Metehan Yesilyurt
AMA 問答

I wrote 100 questions to understand the difference. I'm not a guy who thinks GEO is completely new and different than SEO. I still believe SEO fundemantals are the key for AI visibility. GEO, AEO, LLMO, SXO or anything you name it. https://searchengineland.com/seo-ai-search-questions-464550

我寫過 100 個問題來理解兩者的差異。我不是那種認為 GEO 完全是新東西、和 SEO 截然不同的人。我仍然相信 SEO 基本功是 AI 能見度的關鍵。不管你叫它 GEO、AEO、LLMO、SXO 還是別的名字。https://searchengineland.com/seo-ai-search-questions-464550

What's same? Discoverability, crawling, semantics, reranking, personalization paths.

相同的是什麼？可發現性、爬取、語意、重新排序（reranking）、個人化路徑。

What's different in general? RAG, hybrid approaches, interactive outputs (like artifacts or basic calculator tools that Google created instantly and provide it in AIO or AIM answer)

整體來說不同的是什麼？RAG、混合式做法、互動式輸出（例如 Google 當場生成、放進 AIO 或 AIM 答案裡的 artifact 或簡單計算機工具）。

+ agentic RAG, search & shopping.

再加上 agentic RAG、搜尋與購物。

Thank you! The most dangerous GEO advice is "publish more listicles" or "do programmatic SEO at scale" Most of the people are just using same prompts or rules for it with cheap models. It. Won't. Work.

謝謝！最危險的 GEO 建議是「多發布清單文」或「大規模做程式化 SEO」。多數人只是用同樣的 prompt 或規則，配上便宜的模型去做。這。行。不。通。

I tried this in an experiment and my website literally tanked by Google. Website is still performing well on ChatGPT but what about others? If my audience lives in only ChatGPT, it's okay for me. Otherwise...

我在一個實驗裡試過，我的網站真的被 Google 打趴了。網站在 ChatGPT 上仍然表現不錯，但其他平台呢？如果我的受眾只活在 ChatGPT，那對我來說沒問題。否則……

What does move the needle? This is a great question.

什麼真的有效？這是個很棒的問題。

I'm trying to predict what users are really asking, then start tracking and analyze query fanouts. I am pretty sure this is the most underrated gem in AI visibility.

我試著預測使用者真正在問什麼，然後開始追蹤並分析 query fan-out。我很確定這是 AI 能見度裡最被低估的寶藏。

Specific use case: Many brands publish lists for ChatGPT visibility, and these lists do perform well. However, they prioritize themselves first, followed by other players in the industry(not the industry leading competitors). But in fanout patterns, if there are large players involved, it means you'll likely lose during the retrieval phase because you don't have the coverage that addresses the fanout.

具體案例：很多品牌為了 ChatGPT 能見度發布清單文，而這些清單確實表現不錯。不過，他們把自己排在第一，接著才是產業裡的其他業者（而不是產業龍頭的競爭對手）。但在 fan-out 的模式中，如果有大型業者參與，代表你很可能會在擷取階段輸掉，因為你沒有涵蓋 fan-out 所需要的範圍。

Thank you!

謝謝！

If I have this case, I absolutely start with on-site. Publish great content. Start tracking, analyze fanouts.

如果是我碰到這種情況，我絕對從站內開始：發布很棒的內容，開始追蹤、分析 fan-out。

Then continue with Reddit/forums/LinkedIn/Substack. It's shortest path for creating a consensus. When it comes grounding on topic, you absolutely need consensus.

接著繼續做 Reddit／論壇／LinkedIn／Substack。這是創造共識的最短路徑。說到對主題的 grounding，你絕對需要共識。

If you are operating in local, press releases work really well.

如果你做的是在地市場，新聞稿效果非常好。

Thank you! Great questions!

謝謝！很棒的問題！

I didn't see this pattern. llms.txt is useful for agents, not for AI visibility. If you read anything around AI visibility, just run.

我沒看過這種模式。llms.txt 對 agent 有用，對 AI 能見度沒用。如果你讀到任何圍繞 AI 能見度談 llms.txt 的東西，就快跑。

This is a unique configuration. I believe Google can understand and parse everything that you're doing right now but for ChatGPT and others? It's not even canonical tags aware. So, we still have time to see/track progress with OpenAI. However OpenAI is able to parse your all <head> tags (100% sure about it, I have proof) but I don't know if they respect for noindex tag(if they aware of noindex tag, probably they respect it). OpenAI also checks robots.txt.

這是個很特別的設定。我相信 Google 能理解並解析你現在做的一切，但 ChatGPT 和其他的呢？它甚至不認得 canonical 標籤。所以我們還有時間觀察／追蹤 OpenAI 的進展。不過 OpenAI 能解析你所有的 <head> 標籤（我 100% 確定，我有證據），但我不知道他們是否遵守 noindex 標籤（如果他們認得 noindex，大概會遵守）。OpenAI 也會檢查 robots.txt。

2.1) I'm testing a new approach for serving different rendering for LLMs, I'll have better data soon.

2.1）我正在測試一種針對 LLM 提供不同呈現方式的新做法，很快會有比較好的資料。

Thanks for asking! Great question.

謝謝你問這題！好問題。

There is technique named "query decomposition" and LLMs are using it. We never seen fanouts in AI Mode & AI Overviews so we're partially Google agnostic on this topic (It's possible to see them in Gemini)

有一種技術叫「query decomposition（查詢分解）」，LLM 都在用。我們從沒在 AI Mode 和 AI Overviews 裡看到 fan-out，所以在這個主題上我們對 Google 部分是「不可知」的（在 Gemini 裡可以看到）。

Query decomposition source: https://haystack.deepset.ai/blog/query-decomposition

查詢分解的出處：https://haystack.deepset.ai/blog/query-decomposition

2) It's most likely seems "hybrid approach" What I mean here is LLMs are trying to understand the "intent" first. If it's addressing a shopping intent, switches to shopping retrieval path instantly or maps/local or academic research or medical queries. Or just web.

第 2 點）這看起來很可能是「混合式做法」。我的意思是，LLM 會先試著理解「意圖」。如果是購物意圖，就立刻切換到購物擷取路徑；或是地圖／在地、學術研究、醫療類查詢；不然就只是一般網頁。

3) For web retrieval, they are using a tiny model for creating sub-queries/fanouts for the prompt (it's cheaper on computing side) if they trained a different model. Otherwise, if the model is intelligent enough, it's also able to create fanouts and make live searches. The important point is here that the model that you use and context size. For example; ChatGPT has 3 options; Instant, Medium and High. Each selection will you bring different retrieval pool. High will bring you more citations and better context.

第 3 點）網頁擷取時，如果他們另外訓練了模型，就會用一個很小的模型來為 prompt 產生子查詢／fan-out（運算成本較低）。否則，如果模型夠聰明，它自己也能產生 fan-out 並進行即時搜尋。重點在於你使用的模型與 context 大小。例如，ChatGPT 有 3 個選項：Instant、Medium 和 High。每個選項會帶來不同的擷取池，High 會帶來更多引用與更好的上下文。

4) Why it matters? Google is doing this for over two decades, so it's all fine. But if your goal is building the AGI world, you need to understand what user is asking for and find the most relevant documents online. So, that's why query fanout matters more. Semantics, understanding, computing.

第 4 點）為什麼重要？Google 做這件事超過二十年了，所以沒問題。但如果你的目標是打造 AGI 的世界，你就需要理解使用者在問什麼，並找到網路上最相關的文件。所以 query fan-out 才更重要。語意、理解、運算。

Every LLM is different. Even AI Overviews and AI Mode. AI overviews is more grounding context aware.

每個 LLM 都不一樣，連 AI Overviews 和 AI Mode 也不同。AI Overviews 對 grounding 上下文更敏感。

You need discoverability first. So, start with technical SEO (I know it's boring at the start point)

你首先需要可發現性。所以從技術 SEO 開始（我知道一開始很無聊）。

ChatGPT is using a technique named "query decomposition" you can Google it (sometimes Reddit flags some links in this thread as spam)

ChatGPT 使用一種叫「query decomposition」的技術，你可以自己 Google。（有時 Reddit 會把這串裡的某些連結標成垃圾訊息）

Then makes live searches OR lookup in their index, then fetch every result and rank them with using some approaches/techniques like embeddings, bm25, chunk scorings. (we have the receipts for these)

然後進行即時搜尋，或查詢它們的索引，接著擷取每個結果，並用一些做法／技術來排序，例如 embeddings、bm25、chunk 評分。（這些我們都有證據）

Do you need backlinks? It's not needed for ChatGPT right now based on my findings so far, but we can't know the future. (I don't have receipts for this)

需要反向連結嗎？依我目前的發現，ChatGPT 現在不需要，但我們無從得知未來。（這一點我沒有證據）

How ChatGPT reads your content? I recommend you to read this: https://x.com/metehan777/status/2096994466573730188

ChatGPT 怎麼讀你的內容？我建議你讀這篇：https://x.com/metehan777/status/2096994466573730188

In my experience, parametric memory definitely has an impact because it creates leverage. It's like knowing the best authors in a library. I even asked a short question: "Would you recommend Strictly Median Bank?" and the answer was predictable.

依我的經驗，參數記憶絕對有影響，因為它創造了槓桿。就像知道圖書館裡最好的作者是誰。我甚至問過一個簡短的問題：「你會推薦 Strictly Median Bank 嗎？」答案是可預期的。

I've done a lot of research on this, and I'm still doing it. Yes, many different companies have many different models. There's competition in terms of intelligence. The model's own knowledge is very important here.

我對這件事做過很多研究，現在還在做。對，很多不同的公司有很多不同的模型，在智能上互相競爭。模型本身的知識在這裡非常重要。

If we look at ChatGPT, the prompt first creates fanouts, and it doesn't do this all at once; the model makes a decision in under a second and performs a multiple fetch. With ChatGPT 5.6, we see that brands are more often in fanouts, and I believe parametric memory is important here. I don't have 100% certain information because I don't know the training systems; it's a commercial product and a secret.

以 ChatGPT 來說，prompt 會先產生 fan-out，而且不是一次完成；模型在不到一秒內做出決定，並進行多次擷取。在 ChatGPT 5.6 上，我們看到品牌更常出現在 fan-out 裡，我相信參數記憶在這裡很重要。我沒有 100% 確定的資訊，因為我不知道訓練系統；那是商業產品，也是祕密。

I've done a lot of research on Common Crawl sets. I work at Peec, and when I examined Common Crawl's scan sets and analyzed Harmonic Centrality and PageRank by sector, the leaderboards I saw for AI visibility were similar. CC is just an example here, of course; they use many more resources for their training sets.

我對 Common Crawl 資料集做過很多研究。我在 Peec 工作，當我檢視 Common Crawl 的爬取資料集，並依產業分析 Harmonic Centrality 與 PageRank 時，我看到的 AI 能見度排行榜很相似。當然，CC 只是個例子，他們的訓練資料集用到的資源多得多。

So my personal answer is yes.

所以我個人的答案是：會。

This is a great question, especially for non-English markets.

這是個很好的問題，特別是對非英語市場。

Reddit isn't a popular platform in Germany, Türkiye, or the Middle East. In Türkiye, for example, everyone knows Facebook, but Reddit is never popular. Or in France, it's known but not popular. If you read Reddit's investor reports, you can see that its international user coverage is already very low. But visibility? Of course, it's top-notch.

Reddit 在德國、土耳其或中東並不是熱門平台。例如在土耳其，大家都知道 Facebook，但 Reddit 從來不熱門。在法國則是有名但不流行。看 Reddit 的投資人報告就能發現，它的國際使用者涵蓋率本來就很低。但能見度呢？當然是頂級。

We also see different platforms in niche areas, forums for example. Looking at current trends, I'd call them moving targets.

我們在利基領域也看到不同的平台，例如論壇。從目前的趨勢來看，我會稱它們是「移動的目標」。

When it comes to grounding/web search, the model looks for consensus. Therefore, third-party mentions are important.

說到 grounding／網路搜尋，模型尋找的是共識。因此第三方的提及很重要。

I've been following fanouts for a very long time, and they've never let me down.

我長期追蹤 fan-out，它們從沒讓我失望過。

Consensus, consensus and consensus!

共識、共識，還是共識！

Some things never change. Real user value, lightning fast websites. I always recommend that please start with your technical SEO. Be sure, you have a solid technical usability on your website. I know this is also a generic advice but still matters.

有些事永遠不變：真實的使用者價值、飛快的網站。我總是建議從你的技術 SEO 開始。確保你的網站在技術可用性上很扎實。我知道這也是個通用的建議，但它仍然重要。

What I see at most; all LLMs (besides Google) are behaving like a semantic ranker + semantic reranker (including recency boost)

我看到最多的是：所有 LLM（Google 除外）的行為都像是語意排序器加語意重排序器（包含新近度加權）。

So, end of the day; if you are able to provide fresh & up-to-date information + direct answers + entity rich passages, you'll likely win that citation or at least you'll start seeing your URLs in retrieved URLs.

所以，說到底，如果你能提供新鮮、最新的資訊加上直接的答案加上富含實體的段落，你很可能贏得那個引用，或至少會開始在被擷取的網址裡看到你的網址。

Be sure you check your technical SEO performance

確保你檢查自己的技術 SEO 表現。

Analyze fanouts monthly, you'll often find the content gaps in there

每個月分析 fan-out，你常常會在裡面找到內容缺口。

One day is not a trend, so make sure you track your progress weekly & monthly. (I don't say you need to "something" every week.

一天不是趨勢，所以要確保以週和月為單位追蹤你的進展。（我不是說你每週都需要做「什麼」。）

Repeat

重複。

My friend David is doing great research for this topic https://lnkd.in/p/dFrvTY2A These are the first insights so far.

我的朋友 David 正在針對這個主題做很棒的研究 https://lnkd.in/p/dFrvTY2A 這些是目前最初步的洞察。

Google always love experimenting/testing stuff. So, if the brand mention is left unlinked, it's not good for me. I never seen a relation between knowledge graph with brand mentions. (Maybe I'm wrong.)

Google 總是喜歡實驗／測試各種東西。所以，如果品牌提及沒有附連結，我覺得不太好。我從沒看過知識圖譜和品牌提及之間有什麼關聯。（也許我錯了。）

It's easy to track in Peec dashboard, almost every tool also provides this. Being cited matters more of course :) I always check retrieval performance first, it's the first step. Then start analyzing top cited URLs, update your page.

在 Peec 的後台很容易追蹤，幾乎每個工具也都有提供。當然「被引用」更重要 :) 我一向先看被擷取（retrieval）的表現，這是第一步；接著分析被引用最多的網址，再更新你的頁面。

(I started to leave API analysis. It's useful, I know.)

（我已經開始不太用 API 分析了。我知道它有用。）

The best practice is checking the server logs/user agent analytics.

最佳做法是檢查伺服器日誌／user agent 分析。

If you are using Cloudflare, it's the easiest way to see the performance because Cloudflare has an MCP. So, if you are non-technical, you can learn everything talking with ChatGPT Work or Claude + Cloudflare MCP.

如果你用 Cloudflare，那是最容易看到成效的方式，因為 Cloudflare 有 MCP。所以就算你不懂技術，也可以透過 ChatGPT Work 或 Claude 加上 Cloudflare MCP，邊聊邊把一切弄清楚。

We have a great feature, you just connect your log service once, after that we handle everything. https://docs.peec.ai/crawl-insights

我們有個很棒的功能：你只要連接一次日誌服務，之後一切由我們處理。https://docs.peec.ai/crawl-insights

do you use MCPs?

你有在用 MCP 嗎？

It always reminds me the dead internet theory...

這總讓我想起「死網理論」（dead internet theory）……

Right now, as webmasters, SEOs, GEOs, marketers we have a website but do we check which pages are performing well for "bots" and "humans(GA4)" ? You can see the difference yourself.

現在，身為網站管理者、SEO、GEO、行銷人員，我們有網站，但我們有檢查過哪些頁面對「機器人」表現好、哪些對「人類（GA4）」表現好嗎？你自己就能看出差異。

Log file analysis will give you a starting point only. It's super complicated because even you use Cloudflare, AWS or GCloud or Azure, these companies are giving you "sampling" not the full data. Otherwise you need to take care of GBs of data for every week alone. https://developers.cloudflare.com/analytics/sampling/

日誌分析只會給你一個起點。它非常複雜，因為即使你用 Cloudflare、AWS、GCloud 或 Azure，這些公司給你的也是「抽樣」資料，而不是完整資料。否則你每週光是資料就要處理好幾 GB。https://developers.cloudflare.com/analytics/sampling/

We developed Agent analytics to solve this issue. https://docs.peec.ai/crawl-insights Just choose your log provider and we handle the rest.

我們開發了 Agent analytics 來解決這個問題。https://docs.peec.ai/crawl-insights 你只要選擇你的日誌供應商，其餘由我們處理。

There are various resources on the web where you can do this. The easiest way to do this is to look at clickstream data platforms. I'd even like to leave this link here: https://www.similarweb.com/blog/marketing/seo/clickstream-data/

網路上有各種資源可以做到這件事，最簡單的方式是看 clickstream（點擊流）資料平台。我也想在這裡放上這個連結：https://www.similarweb.com/blog/marketing/seo/clickstream-data/

Besides that, there are platforms like Ahrefs and Semrush that show organic SEO traffic. These also use various prediction and modeling methods.

除此之外，Ahrefs、Semrush 這類平台會顯示自然搜尋流量，它們也是用各種預測與建模方法。

We also have a resource on prompt volume at Peec AI: https://docs.peec.ai/metrics/prompt-metrics/prompt-volume, but we use it as a demand score instead of monthly volumes. I believe this is the most reliable approach. Because as a user, I can ask my question in hundreds of different ways. For a very long time, we've believed in the effectiveness of long-tail keywords in the SEO world, and I've achieved success in that direction. Because the competition is brutal in GEO/SEO. It's technically impossible to see a structure for this right now. Also, the main searches are on the fanout side, and they aren't deterministic. Even for the same prompt, fanouts change.

Peec AI 也有一份關於 prompt 搜尋量的說明：https://docs.peec.ai/metrics/prompt-metrics/prompt-volume ，但我們把它當作「需求分數」，而不是每月搜尋量。我認為這是最可靠的做法。因為身為使用者，我可以用成百上千種方式問同一個問題。在 SEO 的世界裡，我們長久以來都相信長尾關鍵字的效果，我自己也在這個方向上成功過。因為 GEO/SEO 的競爭非常殘酷。現在在技術上不可能看出這裡的結構。而且主要的搜尋發生在 fan-out 那一端，它們不是確定性的，即使是同一個 prompt，fan-out 也會變。

Ultimately, what matters is having the platform that allows you to follow the right questions that will drive that transformation. At GEO Research Team, we pay close attention to this on the product side and constantly strive to support our product team.

歸根究柢，重要的是有一個平台，能讓你追蹤那些真正會帶來轉變的問題。在 GEO 研究團隊，我們很在意產品面的這件事，也持續努力支援我們的產品團隊。

The best data is held by OpenAI, Claude, and Google itself. The more they share, the more we can truly understand the actual volume. Hopefully, we'll see better Search Console / Generative Insight dashboards by giants.

最好的資料掌握在 OpenAI、Claude 和 Google 自己手上。他們分享得越多，我們就越能真正了解實際的搜尋量。希望巨頭們能推出更好的 Search Console／生成式洞察儀表板。

Good question.

好問題。

1 & 2) I always start where my audience lives.

第 1、2 點）我總是從我的受眾所在的地方開始。

Let's assume, theoretically, that everyone uses Google search daily in 2026 (and they actually do).

假設，理論上，2026 年每個人每天都用 Google 搜尋（事實上也確實如此）。

But 15 years ago, everyone actively used Facebook, then Instagram came along, YouTube became more popular, and now there's TikTok. Various communities are more active on Telegram, etc.

但 15 年前，大家都活躍在 Facebook，後來 Instagram 出現、YouTube 變得更熱門，現在又有 TikTok。各種社群在 Telegram 等平台上更活躍。

I previously worked at a mobile app advertising company, and collecting Clickstream data and processing it accurately over a long period is a very laborious task. Ultimately, modeling and prediction will still come into play. The number of devices currently in use is also high; is data being collected from mobile devices or only desktops? That's why I believe providing a demand score, as we do, is better. But please don't interpret this as a promotion of PEEC. Let me explain the reason with a very clear example. When researching a non-English market, you must also be familiar with the structure of that country's language; therefore, the intensity of native searches is naturally at its peak. This is the biggest advantage of our research team. We know the American market very well; we can see what works, and similarly, we know non-English markets very well. But ChatGPT still searches using English fanouts. This raises questions in my mind regarding clickstream data. If the data is very robust, there's no problem, but since this is a trade secret for platforms, we naturally won't get a transparent answer. https://www.searchenginejournal.com/chatgpt-search-often-switches-to-english-in-fan-out-queries-report/567811/ We mostly examine US data, but the situation is very different in non-English markets. Therefore, we come back to the same point: demand. https://docs.peec.ai/metrics/prompt-metrics/prompt-volume

我以前在一家行動 App 廣告公司工作，長期收集並準確處理 clickstream 資料是非常費工的事。最後仍然要靠建模與預測。目前使用的裝置數量也很龐大；資料是從行動裝置收集的，還是只有桌機？這就是為什麼我相信像我們這樣提供需求分數會更好。但請不要把這當成 PEEC 的宣傳。讓我用一個很清楚的例子說明原因。研究非英語市場時，你必須熟悉該國語言的結構，因此母語搜尋的強度自然是最高的。這是我們研究團隊最大的優勢：我們非常了解美國市場，能看到什麼有效，同樣地，我們也很了解非英語市場。但 ChatGPT 仍然用英文 fan-out 去搜尋。這讓我對 clickstream 資料產生疑問。如果資料非常穩健，那沒問題，但這是各平台的商業機密，我們自然得不到透明的答案。https://www.searchenginejournal.com/chatgpt-search-often-switches-to-english-in-fan-out-queries-report/567811/ 我們大多檢視美國的資料，但非英語市場的情況非常不同。所以我們又回到同一點：需求。https://docs.peec.ai/metrics/prompt-metrics/prompt-volume

Along with this, I am also in favor of generating prompts that correctly address the funnel and tracking them. As a research team, we are constantly thinking about this and working on what we can present and model more accurately on the demand side, and our product team is working in the same way. (I'm not against using prompt volume.) Here, the focus should be on the quality of prompts generated when first setting up a project. An AI search/SEO expert can understand this when reviewing the dashboard, and we are trying to provide and improve this.

此外，我也贊成建立能正確涵蓋整個漏斗的 prompt 並追蹤它們。身為研究團隊，我們一直在思考並研究在需求端能更準確呈現與建模的方式，產品團隊也是如此。（我並不反對使用 prompt 搜尋量。）重點應放在專案剛建立時所產生的 prompt 的品質。AI 搜尋／SEO 專家在檢視儀表板時能看出這一點，而我們也正在提供並改進它。

GSC is still not good; there is very little data. Bing WMT and Clarity are much better. I also check them weekly.

GSC 仍然不夠好，資料非常少。Bing WMT 和 Clarity 好得多，我每週也都會看。

3) I have been using DataForSEO for a very long time and recommend its use. By the way, related to the first question: they have also been providing a modeling system for prompt volume for a very long time. https://dataforseo.com/help-center/what-is-ai-search-volume-in-dataforseo

第 3 點）我用 DataForSEO 很久了，推薦使用。順帶一提，和第一個問題有關：他們也提供 prompt 搜尋量的建模系統很久了。https://dataforseo.com/help-center/what-is-ai-search-volume-in-dataforseo

Prompt tracking delusion: Sometimes it's better to think in reverse or from a manager's perspective. Even within Google, 3 different areas have emerged: AIO, AIM, and classic search. How do we report this, is it really possible to get an insight? Yes, it's possible through AI visibility dashboards.

Prompt 追蹤的幻覺：有時候最好反過來想，或從主管的角度想。光是 Google 內部就出現了 3 個不同的區塊：AIO、AIM 和傳統搜尋。我們要怎麼回報這些？真的可能得到洞察嗎？可以，透過 AI 能見度儀表板就做得到。

What excites me in this field is parametric memory, how citations change according to the model, semantic links and grounding, the RAG world. In the 2024 Google algorithm leak, how many SEO experts, including myself, managed to get our websites to the top in less than a month? The answer is probably zero. But our work continued, the reports didn't stop. Being able to present concrete data is always important, and I think tools fundamentally fill this gap first.

這個領域讓我興奮的是參數記憶（parametric memory）、引用如何隨模型而改變、語意連結與 grounding，還有 RAG 的世界。2024 年 Google 演算法外洩時，包括我在內，有多少 SEO 專家能在不到一個月內讓自己的網站衝到第一？答案大概是零。但我們的工作仍在繼續，報告也沒有停。能夠呈現具體資料永遠很重要，我認為工具基本上就是先填補這個缺口。

Be sure your technical SEO is great. It's the first gate of retrieval.

確保你的技術 SEO 很好。它是擷取的第一道關卡。

Make sure you're not blocking LLM user agents. (It's simple but it's the most missing part even in enterprise projects)

確保你沒有封鎖 LLM 的 user agent。（這很簡單，但即使在企業專案裡也是最常被漏掉的部分）

Set your prompts, rotate in every 3 or 6 months (check keyword trends, social platforms, Reddit & Quora real user questions)

設定好你的 prompt，每 3 或 6 個月輪替一次（看關鍵字趨勢、社群平台、Reddit 與 Quora 上的真實使用者問題）。

Make sure you check fanouts monthly. The most winning opportunities are living there. You'll find the serious gaps between your competitors. (I know it's not possible for AIM, AIO but Gemini)

確保每個月檢查 fan-out。最大的機會就藏在那裡，你會發現和競爭對手之間的重大缺口。（我知道 AIM、AIO 做不到，但 Gemini 可以）

Monitor anomalies or up/downs after every new model release, trending/losing domains

每次新模型發布後，監看異常或起伏，以及上升／下滑的網域。

Make your content semantically great. Always use direct answers, be entity rich, do not target only bots.

讓你的內容在語意上做到很好。一律給出直接的答案、富含實體（entity），不要只針對機器人。

Learn more about RAG, LLM dynamics. Always consider these are actually expensive products to give access to the world. So everything needs to be efficient and fast.

多了解 RAG 和 LLM 的運作。永遠記得，這些其實是開放給全世界使用、成本很高的產品，所以一切都必須有效率且快速。

Drink a coffee, review everything, repeat

喝杯咖啡，檢視一切，重複。

ROI & attribution.... The hardest topic alive.

ROI 與歸因……是現存最難的主題。

We're constantly working on a really good case study on this, but our clients often don't want to share their benefits if they see any, so hopefully we can convince a client and share it soon.

我們一直在做一份很棒的案例研究，但客戶即使看到效益，通常也不願意分享，所以希望我們能說服某個客戶，很快把它發表出來。

ROI and AI visibility. They're definitely related; you need a really well-structured analytics system to measure this. You can't achieve this with just a GA4 setup. Moreover, you need to properly classify these traffic patterns and look at direct/dark traffic patterns to accurately measure their impact.

ROI 與 AI 能見度。它們肯定有關聯；你需要一套結構良好的分析系統才能衡量。光靠 GA4 的設定做不到。而且你需要正確地分類這些流量模式，並檢視直接／暗流量（dark traffic）的模式，才能準確衡量它們的影響。

It does have an impact; in the cases I've seen, everyone was developing their own solution.

它確實有影響；在我看過的案例裡，每個人都在開發自己的解決方案。

Fair. Attribution is still the hardest part. Direct and dark traffic still matters and break everything.

沒錯。歸因仍然是最困難的部分。直接與暗流量仍然重要，而且會把一切搞亂。

Within the agentic commerce, it'll be much better.

在 agentic commerce（代理式商務）裡，情況會好得多。

Persona injection is fine. I like the see different approaches. However the biggest issue is LLM memory part. Localization also matters and you're right to address Peec's location injection. We always love reading studies but most of them are addressing the US origin. In Europe, MENA, Asia; even location matters more for some prompts.

Persona injection（角色注入）是可以的，我喜歡看到不同的做法。不過最大的問題是 LLM 的記憶部分。在地化也很重要，你提到 Peec 的 location injection（地點注入）是對的。我們都喜歡讀研究，但多數研究針對的是美國。在歐洲、中東與北非、亞洲，對某些 prompt 來說，地點甚至更重要。

"With memory, past activity, and signed-in context shaping answers more over time, do you thinking trackers will eventually build up context the way a real user’s develops, tracked by market and persona over time? Or is personalization still too weak or inconsistent to change which sources get picked?" It's not possible at the moment. We're working hard for this topic and plan to launch a new experiment. I hope I can share all details with better explanation and results.

「有了記憶、過去的活動與登入後的情境，答案會隨時間越來越被塑造，你認為追蹤工具最終會像真實使用者那樣，按市場與角色隨時間累積情境嗎？還是個人化仍然太弱或太不一致，不足以改變被選中的來源？」目前還不可能。我們正努力研究這個主題，並計畫推出一個新實驗。希望我能以更好的說明與結果分享所有細節。

I saw a pattern in ChatGPT Leak that we found, personal memory matters and if affects to followup questions. Google is doing in a different way. (Connected Google services, Gmail, history, etc)

我在我們發現的 ChatGPT 外洩資料中看到一個模式：個人記憶很重要，而且會影響後續追問。Google 的做法不同（連結 Google 服務、Gmail、歷史紀錄等）。

I can recommend you to follow https://www.linkedin.com/in/garrettsussman/ He's doing amazing research on personalization.

我推薦你追蹤 https://www.linkedin.com/in/garrettsussman/ ，對方在個人化方面做了很棒的研究。

Good question. This question isn't only about AI visibility/answers. It's also about stock market, business side. OpenAI and Google have partnership with Reddit. I always try to follow investor reports from Reddit. You can learn what are their plans for future, some partnership details, etc.

好問題。這個問題不只關乎 AI 能見度／答案，也關乎股市與商業面。OpenAI 和 Google 都和 Reddit 有合作。我總是會追蹤 Reddit 的投資人報告，你可以從中了解他們未來的計畫、一些合作細節等等。

I believe Reddit is valuable, good but it's under AI slop attack which is bad. If Reddit can manage well the AI slop attack, it'd be great for everyone. Otherwise? No value.

我認為 Reddit 很有價值、很好，但它正遭受 AI 垃圾內容的攻擊，這很糟。如果 Reddit 能妥善處理 AI 垃圾內容的攻擊，對大家都好。否則？就沒有價值。

For AIO, AIM, It still matters for me. For ChatGPT? Not important for now, maybe it'll be different next month.

對 AIO、AIM 來說，它對我仍然重要。對 ChatGPT 呢？目前不重要，也許下個月又會不同。

Hey Bogdan! Nice to e-meet you!

嗨 Bogdan！很高興（線上）認識你！

Reddit is still low performing in ChatGPT. Will we see a comeback? Probably. It's still useful for a long time in Google. (we know)

Reddit 在 ChatGPT 裡的表現仍然很低。會不會捲土重來？大概會。它在 Google 裡長期仍然有用。（這我們知道）

My take is it's also about business agreement, I replied below in some few questions, It's always useful to read Reddit's official investor reports. They always share their goals, monthly active users and next plans. Google & OpenAI have agreements with Reddit, maybe they are forcing Reddit on the negotiation side. I don't know.

我的看法是，這也牽涉到商業協議。我在下面的幾個問題裡回答過，讀 Reddit 的官方投資人報告總是很有用，他們都會分享目標、月活躍使用者數與下一步計畫。Google 和 OpenAI 都與 Reddit 有協議，也許他們在談判上對 Reddit 施壓。我不知道。

Is Reddit useful? Yes, it still has effect for influencing people. I just don't like the AI slop just like in X(Twitter) We shouldn't ruin a platform just for AI visibility goals. This is my honest perspective.

Reddit 有用嗎？有，它仍然有影響人的效果。我只是不喜歡 AI 垃圾內容，就像在 X（Twitter）上那樣。我們不應該為了 AI 能見度的目標毀掉一個平台。這是我誠實的看法。

So, I wouldn't leave it, keep monitoring.

所以我不會離開它，持續監測。

It's still new. OpenAI started with partners first, requested huge amount of money. It's only available for free & Go users.

還很新。OpenAI 先從合作夥伴開始，要求非常高的金額。目前只開放給免費與 Go 的使用者。

It's not good at ROI & attribution side. Meta & Google & Amazon are still far beyond that. So, we need to see case studies.

在 ROI 與歸因方面表現不好。Meta、Google 和 Amazon 在這方面遠遠領先。所以我們需要看到案例研究。

We follow every trend, every market. So we always try to keep everything up-to-date. Recently we added Meta Spark, Naver (Korea)

我們追蹤每個趨勢、每個市場，所以總是盡量保持一切是最新的。最近我們新增了 Meta Spark 和 Naver（韓國）。

The overall Peec's tracking list covers on 29th September;

截至 9 月 29 日，Peec 整體的追蹤清單涵蓋：

UI Models:

UI 模型：

ChatGPT, AI Overviews, Gemini, AI Mode, Naver AI Briefing, Perplexity, Copilot

ChatGPT、AI Overviews、Gemini、AI Mode、Naver AI Briefing、Perplexity、Copilot

API Models: Perplexity API, DeepSeek API, Qwen API, Claude Haiku, Mistral Small, Mistral Medium, OpenAI Luna, Claude Sonnet, Meta Spark, Grok API, OpenAI Terra, Claude Opus, OpenAI Sol, Claude Fable, OpenAI Astra (all required credits)

API 模型：Perplexity API、DeepSeek API、Qwen API、Claude Haiku、Mistral Small、Mistral Medium、OpenAI Luna、Claude Sonnet、Meta Spark、Grok API、OpenAI Terra、Claude Opus、OpenAI Sol、Claude Fable、OpenAI Astra（全部都需要點數）

Yes! Last year it was a blast when I first published Perplexity ranking factors. I even asked to Perplexity if it's okay to share. They said yes.

有！去年我第一次發表 Perplexity 排名因素時，引起很大的轟動。我甚至問過 Perplexity 可不可以公開，他們說可以。

It was base64 hashing in the network requests, I asked to the Opus to decode it them all. Then...

那是網路請求裡的 base64 雜湊，我請 Opus 把它們全部解碼。然後……

I have some friends but they are not developers. It's surprising but asking questions in LinkedIn works well! (DM)

我有一些朋友，但他們不是開發人員。說來令人意外，在 LinkedIn 上發問其實很有效！（私訊）

Haha, yes! But I mostly use it for comparison. Recently? An cloud GPU provider.

哈哈，有！不過我大多是用它來比較。最近的一次？一家雲端 GPU 供應商。