No Link in ChatGPT = Not Referenced? A Misconception About How AI Gathers Information
機械翻訳 / Machine-translated
A
AI Friends
1日前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "I asked ChatGPT about our company and our site's link didn't appear — maybe it's not being read at all"? That conclusion may actually be premature. This article breaks down what AI information gathering actually looks like and how you can verify it.
On November 7, 2026, Keyman's Net published an article explaining that "no link in a ChatGPT response = not being read" is a misconception.
The key point: the pages an AI reads when composing a response are not necessarily cited within that response.
In other words, even if you look at the links shown on screen, it's hard to know which information actually informed the AI's answer. Companies working on AI search optimization (known as GEO or LLMO) are wrestling with exactly this problem.
A reference URL is the URL of a page the AI visited as a source while composing its response.
A citation URL, on the other hand, refers to the URL displayed in the completed response as a source reference.
These two don't necessarily match. Some pages that were referenced never get cited, and the reverse can happen too.
Imagine someone who reads ten books at a library and then lists only three of them at the end of their report. The other seven may still have influenced the content.
Furthermore, CINC's analysis found cases where information included in a response could not be verified on the cited page. Relying solely on what's displayed as a source can lead to faulty conclusions.
On October 26, 2026, CINC announced the addition of a new feature to its in-house GEO/LLMO tool. The feature retrieves the URLs of web pages that ChatGPT and Claude visited when composing a response.
For each specified prompt (the question posed to the AI), users can view the following four items in a list:
Three use cases are highlighted:
Say a PR manager at a mid-sized manufacturer asks ChatGPT every week, "What products do you recommend?" Their company's website link never appears in any response.
The manager concludes, "AI is ignoring us," and nearly gives up on the budget for a site overhaul.
But when they check the reference URLs, they find that the company's product comparison page was being read every single time. The response text even contained spec figures that only existed on that page.
In this scenario, what's needed isn't a strategy to get the site read — it's a strategy to get it cited. That means refining how headings and conclusions are written, and making the source of information clearer.
Note: This is a hypothetical example for illustrative purposes.
Incomplete source attribution isn't unique to Japanese cases. As reported by TechCrunch, Columbia University's Tow Center examined 200 citations — 10 articles each from 20 publishers.
The result: ChatGPT gave partially or entirely incorrect responses in 153 out of 200 cases. It acknowledged being unable to answer in only 7 cases.
The researchers themselves acknowledged the study's limited scale, and OpenAI pushed back, saying the usage patterns tested weren't typical. Since this is a 2024 study, results may not reflect current behavior.
Even so, it offers evidence supporting the view that judging AI behavior by source attribution alone is risky.
GEO optimization tools fall into two broad categories. A comparison article by Jinrai that looked at domestic tools organizes them as follows. Note that the company offers its own tool, so their assessments should be read with that in mind.
Pricing ranges widely — from free one-time checks, to around ¥10,000/month (Brand UP, GEO Hack Suite, SUPER ACT, etc.), to tens of thousands of yen per month (Mieruca GEO, Pascal, etc.).
CINC's new feature adds the perspective of "looking all the way to reference URLs" to this measurement landscape. However, it is part of CINC's internal tool and is positioned for consulting use. Public availability and pricing have not been confirmed.
For companies in Japan, the most important takeaway is: don't assume "not cited = not read."
Before writing "zero citations" in an AI search optimization report, it's worth considering whether a site may still be referenced. Conversely, a page that is cited may contain outdated or incorrect information.
Here are some checks you can do starting today:
Retrieving reference URLs is difficult without a dedicated tool. Starting with manual checks is the most practical first step.
Not necessarily. Pages that were referenced may not be cited in the response, so the presence or absence of a link alone isn't enough to judge.
They are generally difficult to identify from a standard interface. CINC has now added a feature to retrieve them within its own tool.
Yes. Both ChatGPT and Claude are supported.
Not always. In the Tow Center's 2024 study, errors were found in 153 out of 200 cases. Always verify important information against the original source page.
Start by asking an AI about your company name and checking whether the response content aligns with your own pages.
This article is a cross-post from AI Friends.
