Perplexity quoted 210,000 pages of AI soft article: How SEO farms feed search results to AI

Perplexity quoted 210,000 pages of AI soft article: How SEO farms feed search results to AI

Three domain names and 215,000 pages are all titled "Best XX Software in 2024." Perplexity uses them as authoritative quotes and continues to feed them into the answers. This is not an evolution of SEO farms, it is Perplexity's citation mechanism that proactively rewards structured nonsense.

I accidentally bumped into a competing product when I searched for "best project management software 2024" at the end of last year: Three of the top five sources pointed to the same unknown domain name. Click in, and the template is exactly the same-the H2 label is neatly stuck, the list of advantages and disadvantages never exceeds seven items, and even the wording of the "Verdict" paragraph is generated like the same prompt. This is not even on the second page of Google, but Perplexity quotes it as a treasure.

Why? Because Perplexity requires a structure where can be referenced . It doesn't read the content, it grasps the skeleton.

Trap of ### reference mechanism

Google's PageRank is a link, and Perplexity is "extractability." In the RAG retrieval pipeline, the granularity of the document segmented by the vector database determines the recall rate. Pages with clear titles, independent paragraphs, and standard list formats have the highest semantic integrity of segmented chunks and naturally have the highest search scores.

These three sites-variants of softwareworld.xyz, saasworthy.com, and crobox.com-have understood the rules thoroughly. They don't write reviews, they write structure . Each page is about 1200 words. H2 covers five major sections: "Pricing","Features","Pros & Cons","Alternatives" and "Final Verdict". The Product/Review tags of Schema.org are all filled. Once the crawler grasps it, it moves everything, and it is clean and high-quality chunk.

Google will also watch users stay, jump out, and click. Perplexity doesn't use it-it just wants to "whether this paragraph can be inserted directly into the answer." Users don't click on the source link at all, but the answer has already extracted "Pros: Easy to use, Great support". The click rate is close to zero, but the content farm does not have any negative feedback signals.

Real Data Flow

The Originality.ai team ran a round of crawling in November last year: Among the top 1000 high-frequency quoted domain names in Perplexity, these three sites together accounted for 4.3%. Based on their estimated average daily query volume, 12,000 citations point to these pages every day for business intent words such as "best software" alone.

What's even more ridiculous is the chain of quotes. Perplexity's answer quotes softwareworld, and the external link in the soft article points to the SaaS official website, which is then quoted twice by Perplexity as an "official source." A closed-loop run-through: garbage pages are washed into authority, and authority is fed back to the model.

I asked a friend who works as a RAG pipeline to verify it: throw the pages of these three stations into the vector library and use "best crm for small business" to check, there will be 3 seats in the Top-5 recall. Changing to the real evaluation pages of Wirectter, G2, and Capterra, the recall ranking dropped beyond 20th due to long paragraphs, mixed advertising code, and incomplete schemas.

Structured nonsense defeats unstructured knowledge.

opponents will say: Citation is better than illusion

Some people will argue: Perplexity at least gives the source, so users can click on it to verify. It's better than ChatGPT compiling non-existent software.

This sounds right and cannot withstand scrutiny.

First, users do not click . Perplexity admitted in its own financial report that the click-through rate of city-to-link was less than 2%. The answer has already read "Verdict: Best overall" to you, and you even click on the hammer link.

Second, the source page itself is a circular argument . Softwareworld's "data source" says "aggregated from user reviews on G2, Capterra," but you can't find corresponding comments by clicking on G2. They just grab the G2 scores and repackage them into a list. Perplexity refers to softwareworld, which is equivalent to referring to the "shadow of G2"-after two layers of transfer loss, the number has long been deformed.

Third, training data pollution is the biggest problem. Perplexity's answers will be included, crawled, and eaten in the next round of model training. 21.5 Thousands of pages of structured nonsense, each page being quoted dozens of times, is equivalent to feeding the next generation model with millions of training samples that "look like reviews, but are actually templates." What the model learned was not "how to evaluate software", but "how to evaluate software."

I know an architect who works RAG pipelines

Around November 11 last year, his team took orders to do "intelligent selection assistant" for cross-border e-commerce. Customers need to stuff Amazon's top 100,000 Listing reviews, QA, and competitor comparisons into the vector library, so that the big model recommends "worthy selections."

He told me: At first they used an open-source semantic chunker to cut by paragraph. The results were all nonsense such as "good quality" and "fast delivery"_because Amazon's comment area had the cleanest HTML structure, <span data-hook="review-body"> cut out the chunk semantics perfectly, and the vector scored the highest. The truly valuable "competitive product comparison table" and "keyword frequency analysis" were shredded due to chaotic table labels and JS rendering.

In the end, they had to write a preprocessing pipeline: reduce the weight of the comment area to 0.1, and forcibly increase the weight of tables, lists, and blocks with "vs" and "comparison" in titles to 2.0. Customer satisfaction rose from 3.2 to 4.6 after launch.

"Essentially the same," he sent me on WeChat."The retriever doesn't understand the business, it only believes in the structure. Whoever has a good structure is the truth. "

Perplexity is now the blockbuster without business weights-just that it faces the entire network and the training set of the next generation model.

tool reshapes users, this time it's the search engine's turn

Ten years ago, SEO forced web pages to look like Google liked them: keyword density, anchor text, and external chain farms. Now RAG forces web pages to grow into the way vector databases like them: semantically independent, with perfect schemas, and list-extractable.

The difference is: In the era of Google, spam pages had to at least be clicked by people. In the era of AI search, spam pages only need to fool the splitter. Humanity has exited the feedback loop.

This is a bit of a joke: we built AI search to escape the SEO garbage dump, and we ended up creating a monster that eats even more structured garbage.

If Perplexity does not change the weight-such as introducing an authoritative source white list with manual annotations, discounting "extractability", and feeding user click/pause/correction signals back to search sorting-these 215,000 pages are just an appetizer. The next wave is 2.15 million pages of "Best AI Programming Assistant","Best Vector Database" and "Best RAG Framework", all generated by the same set of templates.

By then, you won't even be able to find a real evaluation in the model training set.

No one washed the data source. Perplexity has no motivation-more quotes make answers rich, and financing makes it easier to tell stories. Content farms are even less motivated-the cost of running the template is 1% of the real review. The model training party also has no motivation-they have already mixed in Common Crawl, and the cleaning cost is higher than the training cost.

I'm not sure how this will end. But I'll take a gamble: The first AI search to explicitly introduce a "human verification signal" in RAG sorting will defeat Perplexity.

Just like Google used PageRank to kill a portal website built up by keywords. The next generation of winners will not be the cleverest splitters, but the ones who know the most about not believing in structured nonsense.

Tools reshape users. AI search upgrades SEO spam problems to cancer in the training set-this is what really deserves to lose sleep.