” column subsequent to every immediate.
To check them in bulk, filter the report on the fan-out textual content, resembling web site:, “official”, “worth”, “evaluation”, or “finest”. You possibly can then examine them in-platform or export to a spreadsheet.
Once we analyzed “trainers”, Model Radar returned 1,474 US ChatGPT responses with fan-out queries.
“Greatest” appeared 206 instances, “2026” 160 instances, “evaluation” 117 instances, “vs” 105 instances and “official” solely 65 instances.
So for trainers, ChatGPT principally hunts for latest, third-party opinions and head-to-head comparisons, which is a unique image from the general information above.
A operating shoe model would get extra from evaluation protection and comparability pages than from its product pages alone.
That is based mostly on the present prompts within the Model Radar database, however you may also run this evaluation by yourself customized “Tracked prompts”.
Let Ask Ahrefs group fan-out queries for you
Ask Ahrefs can do the grouping and counting that may in any other case take hours in a spreadsheet.
For instance, you’ll be able to simply paste within the immediate under:
Pull the ChatGPT fan-out queries from Model Radar for my tracked prompts/prompts about [your category]. Group them by intent (pricing, specs, opinions, comparisons, how-to), rely the commonest phrases and
web site:domains, and inform me which teams my web site solutions and which it doesn’t.
Then run the identical immediate every month to see which question sorts are rising.
Ask Ahrefs will run a full evaluation, telling you the breakdown of fan-out question by subject, how your web site: presence compares to competitor domains, and advising you on which gaps to fill.
As an example, we have to enhance our opinions if we wish our Model Radar device to be talked about extra in “AI visibility device” associated responses.
How our information compares to different research
Our information agrees with different analysis on the course ChatGPT is heading: towards official sources, web site: searches, and fewer citations.
The primary distinction is scale, since different research discovered much more fan-outs and web site: searches per immediate.
The prompts in our examine are broad.
They stem from actual search demand throughout each subject and nation, whereas many different research used business “finest X” prompts that naturally set off model lookups.
We additionally tracked free, logged-out periods, which search lower than paid or reasoning fashions.
These numbers are a snapshot of 1 mannequin between August and September 2026, and OpenAI modifications how ChatGPT retrieves data consistently.
Should you observe fan-outs often, you’ll be able to spot modifications as they occur, moderately than months later.
Ultimate ideas
ChatGPT’s free mannequin solutions many questions from what it already is aware of.
And, when it does search, it more and more goes to whoever owns the data, just like the model that units the worth.
That leaves two methods into its solutions: be properly sufficient identified that ChatGPT does not must look you up, or be the unique supply it finds when it does.
Pages that solely summarize another person’s work are presently those dropping citations, so the pages value investing in are those solely you’ll be able to publish, like your product information or authentic analysis.
However, keep in mind, this might all shift once more with the following mannequin replace, so preserve a watchful eye by yourself presence in fan-out queries and AI responses!
