跳到正文
DEV Community· Ke Gao·· 13 天前AI 評分59

LLM 實際會引用哪六類網頁,以及如何在一個下午找出自己的引用來源

The six kinds of pages LLMs actually cite (and how to find yours in an afternoon)

AI 導讀

Ke Gao 提出一套找出 LLM 引用頁面的審核方法,將答案引用的 URL 分成六類,並找出品牌尚未覆蓋的類別。方法以 12 條、分屬類別、問題、比較和品牌四組的提示詞,在 ChatGPT(開啟搜尋)、Perplexity 和 Gemini 收集引用 URL,再統計各類分佈。文章建議先檢查爬蟲存取和 Bing 索引,選三類改善,並在兩至四星期後用同一批提示詞重測。

正文

Last time I wrote about auditing whether AI mentions your brand. The part people ask about is always the same: which pages get cited, and how do I get on them?

Here's the short version: run your buyer prompts, collect every URL the answer cites, and cluster them. In every audit I've done, the citations fall into six buckets. Your job is to find the buckets where you're absent but reachable.

Step 1: collect the citations (30 minutes)

Twelve prompts, four families: category, problem, comparison, brand. Run them in ChatGPT (with search), Perplexity and Gemini, logged out. Copy every source URL each answer cites into one column. Don't judge yet, just collect — 60 to 100 rows is normal.

Step 2: cluster into six buckets

Every URL will land in one of these:

Bucket What it looks like How you get in
Directories "Top 40 AI tools", niche tool indexes Free/paid submission forms
Listicles & comparisons "Best X for Y in 2026", "A vs B" Pitch the author, or publish your own comparison page
Docs & changelogs Product docs, release notes, pricing pages Yours to write — make one page answer one question
Community threads Reddit, Quora, HN, Indie Hackers, forums Answer real questions, disclose you're the maker
News & editorial Trade press, launch coverage Harder; usually needs a story, not a product
Code & lists GitHub repos, awesome-lists, package docs A PR, if your tool is developer-adjacent

Step 3: read the distribution, not the individual URLs

Count rows per bucket. Two patterns come up constantly:

  • Directories + listicles dominate. That means the model isn't reading your site, it's reading the pages about your category. Ten good listings beat one beautiful landing page.
  • Community threads show up more than people expect, especially for "what tool should I use" style questions. One useful Quora answer can be pulled into answers for months.

Now compare with where you actually exist. The gap list is your roadmap — and it's usually shorter than it feels.

Step 4: fix the boring blockers first

Before any of this works, two things are often quietly broken:

  1. Crawler access. Check robots.txt for GPTBot, ClaudeBot, PerplexityBot, Google-Extended. Plenty of sites block them by accident via a copied template. Two years of content behind a crawler block is a common, embarrassing finding.
  2. Indexing in the engine that powers the answer. Copilot runs on Bing. If you have no Bing index, you don't exist for a slice of users regardless of content quality.

Step 5: pick three buckets, not six

The failure mode is trying all six at once. I'd pick: one directory pass (10 submissions), one community pass (3 thoughtful answers, disclosure included), and one content pass (5 answer-shaped pages — one question per page, answer in the first 40 words).

Then re-run the same twelve prompts two to four weeks later and count the same buckets. If the distribution shifted toward sources you control, it's working. If it didn't, you were probably working on the wrong bucket — which is exactly the thing measurement is for.

If you'd rather not build the spreadsheet, I made bidvix for the measurement half (mention rate + cited URLs across five models). The clustering above is the part that actually changes what you do, and it works fine in a notebook.

來源:DEV Community · dev.to