OpenAI's help docs describe ChatGPT workspaces using an indexed, cached web. What's confirmed, what's vendor inference, and why discovery and links matter.
"ChatGPT is building its own search index" has been doing the rounds again this week, mostly off the back of a September 30 episode of Edward Sturm's show with David Quaid that passed 10,000 views in its first few days. The headline is true. But most of the posts repeating it mix three very different kinds of evidence: what OpenAI has said in its own documentation, what an OpenAI executive said under oath, and what vendors have inferred from poking at ChatGPT's network traffic. Those deserve different levels of trust, and the gap between them matters if you're deciding what to change on your site.
What OpenAI has confirmed in writing
The firmest source is OpenAI's own help center. Its article on offline web search for ChatGPT workspaces says the feature "allows ChatGPT to use OpenAI's indexed and cached web content instead of using live web search at the time of each request." It's limited to eligible workspaces (certain Enterprise, Edu, Healthcare, regulated and federal setups), and it's framed as a compliance option, not the default consumer experience.
So OpenAI does maintain an index and cache of web pages large enough to answer searches without calling an outside provider. That much is documented, not guessed.
The same article is also unusually frank about what that index misses. It says coverage "does not guarantee coverage of the entire public web," that there's "no refresh SLA for a specific URL," and it lists why a page might be missing or stale:
- The site blocks crawling through robots.txt or similar controls
- CDN or bot-blocking behavior
- Login or personalization requirements
- Heavy reliance on scripts or dynamic loading
- "The page is new, rarely accessed, or low-signal"
- "The site structure makes deeper pages harder to discover"
Read that list as an SEO and it looks very familiar. It's close to the reasons a page fails to get crawled and indexed by any search engine.
What was said under oath
The second layer is Nick Turley's testimony at the Google search remedies trial in April 2025. Courthouse News reported that Turley said ChatGPT pulled from Bing, that OpenAI had "significant quality issues" with a provider he called "Provider No. 1," and that Google turned down OpenAI's request for API access. The same report says OpenAI "hoped that ChatGPT would be able to use the index for 80% of its searches by the end of 2025, but that goal is now years down the line."
One small correction for anyone citing this. Some recent vendor write-ups place that 80% target at the end of 2023. Contemporaneous trial coverage from Courthouse News says end of 2025, and that the goal was by then years away. Either way, the takeaway is the same: as of the testimony, OpenAI's own index was a goal it hadn't hit.
What vendors have inferred
The third layer is the most detailed and the least official. Peec AI's September 4 analysis reports that between May 21 and July 21, 2026, a field called result_source appeared in ChatGPT's server-side events with four values: Labrador, Bright, Oxylabs and SERP. That's where the "Labrador" name for OpenAI's index comes from. Peec also reports shopping experiments with names like "prefer-index-over-serp-v3" running on slices of traffic, and a test by its CPO Malte Landwehr in which asking ChatGPT about a zero-traffic site produced matching spikes in that site's Google Search Console on the same days.
This is useful work, and the Search Console test is clever. But it's observation from outside, not a statement from OpenAI. It tells you OpenAI is mixing its own index with outside results (including, apparently, Google's), and testing how much to lean on each. It doesn't tell you the ratio, and it doesn't mean Bing or Google stopped mattering.
That fits what we found in our look at why 88% of ChatGPT's citations still trace back to classic search rankings. An own index running alongside third-party results doesn't break that pattern overnight. It adds a second gate.
Why links matter for the second gate
Most coverage stops before this point. OpenAI's help article never mentions links. But look again at two of its reasons for a page being absent: "new, rarely accessed, or low-signal" and "site structure makes deeper pages harder to discover." Crawlers find pages by following links, from other sites and from inside yours. A page with no external links pointing to it and only a weak internal path is, almost by definition, rarely accessed and hard to discover. That's our reading, not OpenAI's wording, but it's the same mechanism that has decided crawl priority for every search engine so far.
So the practical checklist for getting into OpenAI's index looks a lot like an indexation checklist with an AI crawler bolted on:
Let the right bot in. OpenAI's crawler documentation says OAI-SearchBot is the one used "to surface websites in search results in ChatGPT's search features," and that sites opted out of it "will not be shown in ChatGPT search answers." GPTBot is the training crawler and is controlled separately. Blocking GPTBot to opt out of training doesn't block search, but plenty of sites have blocked both by accident. Robots.txt changes take about 24 hours to register, per the same page.
Check your CDN, not just robots.txt. OpenAI lists CDN and bot-blocking behavior as its own reason for missing pages. A firewall rule that challenges unknown user agents can keep OAI-SearchBot out even when robots.txt allows it. OpenAI publishes its IP ranges at the URL on that crawler page, so you can allowlist them.
Don't hide content behind JavaScript. "Content depends heavily on scripts or dynamic loading" is on the list. We covered how often that bites in our post on JavaScript links that AI crawlers can't see.
Give deep pages a path. Make sure the pages you want cited are reachable within a few clicks through plain HTML links, and earn at least a few external links to the ones that matter most. A product comparison or data page buried four levels down with no inbound links is exactly the "low-signal, hard to discover" page OpenAI says may not be there.
What not to do
Don't start tracking a separate "ChatGPT rank" for the index on the strength of one network-traffic field. Don't rebuild your link strategy around Bing because a 2025 deposition mentioned Bing, either. The defensible position, based on what OpenAI itself has written, is simpler: ChatGPT can answer from its own crawl, that crawl misses pages that are hard to find or low-signal, and the work that makes a page easy to find and worth citing is the same work that has always driven indexation.
If you want a quick place to start, pull your server logs for the last 30 days and filter for OAI-SearchBot. If your most important pages never show up, that's your answer before any tool tells you. Then check which of those pages have zero referring domains. A link tracker (SerpCrest's included) will show you which pages are carrying the links and which ones are sitting there with none.
