LLM Writing Is Cheap, Verification Is Expensive: A Content Site's Verification Practice

Having a model write a page of content is the easiest step. Give it a topic and some sources, and it produces smooth-reading text. If that were the whole job, a reference site could be built in a weekend.

The developer made a Spanish political test site this year and built a set of reference pages around it: an introduction to each thinker, a glossary, and articles comparing parties pairwise. Most of the text was drafted by AI agents. Drafting was fast; what actually took time was verification. Here are the lessons he summarized.

Models cite "from memory"

If you ask a model to read a page of content and then cite it, the model will first read the page, then write quotations from memory. The result looks correct, but often is not the original text, and is sometimes even translated.

The author ran several measurements. When asking a small model to do verbatim quotations, only 24% to 35% of quotations passed source-text verification. In an early glossary, 141 of 285 quotations were not in their claimed sources at all. One entry translated all quotations into Spanish, while its sources were in English, French, and Russian, so nine out of ten quotations failed.

This means that a large amount of LLM-generated content on the internet is completely wrong.

The "from memory" citation referred to here means that Transformer-type models predict word by word according to token probabilities when generating text, rather than retrieving the original text precisely like a database. They have seen massive amounts of text during training and can reproduce passages highly similar to the original, but specific wording is "smoothed" by the statistical preferences of the internal language model, producing unconscious substitutions or translations. Therefore their output is often semantically plausible but literally unreliable — this is also why such systems are a typical case of "high confidence but low accuracy" on citation issues: the model itself does not know which parts are original and which are reconstructions. The solution is to separate retrieval from generation, letting the model responsible for generation handle only the file contents it has already obtained, and leaving verbatim quotations to retrieval or string-matching scripts.

Download the sources first, then let the model write

The solution is to take "citation" out of the model's hands:

1. Before any model sees the source, use curl to download the source file.
2. Check that each file really is what it claims to be. The author once saved a Finnish poetry collection as Kant's Enlightenment essay, and another time three PDFs were all the same 25 KB error page.
3. Give the files to the writer and tell it not to quote anything not in the files. For research claims, the model does not quote at all: it returns the claim, URL, and a few keywords, and then a script picks out the sentence from the downloaded HTML.

As a result, 36 of 39 research quotations were verbatim correct, and 95% of quotations in thinker biographies passed source verification.

Before trusting a verifier, test it first with quotations you have confirmed are correct.

When quotations do not support the claim

Verbatim quotations can only prove that the text is in the source; they cannot prove that it supports the sentence next to it.

The claim "PP supported Sánchez's investiture in 2023" is false, yet it was paired with a perfect verbatim quotation — a sentence about Sánchez being appointed. So a second check is needed: a model acting as judge reads each claim and quotation and determines whether the quotation supports the claim. Partial matches can only be kept after the judge rewrites the claim into a statement that the quotation actually supports.

Later, the author used an agent to run a review pass on the 26 highest-risk pages. It caught problems the scripts could not: a page attributed a court case and a conviction to a living person without a source; a quotation about Putin actually came from a source about Kennedy; and an argument attributed to Hayek was actually Mises's. The reviewer also over-reported, so every finding was sent to a second agent to rebut it, and 45% of findings were overturned.

Whether "a quotation supports a claim" is usually called an entailment judgment in information retrieval and fact-checking: it is not only about whether the sentence appears in the source, but also whether the source semantically guarantees the claim is true. Using a model as judge here is essentially having an LLM perform a text entailment classification task: input a premise (the quotation) and a hypothesis (the claim), and output supported, partially supported, or irrelevant. This kind of judgment is much harder than verbatim citation, because it requires understanding nuances such as negation, time, conditions, and subject reference, and the model easily mistakes "mentions the same person or event" for "supports the claim." Introducing a second agent to rebut is meant to reduce the confirmation bias of a single judge — two models challenging each other can filter out a considerable proportion of surface matches.

Wording and context

Some edits have nothing to do with facts. In a sentence about a named journalist, "fabricated" was changed to "started"; "the best ammunition against something" was changed to "standard reference."

No check can flag these kinds of problems. A human has to read them.

What to prepare before starting a content site

Budget for verification, because writing is now the cheap part. Do not let the model produce content you cannot verify; give it files rather than URLs, and let code extract quotations.

The author also recommends testing a verifier before trusting it, and keeping a threshold.

In many cases, if an entry verifies fewer than four quotations, it is better to choose a shorter handwritten version.

Search engines are increasingly able to distinguish thin generated content, so this also affects rankings — if a site is basically low-quality AI-generated content, Google may flag it.

The author says that next time he builds a content site, the verifier will be the first thing he writes. In fact, most of the site's content has undergone extensive review, or is simply handwritten.

Search engines have developed a relatively mature quality-assessment approach to AI-generated content. Google's search quality guidelines emphasize E-E-A-T, namely experience, expertise, authoritativeness, and trustworthiness, and explicitly treat large-scale automatically generated content lacking original value as a low-quality signal. Ranking systems do not directly penalize text simply because it was "written by AI," but they do identify whether the content provides unique information, is verifiable, and is maintained by an entity with real experience. Therefore, a content site centered on reference pages can still achieve good positions in search results if its quotations are traceable, sources are clear, and signs of human review are evident; conversely, if the site is mainly generated text with a low verification rate, the risk of being demoted rises significantly.

One-line Summary

What is truly expensive about an AI-drafted content site is not writing, but verifying quotations that models fabricate from memory.

Paragraph Summary

Using a Spanish political test site as a practice case, the author explains that having a model write a page of content is fast, but verification is extremely slow. Models cite "from memory," with small models passing verbatim citation checks only 24% to 35% of the time; in an early glossary, 141 of 285 quotations were not in their sources. The solution is to take citation out of the model's hands: first use curl to download sources, check file authenticity, then have the model write based on files, with research claims extracted from HTML by scripts. Even so, verbatim quotations do not equal support for claims, so a model judge is still needed to assess entailment, with a second agent rebutting to reduce false positives. Wording and context issues still require human review. The author recommends budgeting enough for verification, and preferring a short handwritten version when fewer than four quotations can be verified.