On 8 August ChatGPT changed how it searches. You would not have noticed anything in the UI. The change is in how it finds sources.

We track fan-out queries at Obsero, so we saw it happen in the data rather than in a changelog. The first signs filtered through on 7 August. By the 8th it was everywhere.

Answers went from citing about 12 sources to about 60. And a third of the answers we assessed started going directly to specific, named websites.

If you work in PR, GEO or organic growth, this one is up your street.

When you give ChatGPT a prompt, it does not run one search.

It breaks your prompt into several sub-queries, runs each one through a search engine, and pulls documents back from the results.

Ask for "best fitting jeans for women" and it searches price, comfort and fit as separate queries, then gathers sources for each to build the answer. Commerce prompts get their own shopping fan-outs on top of that.

You never see any of this. A brand can still make the answer without one of these searches. What the fan-out decides is which brands get checked directly, and how much ChatGPT can say about them.

OpenAI released GPT-5.6 Sol last week. The announcement talked about "more reliable facts" by "better using the sources it finds". We can go further than timing on this. Every answer we capture records the model that served it, and the behaviour splits on the version. The answers before are 5.5. The answers after are 5.6.

Two things changed.

Sources per answer went up roughly 5x. From around 12 to around 60.

And the site: operator started appearing in fan-outs. SEOs have used site: for years to check which pages Google has discovered on a domain. (Before any SEO corrects me, that is not the same as what is indexed. Search Console is the place for that.)

Before 8 August, ChatGPT 5.5 barely touched it. Since then it turns up in 37% of answers that have fan-outs.

In plain English: ChatGPT went from taking whatever the open web handed it to working off a shortlist of sources it already trusts, and going to them by name. That trust is decided upstream, in the training data.

Remember: this is ChatGPT only. We ran the same prompts across Google AI Overviews, AI Mode, Gemini, Perplexity and Copilot over the same window. Source numbers were flat. Nothing moved anywhere else.

Not site:, as it turns out.

Answers using site: do cite about 50% more sources. But site: is not the cause. Those answers simply run more searches, and more searches means more sources.

The real story is the shortlist. ChatGPT can only write "site:yourbrand.com" if it already knows and trusts your domain.

That is the big shift. Retrieval used to be a discovery issue. Now, on the prompts that matter commercially, it is an entity association one.

We ran this on Levi's. Over 1,500 fan-out queries across 134 jeans prompts.

The site: behaviour is almost entirely a buying-mode thing.

Those percentages are blended across the whole window, which includes three days before the change. Narrow it to "best X" prompts either side of the 8th and it goes from 1.9% of answers to 36.4%.

Style advice stays on the open web. Buying prompts get a dedicated and trusted shortlist.

The brands ChatGPT checked directly: Levi's, Madewell, Abercrombie, Good American and Wrangler.

The publishers it checked directly: The New York Times, Good Housekeeping, Vogue and Who What Wear.

If you know which sites ChatGPT targets through fan-outs, you have your new media hit list.

On buying prompts specifically (where to buy, who sells X near me, how much are X jeans) site: usage went from nothing before 8 August to 11% of answers. Sources nearly doubled.

And these are not review sites like Trustpilot. They are brand sites. Real examples using Obsero data.

ChatGPT is checking your prices, your stock and your store locator directly. If that information is buried in JavaScript, it cannot hand it to a potential customer.

Which means you can lose a potential customer without ever knowing it happened. There is no bounce rate to look at, because nobody got as far as your site. If your own assets are not optimised for this, you are not in the answer, and this is the sort of thing the fan-out data is for.

If you do PR. The value of that publisher tier has not changed. What has changed is that we can now watch ChatGPT searching those sites directly, with terms like "women's jeans", and using what comes back. Whether you are on that page shapes whether you make the answer, and how much ChatGPT can say about you when you do. Earned coverage is now a retrieval play, not just a reputation one.

If you do GEO or organic. Getting onto the trusted shortlist is an entity problem, not a content problem. Press coverage, consistent brand and topic association, and citations from sources the model already trusts.

If you run the site. Assume your own domain is the target of a site: fan-out. Pricing, stock and product detail need to be accurate and crawlable, in the HTML, not rendered client-side.

One point to remember. Fan-outs are generated at the time of the prompt. They are not fixed, they are not the same for everyone, and they move again every time the model does.

Organisations used to live and die by Google algorithm updates. Now it is model updates, and what changes under the hood filters straight through to buying behaviour, usually before anyone announces it.

One release took sources per answer up 5x and handed ChatGPT a shortlist of entities it trusts to answer the question. The person asking that question could be your buyer.

We tracked all of this in Obsero. If you want to see which domains ChatGPT site-restricts to in your category, the free trial is here.

As always, thanks for subscribing. I do love to hear your feedback, so please reply to this email and let me know what you thought of this edition.

If you want me to run the fan-out analysis on your category, reply and tell me the brand. I will have a look.

Until next week,

Andy