Data-Generated Is Not Generative: The Distinction That Decides Whether You Get Trusted

There is a sentence we have heard in almost every conversation with a publisher since 2023: “So it’s AI content.” It is a reasonable assumption. Content produced at scale by software has, for most of the web’s history, meant content produced badly. Generative AI has made the association stronger, because it makes producing plausible text nearly free.

But two very different things are being lumped together. One system takes a prompt and predicts the most likely next words. The other takes verified data and renders it through a template that a person wrote and reviewed. They can both produce a million pages. They do not produce the same thing, and the systems that decide what gets read — search engines, answer engines, and readers — are increasingly good at telling the difference. This report lays out the evidence.

1. What Google actually penalizes

In March 2024, Google shipped a core update alongside three new spam policies. The one that matters here is “scaled content abuse.” Google’s earlier policy had targeted automatically generated content specifically; the new one, in Google’s words, targets “producing content at scale to boost search ranking — whether automation, humans or a combination are involved.” Google projected the combined changes would reduce low-quality, unoriginal content in results by 40%, and in April 2024 reported the actual figure at 45%.

Read carefully, that is not a policy against machines. It is a policy against content that exists to occupy a query rather than to answer it, and it explicitly includes content farms staffed by people. Google’s separate guidance on AI-generated content, published in 2023, says the same thing from the other side: automation is not against its guidelines when it produces helpful, original content.

The operative word in both is original. Google’s 2022 patent on information gain describes scoring a page by how much it adds beyond what the searcher has already seen. A page that restates the consensus in fresh phrasing adds nothing measurable. A page that contributes a number computed from primary data adds exactly the thing the score is designed to find.

2. Why generated text fails on facts

The Tow Center for Digital Journalism’s March 2025 test is the cleanest public measurement of how generative systems handle provenance. Researchers gave eight AI search engines excerpts from real articles and asked for the headline, publisher, date, and URL — 200 queries each, 1,600 in total. More than 60% of the answers were wrong. ChatGPT Search was wrong 67% of the time; Grok 3 was wrong 94% of the time and returned error pages for 154 of its 200 citations. The paid tiers of Perplexity and Grok answered more questions correctly than the free tiers — and also produced more confidently wrong answers, because they declined less often.

This is not a bug that a better model fixes. A language model is a system for producing plausible text. When the plausible text and the true fact coincide, it looks like knowledge. When they diverge, the model has no way to notice, because it has no representation of the fact separate from the words. Every hallucinated statistic, every invented citation, every subtly wrong graduation rate is the same failure: the system is optimizing for fluency, and fluency is not accuracy.

A data-generated page has the opposite architecture. The number is a field in a database with a named source and a release date. The sentence around it is a template a writer authored and an editor reviewed. The template cannot invent a value; it can only render the one it was given. If the value is wrong, it is wrong in the data, where it can be found and fixed once for every page that uses it.

3. The engagement evidence

Readers make the same distinction, whether or not they can articulate it. The most direct evidence we have is our own.

Media Factual grew out of a network of reference sites written the traditional way — thousands of articles, many writers, one at a time. When we built College Factual on a knowledge graph and templates, we ran both kinds of site on the same analytics and the same ad stack, which made a comparison possible. In the period we compiled in 2022, the data-driven site averaged 2.44 pages per visit. The five human-written sites averaged between 1.16 and 1.36.

Source: Media Factual internal analytics, comparison compiled in 2022. Sites shared the same analytics and monetization configuration; bounce and time-on-site were excluded because the ad platform measured them differently across sites.

A 79% lift over the best of the hand-written sites is not a claim that templates write better prose than people. It is a claim about what a page built from structured data can do that a narrative article cannot: show the reader the number, show it against its cohort, chart it, and link it to the twenty related entities they will want next. A reader who has just looked up one college’s graduation rate wants the comparison, the trend, and the alternatives. A knowledge graph has all three one click away. An article has a conclusion.

4. The eight differences, side by side

We have kept a running comparison for partner conversations since 2024. It is worth reproducing, because each row is a place where the two approaches diverge in a way that a search engine, an answer engine, or a reader can detect.

	Generative AI content	Data-generated content
Accuracy	Bias, hallucinations, and confident false statements	Rendered from a verified, referenced knowledge graph
Trust	Low; every published error is a brand risk	Every figure traceable to a named source
Insights	Restates what already exists	New calculations, comparisons, and rankings
Freshness	Static once written; updates are manual	Regenerated when the data or the strategy changes
Uniqueness	Recombines what is already indexed	Original analysis that exists nowhere else
Engagement	A single article with generic charts, if any	Personalized visualizations and data products
Editorial control	None without reviewing every article	Review every sentence pattern once; it holds for every page
Architecture	Isolated articles chasing isolated keywords	Deep, interlinked topical clusters across pages and sites

The two columns are not “AI versus human.” Writers and analysts are more central to the second column than the first: they design the templates, decide which comparisons matter, and review every sentence pattern before it scales. What they do not do is hand-type the ten-thousandth variation of a sentence whose only difference is the data.

5. What the citation data adds

The same distinction is now visible in which pages answer engines choose. Across the studies we reviewed in our report on citations: AI-cited pages are 25.7% fresher than organic results on the same queries (Ahrefs, 16.975 million citations); pages with added statistics gained 33% visibility in generative answers and pages with keyword stuffing lost 9% (the GEO benchmark); and government sources — the canonical holders of specific numbers — appear in Google’s AI summaries at three times their share of ordinary results (Pew).

Each of those is a property a data-generated page has by default. Freshness is a regeneration schedule. Statistics are the content. Provenance is a field. A generative page can be given all of them, one article at a time, by a person checking every claim — which is another way of saying it can be turned into a data-generated page by hand.

6. Where generative AI does belong

None of this is an argument against language models, which we use daily. They are excellent at the tasks that are about language rather than facts: drafting a template’s sentence variations for a writer to edit, summarizing a methodology, translating a page, suggesting which comparisons a reader might want, checking prose for clarity. In every one of those cases the model works on text whose facts are already fixed by the data. That is the line: the model can shape how a fact is said. It should never be the source of the fact.

The Reuters Institute’s 2026 survey found only 44% of publishers describing their AI initiatives as showing promising results, with 42% calling the impact limited so far and 13% transformational. Our reading is that most of those initiatives put the model in the wrong seat — asked it to produce content rather than to assist the people who produce it. The results follow.

Source: Reuters Institute, Journalism, Media, and Technology Trends and Predictions 2026; survey of 280 media leaders.

Methodology and caveats

Our engagement comparison is drawn from internal Google Analytics for sites we owned and operated, compiled in 2022 for a partner presentation. Pages per visit is the only metric shown because the sites’ ad-layout testing platform altered how bounce rate and time on site were recorded, making those metrics non-comparable. The comparison is between one data-generated site and five human-written reference sites on different topics; it demonstrates the engagement pattern of data-driven pages, not a controlled experiment on identical content. Site names other than College Factual are withheld.

External sources. Google’s 40% projection and 45% result are from its March 5 and April 26, 2024 announcements. The Tow Center figures are for source identification from excerpts, a specific task; they measure provenance handling, not the accuracy of every answer. GEO benchmark figures are position-adjusted word count in generated answers. Ahrefs freshness figures cover five platforms. Reuters Institute figures are from a non-probability survey of 280 senior leaders.

What to do about it

1. Ask where the facts come from, not whether a machine was involved. The question that predicts trust is whether every number on the page has a source, a date, and a method attached — not whether a person typed it.
2. Separate the fact layer from the language layer. Keep facts in structured data with provenance; let writers (and, where useful, models) work on how facts are expressed. Never let the language layer originate a fact.
3. Review patterns, not pages. If you generate at scale, editorial control means reviewing every sentence pattern and every insight rule before it ships, and treating an error as a data fix, not a page fix.
4. Build for information gain. If a page could be produced by a model with no access to your data, it will be. Publish what only your data can say.
5. Judge scale by the rules Google actually enforces. Scaled content abuse is about value, and it applies to human content farms too. A million pages that each answer a specific question with a sourced number is not what the policy targets; a thousand pages that all say the same thing is.

Sources

1. Google — New ways we're tackling spammy, low-quality content on Search (Tucker, March 5, 2024; updated April 26, 2024) — Scaled content abuse definition; 40% projected, 45% reported reduction
2. Google Search Central — Google Search's guidance about AI-generated content (February 2023)
3. Columbia Journalism Review / Tow Center — AI Search Has a Citation Problem (March 6, 2025) — 8 engines, 1,600 queries
4. Search Engine Journal — Google's Information Gain Patent (analysis of US patent granted 2022)
5. Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024)
6. Ahrefs — Do AI assistants prefer to cite fresh content? (July 28, 2025)
7. Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results (July 22, 2025)
8. Reuters Institute — Journalism, Media, and Technology Trends and Predictions 2026 (January 2026)
9. Media Factual — internal engagement analysis (compiled 2022) — Pages per visit, College Factual vs. five human-written reference sites on a shared analytics and monetization stack