For years, organizations approaching AI search have worked from a familiar assumption: if their website is visible in Google and Bing, it will probably be visible in ChatGPT as well.
That assumption is becoming increasingly unreliable.
A recent investigation by Peec AI suggests ChatGPT operates its own retrieval infrastructure, including specialized indexes for the web, PDFs, news, YouTube, Wikipedia, local search, finance, legal content, medical information, shopping and images. The system, reportedly called “Labrador” internally, appears to work alongside external providers such as Google, Microsoft, Yelp and TripAdvisor. AI visibility is no longer simply a matter of traditional search visibility.
Search engines are becoming sources, not the whole system
ChatGPT still appears to use external search providers. Google and Bing remain relevant, and Microsoft says its Web IQ grounding infrastructure powers ChatGPT alongside Microsoft Copilot and other products.
But external search is only one retrieval path. ChatGPT can combine its own crawled content with search results, licensed data, specialized vertical indexes, cached pages and ranking systems optimized for conversational answers.
A strong position in Google may help, but it does not guarantee that ChatGPT will select the same page, interpret it in the same way or cite it in the same answer.
AI-generated answer
Several retrieval paths can contribute to one answer.
ChatGPT indexes
Web · PDFs · newsShopping · local
Search providers
Google · BingOther retrieval services
Data feeds
Shopify · YelpTripAdvisor
Your public information
Pages · documents · product data · listings · independent sources
rworks.ai
AI search has its own content architecture
Traditional SEO generally focuses on making pages discoverable, crawlable and competitive for particular keywords. AI retrieval adds another test: whether a source is understandable, extractable and useful as evidence.
A page can rank well in a conventional search engine and still perform poorly in an AI answer. Its key claims may be buried in marketing language, important facts may be scattered across several pages, or product and policy details may be hard to extract without context. Inconsistent terminology, missing dates and weak attribution create the same problem.
This is why AI visibility is not reducible to “ranking higher.” The system has to retrieve a source, determine what it means and decide whether it is reliable and relevant enough to use.
Specialized indexes change the strategy
The reported presence of separate indexes for news, PDFs, local information, shopping and other content types is particularly important. Different content types serve different intents. A company trying to appear in a product recommendation faces a different retrieval environment from a university trying to be cited in a research answer.
Shopping offers an early example. Profound reported that feed-integrated retrieval in ChatGPT Shopping rose from 8.26% to 61.54% on July 10 across about 1.75 million prompts. It also found substantial gains for large feed-integrated merchants, while hundreds of monitored businesses saw sharp visibility changes.
Profound linked the shift to a model release the day before. That attribution is its analysis, not an OpenAI disclosure. The broader lesson is more durable: for some queries, structured feeds and platform integrations may matter as much as, or more than, a page’s position in web search.
A generic content strategy is unlikely to cover every surface. Organizations need to know where their most important information lives, whether each format is machine-readable, and whether first-party pages, PDFs, listings, feeds and independent sources agree.
Caching creates a separate freshness problem
Peec’s investigation also suggests ChatGPT caches pages instead of fetching every source live for every query. That makes retrieval faster, but it creates a gap between what an organization has published and what an AI system currently knows.
A page may have been updated yesterday while an AI system is still relying on an older version. Conversely, an outdated page may continue influencing answers after pricing, leadership, policies or product details have changed.
Important pages need clear publication and update dates, stable URLs, explicit change notices and clean redirects from retired content. For high-risk facts, the useful test is not merely whether the current page exists. It is whether AI systems are actually using the current version.
Bing visibility is not a complete proxy
Bing remains relevant, but ChatGPT may reach information through its own index, an external search provider, a licensed data partner or a cached representation. The route can vary by query and product surface.
Organizations therefore need to measure AI visibility directly: whether they are mentioned, which sources are cited, whether the information is accurate and which competitors appear instead. Testing should also show whether an updated fact makes its way into later answers or whether the old claim persists.
This is closer to observability than conventional rank tracking. The question is not only, “What position do we have?” It is, “What information is the system assembling about us, and where did it come from?”
What organizations should do now
Start with a source inventory: the pages, documents, profiles, feeds and third-party listings that contain your most important facts. Make those facts easy to extract, then remove contradictions across the wider information layer.
Next, test the questions customers, journalists, policymakers or researchers are likely to ask. Record the answer, sources, omissions and inaccuracies, then repeat the test over time. AI answers can change as indexes, caches, ranking systems and outside sources change.
The broader shift: from search optimization to information engineering
The emergence of proprietary AI indexes points to a larger change. Search optimization was largely about competing for attention within an established index. AI visibility is increasingly about making an organization legible across several retrieval systems at once.
That requires more than adding keywords. It requires structured facts, authoritative sources, consistent terminology, update discipline and continuous testing.
Google and Bing still matter. But they are no longer the only gateways between an organization and an AI-generated answer.
The organizations best positioned for AI search will treat their public information as infrastructure: organized, current, verifiable and designed to survive retrieval across multiple systems.
In the age of generative search, discoverability is only the beginning. The real challenge is becoming a source an AI system can confidently find, understand and use.