2006 年，可用性专家 Jakob Nielsen 用眼动仪记录了两百多个人浏览网页。把很多人的注视叠在一起，热图常常长成一个字母 F：顶上横着读一行，往下再横着读一小段，然后沿着左边一路往下扫。

这个发现很出名，常被拿来证明“读者都很懒”。这篇想换个角度：这不是懒，是精明。读一页文字，很像一只动物在一片林子里找吃的。

先说眼睛本身。读书时，眼睛并不是平平稳稳地滑过一行字，而是一跳一停。停下来叫注视，每次大约四分之一秒；跳过去叫跳视，只要几十毫秒。跳的途中几乎收不到新的画面，但大脑还在处理刚刚看到的字。

每一眼能看清的范围也很小。读英文，大约是注视点左边三四个字母、右边十四五个字母。读中文更窄：左边一个字，右边三个字上下。

所以读者每一眼都在做一个小决定：下一眼落在哪里。又短又常见的词、上下文一看就猜得到的词，常常被直接跳过去。读英文时，三个字母左右的短词，大约三分之二会被跳过。

中文没有空格，眼睛照样知道哪里是一个词。有人做过实验：在中文的词与词之间加上空格，读得并不更快；在每个字之间都加空格，反而读得更慢。上一期讲过，婴儿靠统计就能从连读的音节里切出词来，大人的眼睛大概也在做类似的事。

Nielsen 所在的 Nielsen Norman Group（下面简称 NN/g）后来做了更多眼动研究，发现 F 只是其中一种。下面这八种，是现在常被一起提到的。

F 型是在一大块文字里扫。千层蛋糕是只看标题和小标题，像只吃蛋糕里的奶油层。斑点是在关键词、数字、加粗的字之间跳来跳去，找答案时最常见。沉浸是几乎每个字都读，读小说、读证明、读合同时是这样。

另外几种，和页面的样子有关。锚定，是在手机上让眼睛停在一个地方，用手指把文字一行行推上来。跳过，是在列表里，每一项开头都一样时，直接跳过那几个重复的字。之字，是在左边一张图、右边一段字、再反过来的页面上来回切换；NN/g 发现，这种版式里如果图只是装饰，会打断扫读。

NN/g 自己的结论很值得写东西的人记住：F 型多半出现在没有小标题、没有重点的一大块文字上，读者又不打算认真读的时候。F 型不是读者的毛病，往往是页面没给线索的结果。他们还认为，扫读时最有效的是千层蛋糕；理解得最好的是沉浸，只是最花时间。

为什么会有这些形状？1999 年，施乐帕洛阿尔托研究中心的 Peter Pirolli 和 Stuart Card 提出了“信息觅食理论”：人找信息，和动物找食物用的是同一套算计。

在他们的说法里，一个网页、一段文字，是一块“食物斑块”。标题、链接上的字、加粗的词，是“信息气味”：你还没走进去，就先闻到里面大概有没有吃的。气味越准，觅食的效率越高。

觅食有一个经典问题：在一块地里吃多久，才该去下一块？

设想一只鸟在一丛灌木里找浆果。刚开始很好找，越到后面越难找到下一颗。走到下一丛灌木要花时间，在路上什么也吃不到。它什么时候该走？

1976 年，生态学家 Eric Charnov 给了一个很干净的答案，后来叫“边际价值定理”。翻成大白话就是：当这一丛给你的，已经不如“平均每走一趟能吃到的”，就该走了。

刚进一丛灌木，浆果一颗接一颗，比平均水平好得多，留下来；吃着吃着越来越慢，慢到不如平均水平，再待下去就是浪费时间，该去下一丛了。

这个答案还藏着一个不那么显然的推论：下一丛灌木越远，在每一丛里就越该多待一会儿。路上的时间白白花掉，不如把眼前这一丛吃得干净些。

把灌木换成段落，就是读文章。一段话里，第一句往往给得最多，越往后越是细节和重复，收获递减。读者心里有一个“平均速度”：去别处（下一段、下一篇、下一条消息）平均能得到多少。这一段接下来的话，不如这个平均速度值钱，就该走了。经济学家 Herbert Simon 管这种“够好就停”叫满意即止：人不是在找最好的，是在找够好的。如果你是一字一字读到这里的，请在留言区打“蛋糕”两个字，我想数数有多少位沉浸式读者。那个推论也提醒写东西的人：页面上越容易找到下一段，读者在每一段里就越不肯多待，这不一定是坏事，他们只是走得更有效率。

这些道理能不能真的“长出”F 型？我写了一个很小的模型来试。

一页 6 段，每段 4 行。读者对越往下的段落越不抱期望。在一段里，读到的每个词比上一个更不值钱。读者心里有一个“别处的平均价值”，下一个词不如它，就离开这一段，这正是浆果那条规则。如果页面是一堵墙，读者看不出哪里分段，只能每一行先瞄一眼开头两个字，闻闻有没有吃的；有小标题时，看标题就知道这一段值不值得读；有加粗时，看加粗的词就知道。

结果就是开头那张图。同一个读者，同样的耐心：页面是一堵墙，热图就成了 F，上面两三行读得长，下面只瞄每行开头；有小标题，就成了千层蛋糕；有加粗，就成了斑点；换成一个觉得每句话都有用的读者，几乎整页都被读到。四种形状，都是从“下一个词不值得就走”这一条规则里长出来的。

我本来还想顺便证明“有小标题的文章读得更有效率”，模型却不同意。同一个读者，在一堵墙上反而收获得更多。原因不难找：在模型里，瞄一眼行首两个字，就能像读标题一样知道这一段好不好，标题省下的力气不多，自己还要花几个字去读。

NN/g 说千层蛋糕最有效，我的模型却说不出来。这其实说出了一件真事。一个小标题的价值，等于它比“瞄一眼行首”多告诉了读者多少。“引言”“第二部分”这样的标题，几乎什么也没多说；“什么时候离开一段”这样的标题，读者还没进去，就知道里面有什么。现实里的行首也远没有模型里那么能说明问题。至于在真实的页面上，哪种排版让人收获更多，要靠真实的眼动和理解测验来回答，这个玩具模型回答不了。

觅食讲的是在段落之间怎么挑。落到一个个词上，还有另一条规律，来自信息论。

一个词在上下文里越难猜，它带来的“意外”就越大。“今天天气很”后面接“好”，一点也不意外；接“咸”，就很意外。从 2001 年 Hale、2008 年 Levy 的工作开始，心理语言学家发现，人在一个词上停留多久，跟它带来的意外大小差不多成正比。

2013 年，Smith 和 Levy 用阅读实验的数据检验了这件事，发现这条规律在很大的范围里都成立。2024 年，Shain 等人用六套数据、两百多万个数据点又验证了一遍。这意味着，从“十里挑一”到“百里挑一”，和从“百里挑一”到“千里挑一”，让眼睛多停的时间差不多。

要估计一个词在语境里出现的概率，最方便的工具就是语言模型。研究者发现，语言模型越会预测下一个词，它算出的意外就越能预测人的阅读时间。可这个规律有个拐点：2023 年，Oh 和 Schuler 发现，再往上，参数越多的大模型，算出的意外反而越不像人。问题主要出在人名、地名这类专有名词上：大模型对它们熟得过头，一点也不意外，人却会在这些词上停下来。作者认为，这是因为大模型记住了太多训练文本。

把这些放在一起，对写公众号的人大概有四条用处。

第一，每段第一句放结论。读者在每一段里先吃最甜的那口，吃完多半就走了。

第二，标题要有气味。说出这一段讲什么，别用“引言”“第二部分”这种什么也没说的标题。

第三，加粗要少。加粗是气味，满篇都是加粗，就等于没有气味。这个号的规矩是每节最多加粗一处，这篇也照做了。

第四，想让人沉浸，就给他一个想知道答案的问题。沉浸不是逼出来的，是读者觉得每一句都值得。

所以，读者跳着读，不是不尊重你，是在算账：时间有限，注意力有限，他要在一页字里尽快找到值钱的东西。

也正因为这样，一个人肯一字一字地读完你写的东西，是一件很难得的事。他把有限的时间，押在了你这一块地里。

顺便说一句，文中有一段藏了一句话，只有一字一字读的人才会看见。

读者不是懒，
是在算账。

In 2006 the usability expert Jakob Nielsen used eye trackers to record more than two hundred people browsing the web. Stack many people's fixations on top of each other and the heatmap often takes the shape of the letter F: one line read across the top, a shorter stretch read across a little lower down, then a sweep straight down the left side.

The finding is famous, and is often cited as proof that “readers are lazy.” This essay takes another view: it isn't laziness, it's shrewdness. Reading a page is a lot like an animal looking for food in a patch of woods.

Start with the eyes themselves. When you read, your eyes don't glide smoothly along a line. They jump and stop, jump and stop. A stop is called a fixation and lasts about a quarter of a second; a jump is called a saccade and takes only a few tens of milliseconds. During a jump you take in almost nothing new, but the brain is still working on the words it just saw.

The area you can see clearly in each glance is small, too. In English it runs about three or four letters to the left of the point you're fixating and fourteen or fifteen to the right. In Chinese it is narrower: one character to the left and about three to the right.

So every glance involves a small decision: where should the next one land? Words that are short and common, or easy to guess from the context, often get skipped entirely. In English, short words of about three letters are skipped roughly two-thirds of the time.

Chinese has no spaces, yet the eyes still know where one word ends and the next begins. In one experiment, putting spaces between the words of a Chinese text did not make it faster to read, and spacing out every single character made it slower. Last issue showed that babies can carve words out of a stream of syllables using statistics alone; adult eyes are probably doing something similar.

Nielsen's own firm, the Nielsen Norman Group (NN/g from here on), went on to do much more eye-tracking research and found that the F is only one pattern among several. The eight below are the ones most often mentioned together today.

The F-shaped pattern is scanning a big block of text. The layer-cake pattern is looking only at headings and subheadings, like eating just the cream layers of a cake. The spotted pattern is jumping between keywords, numbers and bold words, most common when you're hunting for an answer. The commitment pattern is reading nearly every word, the way you read a novel, a proof or a contract.

The others have to do with how the page looks. The marking pattern is keeping your eyes in one spot on a phone while your finger pushes the text up line by line. The bypassing pattern happens in lists where every item starts the same way: you skip straight past the repeated words. The zigzag pattern is switching back and forth on a page where an image on the left sits beside text on the right, then the other way around. NN/g found that in such layouts, images that are only decoration interrupt scanning.

NN/g's own conclusion is worth remembering for anyone who writes: the F-shaped pattern mostly shows up on a big block of text with no subheadings and nothing highlighted, when the reader isn't planning to read carefully. The F-shaped pattern isn't the reader's fault; it's usually what happens when the page gives no clues. They also judge the layer-cake pattern the most effective way to scan. The commitment pattern gives the best understanding, but it takes the most time.

Why do these shapes appear? In 1999 Peter Pirolli and Stuart Card of Xerox PARC (the Palo Alto Research Center) proposed “information foraging theory”: people looking for information use the same calculus as animals looking for food.

In their terms, a web page or a passage of text is a “patch” of food. Headings, the words on links and bold words are “information scent”: before you even step inside, you can smell whether there's likely to be food in there. The more accurate the scent, the more efficient the foraging.

Foraging has a classic question: how long should you eat in one patch before moving on to the next?

Picture a bird picking berries in a bush. At first they're easy to find; the longer it stays, the harder it is to find the next one. Getting to the next bush takes time, and there's nothing to eat on the way. When should it leave?

In 1976 the ecologist Eric Charnov gave a very clean answer, later called the marginal value theorem. In plain words: once this bush is giving you less than what an average trip gets you, it's time to go.

When the bird first arrives, the berries come one after another, much better than average, so it stays. As it eats they come more and more slowly, until they fall below average. From then on, staying just wastes time, and the bird should move on to the next bush.

The answer hides a less obvious corollary: the farther away the next bush, the longer the bird should stay in each one. Time on the road is simply lost, so it's better to pick the bush in front of it a little cleaner.

Swap bushes for paragraphs and you have reading. In a paragraph the first sentence usually gives the most; after that come details and repetition, with diminishing returns. Readers carry an “average rate” in their heads: how much they would get, on average, by going elsewhere (the next paragraph, the next article, the next message). Once the rest of this paragraph is worth less than that rate, it's time to go. The economist Herbert Simon called this “stop when it's good enough” satisficing: people aren't looking for the best, only for good enough. If you have read every word to get here, please type the word “cake” in the comments; I'd like to count how many committed readers there are. The corollary holds a lesson for writers too: the easier a page makes it to find the next paragraph, the less willing readers are to linger in any one of them. That isn't necessarily bad; they're just moving more efficiently.

Can these ideas really “grow” an F? I wrote a very small model to find out.

A page has 6 paragraphs of 4 lines each. The reader expects less and less from paragraphs further down. Within a paragraph, each word read is worth a little less than the one before. The reader has an “average value elsewhere” in mind, and as soon as the next word is worth less than that, they leave the paragraph: exactly the berry rule. If the page is a wall of text, the reader can't tell where the paragraphs break, and can only glance at the first two words of each line to sniff for food. With subheadings, the heading tells them whether a paragraph is worth reading; with bold text, the bold words do.

The result is the figure at the top. Same reader, same patience. On a wall of text the heatmap turns into an F: the top two or three lines are read a long way across, and further down only the start of each line gets a glance. With subheadings it becomes a layer cake; with bold it becomes spotted; switch to a reader who finds every sentence useful and almost the whole page gets read. All four shapes grow out of one rule: leave as soon as the next word isn't worth it.

I had also hoped to prove, along the way, that “articles with subheadings are read more efficiently.” The model disagreed. The same reader actually gained more from the wall of text. The reason isn't hard to find. In the model, a glance at the first two words of a line tells you whether a paragraph is good just as well as a heading does, so a heading saves little effort and costs a few words of its own to read.

NN/g says the layer cake is the most effective pattern, yet my model can't show it. This actually points to something true. A heading is worth exactly what it tells readers beyond a glance at the start of the line. Headings like “Introduction” or “Part Two” add almost nothing; a heading like “When to leave a paragraph” tells readers what's inside before they go in. And in real life, the start of a line says far less than it does in the model. As for which layout really lets people gain more on a real page, that takes real eye tracking and comprehension tests to answer. This toy model can't.

Foraging is about choosing between paragraphs. Down at the level of single words there is another regularity, and it comes from information theory.

The harder a word is to guess from its context, the more “surprise” it carries; the technical name is surprisal. After “The weather today is very,” the word “nice” is no surprise at all; “salty” is a big one. Starting with work by Hale in 2001 and Levy in 2008, psycholinguists have found that how long people linger on a word is roughly proportional to how much surprise it brings.

In 2013 Smith and Levy tested this against data from reading experiments and found that the rule holds over a very wide range. In 2024 Shain and colleagues checked it again, with six datasets and more than two million data points. So going from “one in ten” to “one in a hundred” adds about as much reading time as going from “one in a hundred” to “one in a thousand.”

The handiest tool for estimating how likely a word is in context is a language model. Researchers found that the better a language model predicts the next word, the better its surprisal predicts human reading times. But the trend has a turning point. In 2023 Oh and Schuler found that past a certain size, the more parameters a large model has, the less human-like its surprisal becomes. The trouble lies mainly with proper nouns such as the names of people and places: large models know them so well that they aren't surprised at all, yet people slow down on these words. The authors put this down to large models having memorized too much of their training text.

Put all this together and there are roughly four lessons for anyone who writes on WeChat.

First, put the conclusion in each paragraph's first sentence. Readers take the sweetest bite of each paragraph first, and most of them leave after that.

Second, give your headings a scent. Say what the section is about; don't use headings like “Introduction” or “Part Two” that say nothing at all.

Third, go easy on bold. Bold is scent, and when the whole page is bold there is no scent at all. This account's rule is at most one bold phrase per section, and this essay sticks to it.

Fourth, if you want people to read every word, give them a question they want answered. Commitment can't be forced; it comes when readers feel every sentence is worth it.

So when readers skip, they aren't disrespecting you. They're doing the math: their time is limited, their attention is limited, and they want to find what's valuable on a page as fast as they can.

That is also why it's a rare gift when someone reads what you wrote word by word. They have staked their limited time on your patch.

By the way, one paragraph in this essay hides a sentence that only word-by-word readers will see.

Readers aren't lazy;
they're doing the math.