ChatGPT indexed a huge pile of new pages. A blank one still will not sell.

On 15 October 2026, Cloudflare and OpenAI published early results from a research pilot they started in July. The writers are Matthew Conroy and Jack Galilee at Cloudflare, and Henry Aspegren and James Balfour at OpenAI. Using signals from about 1 percent of Cloudflare's network, OpenAI's search crawler discovered and indexed hundreds of millions of web pages it had not had before. OpenAI then decides, on its own, which of those pages are relevant enough to show. ChatGPT Search can link back to the original page.

That is a large number. It is not a promise that your company site was in the pile, or that a citation will turn into an enquiry.

The useful part of the post is what the crawler learned to skip. Bots ask, over and over, for files that are not there. Some sites send almost any address to the homepage. Other addresses are copies of the same page. A crawler that spends its time on those is not reading the installation note, the current fee, or the page that says what you actually sell. Cloudflare says a signal about a page is metadata for discovery. It is not a copy of the page handed over for model training. A signal does not grant permission to fetch the page. Domain owners still choose, and Cloudflare points them to AI Crawl Control if they want to see and manage that access.

They are also testing a narrower safety signal. A real site can have one bad page, such as a fake login form planted by someone else. Flagging that page, rather than the whole domain, is the idea. They call the results early. Do not treat it as a shield you can buy.

Two practical consequences follow, and both belong in a website plan, not in a rumor about ChatGPT traffic.

First, a page has to exist, and it has to be the page. If every unknown address jumps to the homepage, you have trained both people and crawlers to ignore the address. A buyer who wants the desktop app, the mobile app, or the care terms should land on that page. A homepage that tries to be all of them is how the specific sentence gets lost. When we build a company website, the written plan names the pages before we start. It also names the price, and who owns the code. After handover, we fix faults in the agreed work for at least 3 months. New features are outside that cover.

Second, the page has to say something a rival page does not. Cloudflare's examples are ordinary and exact: the manual for one washing-machine model, the current passport fee, today's menu and hours. A company site has the same job. The services page should name what this job includes. Ours name a company website, a mobile app, a desktop app, AppSheet, search, and website care, each with its own limits. Search work does not promise a rank. A mobile app does not go to the App Store unless the plan says so, and the store still decides. AppSheet is for the sheets you already use. A desktop app can talk to a printer or a scanner if the plan names that equipment. We do not buy the hardware. Care is for a site that is already live, including one we did not build, after we have looked at the code and the host.

None of that appears because a crawler is clever. It appears because someone wrote it down.

The post also warns that finding new pages can raise crawling at first, while those pages are judged. A quieter site is not always a healthier one, and a busier log is not a sale. Website care can watch the certificate, the updates, and a break in the form. The plan names what that care includes, and the price. Do not send passwords in the form.

Search work, if you want it, is on the pages you already have: titles and descriptions that match the page, a sitemap, and pages that say what you sell. We do not promise a date. A search engine recrawls on its own schedule. We usually need read access to Search Console and access to the site. The same limit applies to any assistant that cites the web. We cannot make ChatGPT quote you. We can make the page worth quoting, and we can stop the site from hiding that page behind a redirect.

If you are on Cloudflare, open AI Crawl Control and see what is asking for your pages. Then open three addresses a buyer might type. If two of them dump the visitor on the homepage, that is the brief. Tell us what you need. We reply by the next business day. You agree the written plan before we start.

Source: Cloudflare and OpenAI, "Making AI discovery better for everyone," 15 October 2026. https://blog.cloudflare.com/cloudflare-openai-discovery-agentic-web/