Something unusual is appearing in website logs: requests from ChatGPT and other artificial intelligence systems for plausible pages that never existed, such as /security, /enterprise-pricing, or /compare/product-a-vs-product-b.
Imagine an enterprise buyer researching software. ChatGPT requests /security/data-residency, although the company publishes that information elsewhere. The request returns a 404 and disappears into a server log.
Most companies treat these requests as they would any other 404. The URL is wrong, the server returns an error, and someone may eventually add a redirect. Marketers should examine the requested path before fixing it, because a plausible URL generated by an outside system can reveal what that system expected the company to answer.
TWO DIFFERENT KINDS OF DATA
A 404 once meant a moved page, stale backlink, or mistyped address. Generative AI adds another possibility: systems can generate plausible URLs based on patterns in the information they encounter. In an analysis of 16 million URLs cited by six AI assistants, Ahrefs found that AI assistants referred visitors to 404 pages 2.87 times as often as Google Search. ChatGPT had the highest 404 rate.
For marketers, invented links may be useful evidence. A request for /enterprise-pricing suggests that the system expected the company to publish enterprise pricing. Repeated requests for /security/data-residency may signal that security proof matters to an upstream decision and is too hard to find.
Not every strange path is meaningful. Focus on requests associated with AI-agent activity that recur or form coherent clusters. Together, they can reveal a shadow information architecture: What customers and their agents expect the website to answer.
THE DISTINCTION BETWEEN PERSONALIZATION AND RELEVANCE
Analytics describe what people did within the website the company created: the pages they visited, what they clicked, and where they left. Marketing technology can identify people through account, industry, purchase history, audience, or propensity data, yet still miss why someone arrived today. An anonymous request for /compare/product-a-vs-product-b may reveal more about the present decision than another field in a customer profile.
Personalization starts with what a company knows about a person. At FibrAI, we define relevance as understanding what a visitor, or an agent acting for them, is trying to accomplish in the moment and supplying the information needed to meet that goal. A failed AI request can expose that intent directly: the path may identify the comparison, concern, proof or next step that is missing. AI-referred visitors often arrive with developed intent, so the website must meet them at the appropriate point in their decision.
THE EFFORT TO TALK TO MACHINES
Companies are trying to tell AI systems what their websites contain. Adoption of llms.txt, a proposed standard for surfacing preferred content to language models, grew 8.8-fold from June 2025 to May 2026, according to Originality.ai. Yet an Ahrefs analysis of 137,000 domains found that 97% of llms.txt files received no requests in May 2026.
Making websites easier for machines to understand remains useful. The missed opportunity is treating the information flow as one-way: Brands publish for machines but pay less attention to what machines try to retrieve. Those requests may reveal unfamiliar terminology, hidden comparisons, or proof buyers expect earlier in the journey.
THE ERROR-HIDDEN INSIGHT
Adobe’s LLM Optimizer monitors CDN logs for 404, 403, and 5xx errors returned to identified AI agents. For 404s, it recommends alternatives with a confidence score and rationale.
That infrastructure is useful, but the marketing question remains after a redirect. If systems repeatedly ask for /integrations/salesforce while the company’s information sits on /connectors/crm, the mismatch still needs an explanation. The answer may be hard to recognize, available only in sales collateral, absent because buyers expect a dedicated offer, or intentionally withheld. Each explanation calls for a different content or customer-journey decision.
There is a second reason to be careful: A redirect makes an assertion about relevance. If an AI system asks for a security resource and receives a generic trust page that does not substantiate the expected information, the company has supplied an adjacent answer and left the interpretation to the machine.
A verified destination should be served when one exists. When it does not, declining to substitute an answer may be the more relevant and trustworthy response.
WHERE THE ERROR LOG BELONGS
The practical change begins with separating a meaningful pattern from ordinary server noise. A failed path becomes worth investigating when the request can be reasonably associated with AI-agent activity, recurs, or belongs to a semantically coherent cluster, and points to a recognizable information need.
The next question is whether an authoritative destination already satisfies that need. A similar URL is not enough. The page must provide the information and proof the request implies. Preserving these signals over time allows teams to see recurring concepts, differences between market language and company terminology, and repeated demand for comparisons, integrations, policies, or evidence that remain difficult to locate.
Those patterns belong alongside site-search behavior, search demand, customer interviews, sales feedback, and win-loss analysis. They can inform information architecture and content planning because they expose the difference between what a brand has chosen to say and what somebody expected it to answer.
This is where the 404 becomes part of a larger change in marketing. AI visibility can make a brand discoverable. But the website still must provide the information, proof, and next step that allow a customer, or an agent working for that customer, to make a decision.
Recurring invented paths should therefore remain available for analysis after the immediate technical response. Depending on what the pattern reveals, the right action may be to route the request to an authoritative resource, improve a page that obscures the answer, or create a resource the market repeatedly expects to find. Preserving the evidence allows marketing and web teams to make that choice deliberately.
Pritam Roy is co-founder of FibrAI.