调研显示仅四成 AI 引用数据可溯源至原始研究

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.1

• 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.

1. 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. 2. 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. 3. 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. 4. 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.

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.

Four in ten trails reached a study that says the same thing

Each square is one in a hundred of the 1,285 statistics we traced, sorted by where its trail ended.

• 40%SourcedWe found the study, and it says the same thing.
• 8%ChangedWe found the study, but the number, who was asked, where or when is different.
• 40%Can’t checkA page names a real source, but the number isn’t in it, or the source is locked or gone.
• 11%NowhereNo study anywhere on the trail. Every page repeats the number and credits no one.

Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. A random sample of each assistant’s cited statistics, weighted by how many each quoted, and corrected with 150 trails checked by hand. Sourced: 40%, 95% range 34% to 46%.

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.

ChatGPT quoted the fewest numbers, and sourced the most

The same hundred squares for each assistant, from its own sample of cited statistics.

1. ChatGPT

57% sourced

1.3 statistics per answer 2. Perplexity

46% sourced

2.2 statistics per answer 3. AI Overviews

39% sourced

4.8 statistics per answer 4. Gemini

38% sourced

6.2 statistics per answer 5. Claude

38% sourced

6.8 statistics per answer

Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. Traced per assistant: ChatGPT 108, Perplexity 277, Google AI Overviews 300, Gemini 300, Claude 300. The four after ChatGPT sit inside each other’s ranges, so they aren’t ranked.

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.

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

Every page an assistant cited first for a number, by kind. Each column’s width is its share of those citations. Green is how often its trails reached a study that says the same thing.

62%

Research firm13% of citations

Research firm or survey: 13% of first citations, 62% of them sourced.61%

Government7% of citations

Government or official body: 7% of first citations, 61% of them sourced.37%

Company blog54% of citations

Company blog: 54% of first citations, 37% of them sourced.37%

News10% of citations

News or magazine: 10% of first citations, 37% of them sourced.35%

Stats roundup8% of citations

Stats roundup: 8% of first citations, 35% of them sourced.41%

Other8% of citations

Other: 8% of first citations, 41% of them sourced.Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. 1,268 traced statistics with a readable first page. Page types judged from the page itself. Outcomes corrected with 150 trails checked by hand.

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.

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 changed26%What was measured changed20%The year changed17%The source changed16%Who was asked changed14%The place changed4%

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.

Watch a number lose its context

Each trail runs from the original study to an answer an assistant gave in October 2026. Step along it: the number stays, and the words around it change.

AskedWhat percentage of first impressions of a website are design-related?

The study · Sillence and colleagues, CHI, 2004 Of the comments 15 women made about rejecting or mistrusting health websites, 94% were about design.

Repeated “Research shows us that 94 percent of people notice the look-and-feel of a website or application first” What changedComments about rejecting health sites become what people notice first

Repeated “Since 94% of first impressions are related to design, leveraging captivating visuals across your brand's website…” What changedNow 94% of all first impressions

The answer “About 94% of a website's first impressions are design-related.” What changedStated as fact for every website

A 2004 study watched 15 British women judge health websites, and 94% of their comments about rejecting a site were about design. Twenty years of retelling made it “94% of first impressions are design-related”, and Google AI Overviews states it as fact.

Step 1 of 4Every step checked by hand on the open web, with archived copies kept. We name the original studies; the pages that repeated them are described by type, not named.

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.

One question, five answers, from 49% to 96%

Asked “What percentage of marketers use marketing automation tools?”, word for word, on the same day.

0%100%**49–51%AI Overviews62%ChatGPT70–79%Gemini81%Perplexity96%**Claude

• ChatGPTThere isn’t one universal figure, it varies by survey and what counts as “marketing automation.” One commonly cited survey found 62% of U.S. digital marketers used automation software, but that data is from 2020, so it may not reflect current adoption.
• Perplexity****Approximately 81% of marketing organizations use marketing automation.
• GeminiAcross major industry benchmarks, between 70% and 79% of marketers and companies use marketing automation tools in some capacity
• Claude****96% of marketers have used a marketing automation platform or plan to use one
• AI OverviewsRoughly 49% to 51% of companies and marketing teams currently use marketing automation tools, according to industry benchmarks from Moosend and related research studies.

Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. Answers collected on 25 October 2026. A range is drawn from its low to its high.

It wasn’t one odd question. The typical gap between the highest figure and the lowest was 16 points.

On 31 of 71 questions, the five figures were 20 points or more apart

Every question where all five gave a percentage, one line each, from the lowest figure to the highest. Closest at the top.

Closest20 points or more apartFurthest apart0%50%100%Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. A range counts at its middle. 2 questions with figures above 100% are left off the scale.

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.

Quoted by all five. Sourced by none.

Four statistics every assistant gave us, each with a trail that ends at a page citing no research.

• 44%

Salespeople who give up after one follow-up

Every trail ends at a page that cites no research, one of them a freelance writer’s sample article. Its full trail opens chapter 1.

Quoted by all five

• 90%

Social media users who follow at least one brand

Stated with no source. No study with this figure was found.

Quoted by all five

• 75%

Searchers who never scroll past the first page of results

Repeated with no source, often credited to an old marketing roundup that no longer carries it. No study was found.

Quoted by all five

• 5 to 25 times

How much more it costs to win a customer than to keep one

A 2014 Harvard Business Review article gives the range “depending on which study you believe”, and names none.

Quoted by all five

Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. Statistics all five assistants quoted whose every trail ended at a page citing no research. Each trail was checked by hand on the open web.

The 29 statistics AI repeats most

Each was quoted by at least three of the five assistants, most often first, and its trail checked by hand.

1. 95%

Generative AI pilots that fail to deliver a measurable financial return

MIT NANDA’s 2025 report says 95% of organizations got no return from generative AI. That isn’t the same as 95% of pilots failing.

5 of 5ChangedStudy from 2025 2. 44%

Salespeople who give up after one follow-up

Every trail ends at a page that cites no research, one of them a freelance writer’s sample article. Its full trail opens chapter 1.

5 of 5Nowhere 3. 4% to 10%

Unhappy customers who actually complain

Goes back to TARP’s 1979 research for the White House consumer office. Its author says rates that low apply to small purchases; in 2016, 22% to 41% complained about their most serious problem.

5 of 5ChangedStudy from 1979 4. 2 hours 21 minutes

Time people spend on social media each day

DataReportal’s Digital 2025 report, from GWI survey data: the typical internet user spends 2 hours 21 minutes a day on social media.

5 of 5SourcedStudy from 2025 5. 46%

Google searches with local intent

Traces to a 2018 social post about a slide at a search conference. Google never published it, as far as we could find.

5 of 5Can’t check 6. 90%

Social media users who follow at least one brand

Stated with no source. No study with this figure was found.

5 of 5Nowhere 7. 94%

First impressions of a website that are about its design

Began as a 2004 study of how 15 women judged health websites. Its full trail is in chapter 4.

5 of 5ChangedStudy from 2004 8. 75%

Searchers who never scroll past the first page of results

Repeated with no source, often credited to an old marketing roundup that no longer carries it. No study was found.

5 of 5Nowhere

Trakkr Research, five AI assistants asked 200 questions that call for a number, October 2026. Statistics grouped by figure and meaning. Original year shown where the trail reached the study.

If you use them

1. 1

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. 2. 2

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. 3. 3

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

1. 1

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. 2. 2

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. 3. 3

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.

Questions200160 everyday questions in 16 topics, plus 40 famous statisticsAnswers1,000From ChatGPT, Perplexity, Gemini, Claude and Google AI OverviewsStatistics quoted3,1672,419 with a citation, in answers to the everyday questionsTrails followed1,285A random sample of cited statistics, up to five pages deep; 180 checked by hand

We asked each assistant 200 questions that call for a number, word for word, through its company’s API with web search on and a US location: 160 everyday questions across 16 topics, and 40 famous statistics we suspected, reported separately. We pulled every statistic out of every answer, with the page it cited.

For a random sample of each assistant’s cited statistics (every one of ChatGPT’s, up to 300 from each of the others), we opened the cited page, found the figure, and followed the page’s own source, up to five pages deep. Where a page named a source without a link, we searched for it. An AI model judged each trail, and a stronger model re-judged 300 of them, agreeing on the outcome 82% of the time.

Then we checked 150 trails by hand, picked at random, with a third check wherever the hand check and the model disagreed. A trail we checked by hand counts as we found it; every other trail counts as the mix of outcomes we found by hand behind the model’s verdict. The ranges carry the uncertainty from both samples.

Rates for each assistant cover its own sample. The five-assistant rate weights each assistant by how many cited statistics it quoted, so it describes the statistics people were shown. Every statistic shown on its own on this page was checked by hand.

The machine judge was stricter than people

The AI judge alone23%Corrected with 150 hand checks40%

Share of traced statistics sourced. People found originals the judge missed, so every rate on this page uses both.

Where it falls short

• API answers

We asked through each company’s API with web search on and a US location. The apps people use can answer differently.

• One day

Every question was asked once, on 25 October 2026. Answers change from day to day.

• A machine judge

An AI model judged every trail and was stricter than a person. We corrected it with 150 hand checks; on its own it found 23%.

• Locked sources

Some originals sit behind paywalls or sign-up forms. We count those as a source we couldn’t check, not as missing.

Notes

1. A 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.
2. Each 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.
3. An 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.
4. Google 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.
5. We 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.
6. We 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.

The data

zombie-stats.json

Aggregates, our own questions and the hand-checked trails. No customer data, and no names of pages that repeat a statistic.

Cite this study

Grenfell, M. (2026, October 28). Zombie stats (Trakkr Research Study 013). Trakkr. https://trakkr.ai/trakkr-research/zombie-stats

Questions,

answered.

The ones we hear most about this study.

Rarely, in our sample. The page an assistant cited contained the figure 82% of the time. The bigger problem is upstream: the pages AI cites are mostly blogs and roundups repeating a number. Of the statistics we traced, four in ten reached a study that says the same thing.

In our sample, ChatGPT did best: 57% of the statistics it quoted reached a matching original, and it quoted the fewest. The other four were between 38% and 46%. Every assistant quoted numbers that led nowhere.

Open the source it cites. If that page is a list of statistics or a company blog, follow its link to the next page, and keep going until you reach the study itself. Check the year, who was asked and how many, then quote the study, not the page that repeated it.

A number that keeps being repeated after its source has changed, aged or never existed. It survives because each page copies the last one, and AI assistants now repeat it to millions of people.

Yes, with attribution. Cite Trakkr Research Study 013 and link to this page. The aggregate data is a free JSON download.

Ask anything about this studyAsk anything about this study

The Agent answers from our docs, methods and pricing, and links to the source.

Pages AI cites for Lope this week

lope.exampleChatGPTYour 2026 runner survey, in 14 answersPerplexityA stats roundup quoting your surveyA copyGeminiYour survey method page, in 9 answersClaudeA news story that cites your surveyA copyAI OverviewsA 2019 blog with an older figureOut of date

Be the source AI can trace.

Trakkr shows which pages ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews cite when they answer questions about your brand, and which of your own pages they pass over.