AI generated content sites: how do buyers tell a content farm from a real publisher?
Part of MFA avoidance
The short answer
AI generated content sites publish large volumes of articles written mainly by language models, with little reporting, review or original value, and earn money by filling those pages with programmatic ads. Buyers tell them apart from real publishers by checking bylines, sourcing, publishing pace, repeated article structures and ad layout on sampled pages, not by guessing how text was written.
Key takeaways
- An AI content farm is defined by its business model: cheap generated pages, bought or search traffic and dense ads, not by the use of AI tools alone.
- The ANA's Q1 2026 benchmark release identified AI slop as an emerging MFA subtype requiring ongoing mitigation, a finding about its benchmark participants.
- Google separates helpful AI assisted work from generating many pages without added value, and bars its ads on generated content without manual review or curation.
- Detector scores and disclosures are context, never proof; record what the page shows and what a claim's source actually says.
What are AI generated content sites?
An AI generated content site is a website whose articles are produced mostly by a language model and published at a pace no editorial team could review. The pages look like a magazine: headlines, categories, an author name, a date. Missing is the work behind them. Nobody interviewed anyone, tested a product or checked a quoted figure. Revenue comes from the ads around that text, which fits the pattern of made-for-advertising inventory.
Definition
- AI content farm
- A website or network of sites that uses generative models to publish large numbers of pages with little human review or original value, mainly to collect search or paid traffic and monetize it with ads.
The pattern is measurable in research. Stanford researchers Hans Hanley and Zakir Durumeric trained a synthetic news detector and classified more than 15.46 million articles from 3,074 misinformation and mainstream news websites. In their ICWSM 2024 paper (opens in a new tab) they found that between January 2022 and May 2023 the relative number of synthetic articles rose 57.3 percent on mainstream sites and 474 percent on misinformation sites, and that the increase was largely driven by smaller, less popular websites. Their analysis of ChatGPT's release found a marked increase on small sites and misinformation sites, but not a corresponding increase on large mainstream news websites.
How did the ANA describe AI slop in its benchmark?
The ANA's Q1 2026 Programmatic Transparency Benchmark release (opens in a new tab), published May 27, 2026 and the latest public edition, reported a rise in MFA exposure among participating marketers, "with the report identifying AI slop as an emerging subtype requiring ongoing mitigation." The benchmark is built from the log-level data of the advertisers that take part, so the finding describes those participants, not the whole market. The exposure figures themselves are set out in our MFA explainer.
Two words carry the practical meaning. Subtype places AI slop inside the MFA definition the trade groups published in September 2023, not in a new category. Ongoing mitigation means the ANA treats it as continuing work, not a problem a one time blocklist settles.
How does an AI content farm turn generated pages into ad inventory?
The business model is the MFA model with paid writing replaced by model output. A content farm pays for model usage and hosting, so it can publish large numbers of pages aimed at search phrases or trends and monetize the ones that draw visits through exchanges.
Google's spam policies (opens in a new tab) name this behavior. Scaled content abuse is "when many pages are generated for the primary purpose of manipulating search rankings and not helping users," and the first example is "using generative AI tools or other similar tools to generate many pages without adding value for users." These are search rules, but they describe pages that later appear in a bid stream.
How can a buyer spot an AI content farm?
Avoid trying to prove authorship. A reviewer can prove how the site behaves: who stands behind the articles, whether claims trace to real sources, how pages are built and how fast they appear. Check the signs below on five or more recent articles from different sections.
| Sign | What a reviewer sees | How to check it |
|---|---|---|
| Hollow bylines | Author pages with no history, a generic headshot, or one name credited across unrelated beats | Open the author page and search the name with the site domain; Google lists fabricated creator profiles as deception |
| Unsupported claims | Statistics, quotes or product specs with no source, or a cited source that does not contain them | Follow one consequential claim to its source and read the passage it depends on |
| Repeated skeleton | The same headings, intro formula and conclusion across articles with only the subject swapped | Lay three articles side by side and compare the heading order |
| Publishing bursts | Dozens of articles on one day, dates clustered, no corrections or updates | Read the XML sitemap or RSS feed and count posts per day |
| Ad heavy layout | Ads between most paragraphs, sticky units and recommendation boxes below the text | Load the page on a phone and note how much of each screen is advertising |
A 20 minute content farm check for one candidate site
- Pick five recent articles from different sections. Skip the homepage, which can be more polished than the articles.
- Open each author page. Note whether the person has a history, a role and contact details, and whether the same name covers unrelated topics.
- Trace one claim per article. Choose a figure or a quote that matters and open the source; record whether it says what the article says.
- Compare the structure. Put two articles next to each other and look for the same headings in the same order with the topic swapped.
- Read the sitemap or feed. Count how many pages were published in the last week and on which days.
- Record the ad layout on a phone. Screenshot the first three screens of one article with the time and device noted.
Is all AI assisted content low quality?
No. Google's guidance on using generative AI content (opens in a new tab) asks publishers to make sure AI assisted work meets the standards of its Search Essentials, to "manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing," and to consider explaining how content was created. Its guide to helpful, people-first content (opens in a new tab) says readers benefit from knowing how a page was produced, including AI assisted content, and that the reason for creating it should be to help people.
Google's publisher policies (opens in a new tab) draw the line for ads in similar terms: "Don't place ads on automatically generated content without manual review or curation." A separate policy on replicated content (opens in a new tab) bars ads beside material copied or modified slightly, for example by substituting synonyms, without added value. The issue is the missing review and the missing value, not the tool.
| Practice | AI assisted publisher | AI content farm |
|---|---|---|
| Who answers for the page | A named editor or writer with a visible history | A stock name or no one |
| Use of the model | Drafting, summarizing, translation, data formatting, with review | Producing the finished article |
| Claims | Traceable to reporting, documents or tests | Plausible but unsourced, sometimes wrong |
| Limits of the signal | A disclosure does not prove every claim was checked | A clean layout does not mean the text was reviewed |
The checks matter because of how models fail. NIST's Generative AI Profile (opens in a new tab) (NIST AI 600-1, July 2024) lists confabulation as a core risk: "the production of confidently stated but erroneous or false content." An unreviewed page carries that risk on every claim, a suitability concern when the error is about health, money or safety.
Why do AI content sites attract programmatic spend?
Standard campaign metrics do not flag them. The trade groups' 2023 definition, published by the ANA (opens in a new tab), notes that MFA websites exhibit high measurability rates, good viewability rates, low levels of invalid traffic and usually brand-safe environments. Generated articles about recipes or gadgets contain nothing a keyword blocklist is built to catch.
Three buying habits widen the opening. Broad category targeting accepts any page tagged with the right topic. Optimizing toward cheap viewable impressions rewards dense generated pages. And exclusion lists lag: a farm can register a new domain, copy its template and sell the same week. An inclusion list built from reviewed sites, as described on our MFA avoidance hub, closes that gap because an unreviewed domain never runs.
Buying settings that reduce exposure to AI content farms
- Open exchange display and native run on an inclusion list, not on category targeting alone.
- New domains reach the list only after a reviewer has sampled several articles.
- The served domain report is compared with the approved list in the first week of every flight.
- Each exclusion records the observed evidence, not a label such as AI site.
What does an AI content review look like on a real brief?
Fictional example
A Texas running shoe brief
A fictional Austin running shoe retailer asks its agency for an English display plan in Texas, beside marathon training plans, shoe fitting advice and race coverage. The first candidate set mixes running club blogs, regional race calendars and a dozen shoe review sites found through search.
Four review sites share one pattern: a best shoes article for every foot type with the same seven headings, one author with no history on every article, and cushioning measurements missing from the manufacturer specifications they link to. Their sitemaps show more than 40 posts on one day. The planner saves screenshots and leaves all four off the list.
A fifth site says in its editorial note that it uses AI to draft race results tables from official timing files, and its training articles carry a coach's byline with a long archive. The planner checks two tables against the official files, finds them accurate, and keeps the site with a note on how its AI use was verified.
How does PlacementGraph check thin and templated pages before they reach a plan?
Frequently asked questions
Does an AI detector score prove a page was machine written?
No. A detector is a classifier that estimates the likelihood a passage was generated; it does not record how the page was made. Treat a score as one clue that sends a site to review, then decide on what the pages show: unsupported claims, hollow bylines, repeated structures and publishing pace.
Should a site that discloses AI use be excluded?
Not for the disclosure. Google encourages publishers to explain how automation was used where readers would ask. A disclosure tells a reviewer where to look. Check whether the AI assisted pages were reviewed, whether their claims trace to sources and whether a named person answers for them.
Can an AI content farm also sit on an older, established domain?
Yes. A domain's age or past reputation says nothing about who publishes on it today. Generated sections can appear on a long running site after a sale, so review recent articles and current author pages rather than the domain's history.
Is a site with accurate AI generated articles still a risk?
It can be if the business model is arbitrage: bought visits, dense refreshing ads and pages written for search phrases rather than readers. Accuracy is one test. The MFA definition also looks at ad load, traffic sourcing and generic content, so a correct but empty page can still fail.
Sources
- ANA Programmatic Transparency Benchmark Reveals Widening Gap Between Top and Bottom Advertisers (Q1 2026) (opens in a new tab)ANA, May 27, 2026
- Leading Trade Groups Define “Made For Advertising” Websites (opens in a new tab)ANA (Association of National Advertisers), September 26, 2023
- Machine-Made Media: Monitoring the Mobilization of Machine-Generated Articles on Misinformation and Mainstream News Websites (opens in a new tab)Hans W. A. Hanley and Zakir Durumeric, Stanford University, ICWSM 2024 (arXiv 2305.09820 v5)
- Spam policies for Google web search: scaled content abuse (opens in a new tab)Google Search Central, last updated August 28, 2026
- Using generative AI content on your website (opens in a new tab)Google Search Central, last updated October 1, 2026
- Creating helpful, reliable, people-first content (opens in a new tab)Google Search Central
- Google-served ads on screens without publisher content (opens in a new tab)Google Publisher Policies Help
- Google-served ads on screens with replicated content (opens in a new tab)Google Publisher Policies Help
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) (opens in a new tab)National Institute of Standards and Technology, July 2024
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.