一个号码，跟随。
我们问了ChatGPT,Perplexity,Gemini,Claude和Google的AI概览200个需要数字的问题，就是人们每天问的那种问题。 附注1A 统计是助理人员声称为事实的份额、比率、平均数、计数或费用。 我们忽略了建议和价格(“节省你收入的20%”)。 然后我们点击引文，继续点击。 这里有一个线索，从答案到结尾。
Google AI 概览
在一次后续调查之后，销售人员放弃的比例是多少？
44%的销售人员在一次后续尝试后放弃。
- 44%
- 44%
- 44%
- 44%
- 44%
44%的销售人员在一次后续尝试后放弃(来源).
80%的销售需要5次会后后续电话；44%的销售人员在一次后续之后放弃
没有调查没有样本没有日期
- 在答复中，每个助理都有一个同样的问题：“在一次后续调查之后，销售人员放弃的比例是多少？” 所有5名助手都给出了同样的数字，只有ChatGPT 称背后的证据不清楚。
- 在Google AI Overviews中一击引用了一家销售公司的统计页面，从2020年开始。 号码就在那儿其“来源”链接涉及一个书面样本。
2017年，数字在它的头条。 它列出了两个幻灯片甲板，一个博客列表和两个新闻报道。 无一是对后续行动的研究。
这条路的尽头就是终点。 没有调查，没有样本，没有研究：我们尝试的每一个通往这个数字的路线都停留在一页没有引用任何研究。 照片里还有些数字，这是僵尸统计。
大部分数字不是僵尸。 但你无法从答案中分辨出：一个僵尸带着同样的自信和同样的引用，来到这里，所有五个助手都同意。 因此，我们在1,285项统计中也做了同样的事情。
不到一半到达那里。
我们随机抽取了每个助手所引用的1,285个统计数据样本，走的每条线索都一样：被引用的一页，然后是该页的来源，最深可达5页。 注2 每个助理的样本都是从所引用的统计数据中随机抽取的：ChatGPT 108,Perplexity 277,Gemini 300,Claude 300,Google AI Overviews 300. 5名助学率根据引用的统计资料对每个助学者进行加权，因此它描述了所显示的统计数据。 在一个没有链接的名为源的页面中，我们搜索了它。
40%的研究发现称也是一样另有8%进入研究范围，但数字或背景在路上有所变化。 其余的人从未到过那里：有40%的人停留在一个我们无法查证的源头，因为它被锁住、消失或对数字保持沉默，有11%的人一无所获。 注 3 A AI模型判断了每一条线索，而更强大的一个则重新判断了300条线索。 然后人们随机重新检查了150个线索，只要两人不同意，就进行第三次检查。 人们发现该模型错过了原作：仅通过模型，23%的来源。 这里的每个费率都通过手检查得到纠正，其范围带有两者的不确定性。
ChatGPT成绩最好，并引用了最少的数字：每回1.3，而Claude为6.8. 其他四人登陆率在38%至46%之间，过于接近排名。
我们的问题并不奇怪。 在Google AI Overviews对真实日常搜索的300个统计中，40%是来源，份额相同。 注：4Google AI Overviews回应了Trakkr在2026年9月25日至2026年10月25日期间追踪到的52,930个日常英语搜索。 4.4%的数据，我们追踪到300个同样的数据。 只有计数和费率离开我们的系统。
大赦国际援引的是副本，而不是研究报告。
大部分时间，AI没有编出数字：它引用的页面有82%的时间。 麻烦是哪页引用的一些网页中只有21%是研究本身。
超过一半是公司博客。 加上新闻报道和统计综述，大部分引用内容都出现在为收集别处的数据而建的网页上
踪迹从何而起决定所起。 从一家研究公司或政府网页上，它达到了匹配研究的61%. 从一个公司的博客，37%；从统计综述，35%. 甚至研究公司也少了38%的时间，常常是因为报告全文在表格后面。
一个引用告诉你AI从哪里找到一个号码，而不是这个号码的来源。
数字在路上变化。
访问研究并不一定意味着所持数字。 在获得的数据中，17%在路上有所变化。 数字四分之一只移动过一次通常是上下文：测量到的，年份，来源，或者问到谁。
改变的，我们记录的每一个变化
- 号码变了
- 26%
- 测量到的改变
- 20%
- 一年改变了
- 17%
- 源头变了
- 16%
- 谁被改了
- 14%
- 地方变了
- 4%
15名妇女判断健康网站成为每个网站的第一个印象。 三家公司成为了商业规则。 这里有我们所跟踪的最清楚的线索，每一个都用手检查。 注：5我们称原研究。 重复它们的页面按类型描述，而不是命名。 这些线索的每一步都用手在公开的网上检查，并保存存档副本。
僵尸们
问五个助手同样的问题，在所有五个人给出一个百分比的71个问题中的28个，我们得到了4个或5个不同的数字。 注6 我们比较了每个助理在同一问题的标题数字。 一个范围在中间算起，2分内的数字是相同的。
这不是一个奇怪的问题。 最高数字与最低数字之间的典型差距为16分。
协议是一种证明。 事实并非如此。 所有五个人给出了同样的数字对于10个问题，而我们追踪到的5个，2是来源。 他们最多重复的、有3个或更多人引用的统计资料没有更好：29个中有11个。
而4个被所有5个助手所引用，在他们身后没有任何研究。
AI统计有什么用？
视之为先导，而非从源。 取出六道习取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取出取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取取来。取取取取取取取取取取取取取取取取取取取取取。取取取取取取取取取取取取取取取取取取取取
如果你用它们
- 跟着引用，然后再跟着它。 如果引用的页面是统计列表或公司博客，请使用自己的链接再往前走一页。 直到你到达研究本身
- 看看是谁问的，有多少，什么时候问的我们发现的大多数变化都发生在背景上，而不是数字上：一个国家的民调，一个十年的研究，15人。
- 不要把协议当作支票。 提供同样人数的5名助理并非5个来源。 他们最多重复的数字没有其他来源更好。
如果你出版的话
- 把方法放在号码旁边取出样本，取出日期和与出处的链接，与图同一页，于是取出下页，再取出下回回回回回回回回回。
- 把发现公开一份表格背后的报告是一个常见的死胡同：即使是从一家研究公司开始的踪迹也不足38%的时间。 将标题编号和方法放入开页。
- 看AI为你们的号码引用的哪一页通常都是一个综合体或博客复制你，那页决定人们听到什么。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
One number, followed.
We asked ChatGPT, Perplexity, Gemini, Claude and Google’s AI Overviews 200 questions that call for a number, the kind people ask every day.Note 1A statistic is a share, rate, average, count or cost that an assistant states as fact. We left out advice (“save 20% of your income”) and prices. Then we clicked the citations, and kept clicking. Here is one trail, from the answer to the end.
Google AI Overviews
What percentage of salespeople give up after one follow-up?
44% of salespeople give up after just one follow-up attempt.
- 44%
- 44%
- 44%
- 44%
- 44%
44% of salespeople give up after one follow-up attempt (source).
80 Percent of Sales Require 5 Follow-up Calls after Meeting; 44 Percent Salespeople Give Up After One Follow-up
No study behind itNo survey, no sample, no date.
- In the answer Every assistant got the same question: “What percentage of salespeople give up after one follow-up?” All five assistants gave the same figure, and only ChatGPT said the evidence behind it was unclear.
- One click in Google AI Overviews cited a sales firm’s stats page, from 2020. The number is right there. Its “source” link goes to a writing sample.
- Two clicks in That leads to a freelance writer’s sample article, online by 2017. The number is in its headline. It lists two slide decks, a blog list and two news stories. None is a study of follow-ups.
- The end of the trail That’s where it ends. No survey, no sample, no study: every other route to this number we tried stopped at a page that cites no research. It’s a number still walking around with nothing underneath it. That’s a zombie stat.
Most numbers aren’t zombies. But you can’t tell from the answer: a zombie arrives with the same confidence and the same citation, and here, with all five assistants agreeing. So we did the same for 1,285 statistics.
Fewer than half get there.
We took a random sample of the statistics each assistant cited, 1,285 in all, and walked every trail the same way: the cited page, then that page’s source, up to five pages deep.Note 2Each assistant’s sample was drawn at random from the statistics it cited: ChatGPT 108, Perplexity 277, Gemini 300, Claude 300, Google AI Overviews 300. The five-assistant rate weights each assistant by how many cited statistics it quoted, so it describes the statistics people were shown. Where a page named a source without a link, we searched for it.
40% reached a study that says the same thing. Another 8% reached the study, but the number or its context had changed on the way. The rest never got there: 40% stopped at a source we couldn’t check, because it was locked, gone or silent on the number, and 11% led nowhere at all.Note 3An AI model judged every trail, and a stronger one re-judged 300. Then people re-checked 150 trails at random, with a third check wherever the two disagreed. People found originals the model missed: by the model alone, 23% were sourced. Every rate here is corrected with the hand checks, and its range carries the uncertainty from both.
ChatGPT did best, and quoted the fewest numbers: 1.3 per answer, against 6.8 for Claude. The other four landed between 38% and 46%, too close to rank.
It isn’t a quirk of our questions. Of 300 statistics in Google AI Overviews answers to real everyday searches, 40% were sourced, the same share.Note 4Google AI Overviews answers to 52,930 everyday English searches that Trakkr tracked between 25 September 2026 and 25 October 2026. 4.4% carried a statistic, and we traced 300 of those the same way. Only counts and rates leave our systems.
AI cites the copy, not the study.
Most of the time, the AI wasn’t making numbers up: the page it cited had the figure 82% of the time. The trouble is which page. Only 21% of the pages cited for a number were the research itself.
More than half were company blogs. Add news stories and statistics roundups, and most citations land on pages built to collect figures from elsewhere, often without a link back.
Where the trail starts decides where it ends. From a research firm or a government page, it reached a matching study 61% of the time. From a company blog, 37%; from a statistics roundup, 35%. Even research firms fell short 38% of the time, often because the full report sat behind a form.
A citation tells you where the AI found a number, not where the number came from.
Numbers change on the way.
Reaching the study didn’t always mean the number held. Of the statistics that got there, 17% had changed on the way. The digits moved only about one time in four. Usually it was the context: what was measured, the year, the source, or who was asked.
What changed, across every change we logged
- The number changed
- 26%
- What was measured changed
- 20%
- The year changed
- 17%
- The source changed
- 16%
- Who was asked changed
- 14%
- The place changed
- 4%
Fifteen women judging health websites become every website’s first impression. Three companies become a rule of business. Here are the clearest trails we followed, each one checked by hand.Note 5We name the original studies. The pages that repeated them are described by type, not named. Every step in these trails was checked by hand on the open web, with archived copies kept.
The zombies.
Ask five assistants the same question and you’ll often get five different numbers. On 28 of the 71 questions where all five gave a percentage, we got four or five different figures.Note 6We compared each assistant’s headline figure for the same question. A range counts at its middle, and figures within 2 points count as the same.
It wasn’t one odd question. The typical gap between the highest figure and the lowest was 16 points.
Agreement feels like proof. It isn’t. All five gave the same figure on 10 questions, and of the 5 we traced, 2 were sourced. The statistics they repeated most, quoted by three or more, did no better: 11 of 29.
And four were quoted by all five assistants, with no study anywhere behind them.
What to do with an AI statistic.
Treat it as a lead, not a source. Six habits, three for each side of the citation.
If you use them
- Follow the citation, then follow it again. If the cited page is a list of statistics or a company blog, use its own link to go one page further. Keep going until you reach the study itself, or until you can’t.
- Check who was asked, how many, and when. Most of the changes we found were in the context, not the digits: one country’s poll, a ten-year-old study, fifteen people.
- Don’t take agreement as a check. Five assistants giving the same number isn’t five sources. The numbers they repeated most were no better sourced than the rest.
If you publish them
- Put the method next to the number. The sample, the date and a link to the source, on the same page as the figure, so the next page, and the next answer, can trace it back to you.
- Keep the findings in the open. Reports behind a form were a common dead end: even trails that started at a research firm fell short 38% of the time. Put the headline numbers and the method on an open page.
- See which page AI cites for your numbers. It is often a roundup or a blog that copied you, sometimes with an older figure. That page decides what people hear.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.