Receipts That Don't Hold: What AI Search Really Cites

The reel opens with a confession. "I've logged every source that my AI has given me since June. Not the answers, but the sources or the citations. And this really weird thing has been happening. I recognize almost none of them." The creator writes a newsletter on technology and was researching how AI recommends products. What follows is a fast chain of explanations. Consumer Reports, which has tested products since 1936, blocks the crawlers of OpenAI, Google and Common Crawl in its robots file. Sixty percent of high-credibility news sites block AI crawlers. Bot traffic now exceeds human traffic. So when an assistant searches, "it's reading a part of the internet that still lets those crawlers in." Researchers checked 26,000 cited sources and found one in six written by another AI. Forty-four percent of what ChatGPT cites comes from best-of listicles, some written by brands ranking themselves first. And there is "post hoc citation": the model "writes the answer first and then goes and looks for sources that support that answer." The close: "When ChatGPT or Claude gives you a link, click it, see what's there. See if your answer is on that page."

I liked it, and I did what it says. I followed every number back to its source. Each one comes from a real study, a real file or a real report. Most were widened on the way into the reel: a narrow finding became a general one, a detector's estimate became a fact, a training rule became a browsing rule. That is the same thing the research says happens between a web page and an AI answer, and it happens in human retellings too.

So the reel is accurate about its most important point. A citation is a receipt for a claim. In an earlier essay I argued that in an age of cheap images, provenance needs a receipt. This one is about what happens when the receipts themselves do not hold.

1. The numbers, followed back

Here is the ledger, before the detail. Consumer Reports was founded in 1936, and its robots file does name the three crawlers the reel reads out; but they control model training, and the file leaves the search and browsing agents of OpenAI, Anthropic and Perplexity unnamed. The 60% figure is real and means something narrower than "block." Bots do outnumber humans in two vendors' measurements, and nothing ties that to why publishers block AI. The "unlocked web" exists, and its borders are drawn per company by contracts and server rules more than by the lines the reel reads aloud. "One in six AI-written" is a detector's estimate, on politics and health questions, varying fourfold between engines. "Forty-four percent listicles" holds for product-recommendation prompts only. And "post hoc citation" names something real that is not how either OpenAI or Anthropic documents its search. The creator's own log of unfamiliar sources cannot be checked, and it is what the citation data would predict.

The final advice, to click and check, needs no correction at all.

2. The locked door that isn't

The reel's first example holds up as a reading of a file. Consumer Reports' robots.txt, which tells automated visitors where they may go, gives a whole-site disallow to GPTBot, Google-Extended and CCBot, Common Crawl's crawler. The lines have been there since at least December 2023, according to the Internet Archive's copies. Consumers Union, the organization behind the magazine, was founded in February 1936.

What the lines do is another matter. GPTBot is OpenAI's crawler for collecting training data. Google-Extended is not a crawler at all: Google's documentation calls it a control token that governs whether content is used to train Gemini and to ground Gemini's answers, and says it "does not impact a site's inclusion in Google Search" (Google). CCBot feeds Common Crawl, the open web archive used to train most of the large language models released between 2019 and 2023. None of the three is the agent that fetches a page when someone asks ChatGPT or Claude a question. The file names no agent from Anthropic or Perplexity and no OpenAI search or browsing agent. Under the robots standard, those fall under the file's general rules, which allow articles and product pages.

As part of the research for this essay, single test requests were sent to Consumer Reports' server on 6 November, from one location, changing only the name each request gave for itself. The server refused the names of OpenAI's training crawler and its browsing agent, and served the names used by Anthropic's and Perplexity's agents in full. That test has limits, since real crawlers arrive from known addresses, but it shows that the actual gatekeeping happens at the server, not in the file. And the thing worth protecting, the test ratings, sits behind a membership paywall, which keeps out every bot and every reader who has not paid. Consumer Reports also runs its own AI advisor for members. "They just don't want their testing scraped" fits all of this. I found no statement from the organization saying so.

3. Sixty percent, and the asymmetry behind it

The 60% comes from a peer-reviewed paper by Nicolas Steinacker-Olsztyn, Devashish Gosain and Ha Dao, presented at the ACM Web Conference in 2026 under the title "Is Misinformation More Open?" (Steinacker-Olsztyn et al. 2026). They classified 3,369 news sites as reputable and 710 as misinformation, using an independent rating service, and read their robots files. "Among websites with a robots.txt file, 60.0% of reputable sites include a DisallowAll directive for at least one AI agent, compared to just 9.1% of misinformation sites." The list had 63 agents. For any single major agent, the share among reputable sites in May 2025 ran from 39% to 53%. The paper does not separate training agents from browsing agents.

Other measurements land in the same range. By the end of 2023, 48% of the most-used news sites in ten countries were blocking OpenAI's crawlers, from 79% in the United States to 20% in Mexico and Poland (Fletcher 2024). Across the web as a whole the picture is different: about 5% of the text in a widely used training corpus was fully restricted by 2024 (Longpre et al. 2024). So "that's happening everywhere" is wrong in an interesting way. It is happening among the sources with the most to lose.

That asymmetry is the finding I would carry from this part of the research. Reputable publishers are more than six times as likely as misinformation sites to ask AI agents to stay out. Whatever stays open, by default, skews toward weaker sources. An assistant that reads the open web reads a web that credible publishers are leaving.

4. A patchwork drawn by contracts

The reel's explanation, that AI search reads only the part of the web that still lets crawlers in, mixes up two kinds of agent. Training crawlers collect text for building models. Search indexers and live fetchers retrieve pages when someone asks a question. The companies even disagree on whether those live fetches obey robots files. OpenAI says that for its browsing agent, "robots.txt rules may not apply"; Perplexity says its fetcher "generally ignores" them; Google says its user-triggered fetchers "generally ignore" them; Anthropic says its bots honor them (OpenAI; Anthropic). A robots block therefore constrains Claude's browsing more than ChatGPT's, the opposite of the reel's pairing.

Licensing redraws the map company by company. The Wall Street Journal's robots file disallows every agent it does not name, then explicitly allows OpenAI's three, after News Corp licensed its content to OpenAI. Reddit's file blocks every crawler, while Reddit licenses its data to Google and OpenAI through its API. The Associated Press licensed its archive to OpenAI in 2023 and still blocks OpenAI's training crawler. In one vendor study of millions of citations, about 70% of ChatGPT's citations to top news sites came from sites that block its retrieval agents, a result that could reflect licensing, search-index snippets, fetchers that ignore the files, or older crawls (BuzzStream 2026). Since July 2025, Cloudflare, which sits in front of a large share of websites, asks the owner of every new domain whether to allow AI crawlers at all. Enforcement is moving from the sign on the door to the lock.

The bot statistic belongs here too. Imperva, which sells bot protection, measured automated traffic at 51% of the traffic it saw in 2024 and 53% in 2025; Cloudflare's data for web pages agreed for much of 2025. But AI crawlers were about 4% of page requests on Cloudflare that year, and the sources tie publishers' blocking to copyright, payment and the imbalance between what AI crawlers take and the visitors they send back, not to total bot volume. Cloudflare measured one AI company's crawler fetching as many as 500,000 pages for every visitor it referred (Cloudflare 2025).

So the unlocked web is real and plural. A ChatGPT user and a Claude user are reading different slices of it, drawn by contracts, server rules and paywalls.

5. What actually gets cited

The "one in six" comes from a 2026 audit by Mowafak Allaham and Nicholas Diakopoulos at Northwestern (Allaham and Diakopoulos 2026). They asked ChatGPT, Copilot, Gemini and Perplexity 712 real questions about politics, health and the environment and collected 26,266 cited links. They could analyze the text of 19,154 of them. An AI-text detector flagged about 16% as likely AI-generated, which is where one in six comes from. The rate ran from 7.3% for ChatGPT to 27.8% for Copilot. A manual check of 200 flagged pages found no disclosure of AI use on any of them. The authors call 16% a lower bound, because their detector missed known AI-written articles in testing; they did not measure how often it wrongly flags human writing on real web pages. Two other studies with different detectors found 10% and 18%. A detector's flag is an estimate, and three estimates agreeing on a range is stronger than any one of them.

The "44%" comes from an SEO software company. Ahrefs examined ChatGPT's answers to 750 prompts of the form "best X" about software, products and agencies, and found that best-of blog lists made up 43.8% of the cited sources (Ahrefs 2025). For that kind of question it is the relevant figure. For ChatGPT in general, a larger dataset of over a million citations puts listicles at just under a fifth (Peec 2026). Lists in which a company ranks itself first are real and smaller: 3.6% of ChatGPT's citations in one software study (Peec 2026). And the supply side is documented. A 2024 paper showed that adding citations, quotations and statistics to a page can raise its visibility in AI answers by up to 40%, with the lowest-ranked pages gaining most (Aggarwal et al. 2024), and Ahrefs' own study ends by recommending that companies publish self-promotional "best" lists.

The creator's unfamiliar sources fit the data. In one analysis of 680 million citations, ChatGPT's most-cited site, Wikipedia, accounted for 7.8% of its citations, and its ten most-cited sites together for about a sixth (Profound 2025). The rest spreads across a very long tail. On average only 12% of pages cited by AI assistants rank in Google's top ten results for the same question (Ahrefs 2025). A log of shopping questions full of names you have never seen is what the citation data predicts.

6. Post hoc, and the three things it can mean

"The model writes the answer first and then goes and looks for sources." As a description of how ChatGPT search or Claude's web search works, this is not what either company documents. Anthropic's documentation describes Claude deciding to search, reading results, possibly searching again, and then giving "a final response with cited sources" (Anthropic); OpenAI's describes citations for links found in search results.

But the reel is pointing at something real, and it helps to separate three things the phrase can mean. The first is search-first citation, the documented design, in which citations point to pages the model was shown and can still fail to support the sentence they sit beside. The second is write-first citation, a real research design in which a draft is written and sources are attached afterward; one 2025 study found it gave better coverage than citing while writing (Saxena et al. 2025), and a Google patent granted in 2023 describes linking parts of a finished summary to documents that verify them, possibly documents that were not used to write it (Google patent US11769017B1). The third is memory-first citation, in which the model already favors an answer and the citations serve as justification. Researchers have shown citations that are correct without being faithful, pointing to a document that supports a sentence the model did not actually take from it (Wallat et al. 2024), and one marketing firm's analysis of 362,188 responses argues that brand recommendations come from memory, with retrieval acting as "the bibliography, not the brainstorm," a claim the firm itself labels behavioral evidence rather than proven architecture (Seer Interactive 2026).

The reel presents the second or third as how the products work. The documentation describes the first. All three produce the same experience for the person reading the answer: a link that may or may not say what the sentence says.

7. Why the advice survives every correction

The advice does not depend on the mechanism, because the failure rates are measured. In a 2023 evaluation of four generative search engines, only 51.5% of their sentences were fully supported by the sources they cited, and 74.5% of the citations supported the sentence they were attached to (Liu et al. 2023). Those systems have since been replaced, and the problem has not gone away. A 2025 study in Nature Communications found that even GPT-4o with web search left about 30% of individual medical statements unsupported by its own citations (Wu et al. 2025). Twenty-two public broadcasters testing more than 3,000 answers in 2025 found serious sourcing problems in 31% of them (EBU and BBC 2025). And in a test of eight AI search tools asked to identify the source of news excerpts, the tools together got more than 60% of 1,600 queries wrong (Jaźwińska and Chandrasekar 2025).

A citation is a claim about evidence, and a quarter to a half of those claims fail, depending on the system and the subject. The only way to know which half you are holding is to open the link and find the sentence.

8. Receipts

The recent essays in this series were researched with AI assistance and then checked source by source. Each research report marks every claim: [V] for verified against the primary source, [S] for known only through a secondary source, [K] for background knowledge not rechecked. Every number in this essay carries one of those marks somewhere in its research notes. It is the reel's advice, done slowly.

Doing it slowly shows that the problem the reel describes is older than AI. In the research for this essay alone: the Northwestern paper gives "about 16%" in its abstract and "one in five" in its limitations section, without reconciling them. The SEO study behind the 44% has 26,283 links in its title, close enough to the Northwestern paper's 26,266 that the two are easy to cross. The widely quoted figure that Wikipedia makes up 47.9% of ChatGPT's citations is a share of its top ten sites only; its share of all citations is 7.8%. In the research for this series I traced a figure that "fewer than 5% of users ever change a setting" to a single blog post describing an undated, unpublished experiment, and a popular claim that a NASA nap study found "34% better performance" to a report where those numbers describe something else. People widened all of these, long before a model read them.

That is why I think of a citation as a receipt. A receipt does not prove the purchase was good. It proves the transaction happened as described, and only if someone checks it against the goods. The reel's last sentence asks its viewers to become that someone.

9. Against myself

The creator made a ninety-second video for a general audience, and almost everything I have corrected was a compression, not an invention. A reel cannot footnote. On that view, my corrections are pedantic: the reel's message was right, and it got its viewers to do the thing that matters.

The evidence I have leaned on has its own limits. Most of the data about listicles and citation shares comes from companies that sell tools for improving visibility in AI answers, which gives them a reason to make the problem look large and the fix look necessary. Detectors make errors no one has measured on real web pages. The citation-accuracy studies test particular systems at particular moments, and the systems change every few months; a 2023 failure rate says little about the assistant on your phone today.

And checking every link is expensive. If a quarter of citations fail, three quarters do not, and nobody opens every link in every answer. The practical version of the advice is narrower: check the citations your decision depends on.

10. At the desk, and how this could be wrong

For my own work, the conclusion is a set of habits I already use and now have reasons for. Open the source and find the sentence. Prefer the primary document to anyone's summary of it, including an AI's and including mine. Date-stamp every figure, because the systems that produce them change. Mark what you could not check, and say so in the text. And when a number is striking, look for the denominator.

The argument could be wrong in three specific ways. First, an independent audit could find that current AI search tools support nearly all of their cited sentences, which would turn the advice into a habit from an earlier era. Second, AI companies could publish how their citations are produced and how often they are faithful, which would replace inference with documentation. Third, a careful study could show that AI-text detectors substantially over-flag ordinary web writing, which would shrink the "AI-written sources" finding to noise.

Until then, I read the reel as a good piece of AI literacy whose numbers did what numbers do in transit, and whose last line holds. Click the link. See what is there. See if your answer is on that page.

References

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. KDD 2024. https://arxiv.org/abs/2311.09735
• Ahrefs (Allsopp, G.). (2025, December 4). Do self-promotional "best" lists boost ChatGPT visibility? Study of 26,283 source URLs. https://ahrefs.com/blog/best-lists-research/
• Ahrefs (Linehan, L.). (2025, August 11). Only 12% of AI cited URLs rank in Google's top 10 for the original prompt. https://ahrefs.com/blog/ai-search-overlap/
• Allaham, M., & Diakopoulos, N. (2026). Synthetic sources?: Auditing generative search engine citations for evidence of AI-generated sources. AIES 2026. https://arxiv.org/abs/2605.23684
• Anthropic. (2026). Does Anthropic crawl data from the web, and how can site owners block the crawler? https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
• Anthropic. (2026). Web search tool [API documentation]. https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool
• BuzzStream (Nero, V.). (2026). Do news publishers that block AI crawlers get cited less often by AI? https://www.buzzstream.com/blog/news-block-ai-bots-citations/
• Cloudflare (Belson, D.). (2025, December 15). The 2025 Cloudflare Radar year in review. https://blog.cloudflare.com/radar-2025-year-in-review/
• EBU & BBC. (2025, October 22). News integrity in AI assistants. https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste
• Fletcher, R. (2024). How many news websites block AI crawlers? Reuters Institute for the Study of Journalism. https://doi.org/10.60625/risj-xm9g-ws87
• Google. (2023). Generative summaries for search results. US Patent 11,769,017 B1. https://patents.google.com/patent/US11769017B1
• Google Search Central. (2026). Google's common crawlers. https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers
• Jaźwińska, K., & Chandrasekar, A. (2025, March 6). AI search has a citation problem. Columbia Journalism Review / Tow Center. https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
• Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. Findings of EMNLP 2023. https://arxiv.org/abs/2304.09848
• Longpre, S., Mahari, R., Lee, A., et al. (2024). Consent in crisis: The rapid decline of the AI data commons. https://arxiv.org/abs/2407.14933
• OpenAI. (2026). Overview of OpenAI crawlers. https://developers.openai.com/api/docs/bots
• Peec AI (Wells, T.). (2026). Content types cited most by engine. https://docs.peec.ai/research/content-types-cited-most-by-engine
• Peec AI (Wells, T.). (2026). Self-promotional listicles. https://docs.peec.ai/research/self-promotional-listicles
• Profound (Lafferty, N.). (2025). AI platform citation patterns. https://www.tryprofound.com/blog/ai-platform-citation-patterns
• Saxena, Y., Bommireddy, R., Padia, A., & Gaur, M. (2025). Generation-time vs. post-hoc citation: A holistic evaluation of LLM attribution. https://arxiv.org/abs/2509.21557
• Seer Interactive (Lovett, J.). (2026, March 24). LLM ghost citations: Why your content is working and your brand isn't. https://www.seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt
• Steinacker-Olsztyn, N., Gosain, D., & Dao, H. (2026). Is misinformation more open? A study of robots.txt gatekeeping on the web. WWW '26. https://arxiv.org/abs/2510.10315
• Wallat, J., Heuss, M., de Rijke, M., & Anand, A. (2024). Correctness is not faithfulness in RAG attributions. https://arxiv.org/abs/2412.18004
• Wu, K., Wu, E., Wei, K., et al. (2025). An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications, 16, 3615. https://doi.org/10.1038/s41467-025-58551-6

Source note: sources were checked on 6 November 2026. Consumer Reports' robots file was read live and through Internet Archive copies; the server test was a single-location measurement that changed only the user-agent name. The Steinacker-Olsztyn, Allaham and Diakopoulos, Liu, Wu, Aggarwal and Fletcher papers were read in full or in their full HTML versions; Wallat et al. at abstract level. Company documentation (OpenAI, Anthropic, Google, Perplexity) was read on the companies' own pages, except where noted in the research notes. Ahrefs, Peec, Profound, BuzzStream, Seer, Imperva and Cloudflare figures are vendor data, not peer reviewed. Quotations from the reel come from auto-generated captions.