Perplexity cited 215,128 generated software-buying pages in tests across 380 categories
- Tests of Perplexity Sonar and Sonar Pro across 380 software-buying categories produced 7,534 citations, and 59.8% pointed to domains ranked below 100,000 in Tranco while 23.4% were outside its top million.
- Three apparently linked domains, wifitalents.com, worldmetrics.org, and gitnux.org, list 215,128 generated
/best/<category>-software/pages in their sitemaps; none existed before December 2023. - Guideflow, a vendor of interactive product demos rather than a software-review publisher, was the third-most cited domain with 194 citations across 96 categories, ahead of Gartner.
- Of 2,055 cited domains, 751 domains were absent from Tranco's top million; their median first Wayback capture was 2020 versus 2011 for ranked cited domains.
- The report tested only Perplexity through OpenRouter, explicitly excluding Google and making no claims about other search or retrieval engines.
Hacker News opinions
I worry that LLM search will train on AI-generated pages until it amplifies its own errors. Maybe pre-2022 human-written material becomes the trusted baseline, unless newer material can be certified as human-made.
I think we may already be at that point. I've gone back to library books, especially technical books published before 2020, because physical publishing is harder to flood with slop.
Product search is basically impossible now. If I do not already know what I want, I am screwed.
Manipulating whatever models train on so they recommend your product is surely already a big business.
I've seen API key requests and sales jump about 5x over the summer despite doing almost no conventional marketing. I suspect AI search tools are recommending my services.
I found the subject interesting, but the report reads like a Claude artifact. Searching for its founder produced nothing, and it is ironic to dismiss low-ranked domains while trellner.com itself is outside Tranco's top million.
I used a free 12-month Perplexity offer. It was initially useful for locating a specific recipe, manual, or help page, but it increasingly returns links that do not match the adjacent text and seems optimized for speed over result quality.
I paid for Perplexity Pro for three years and eventually left after paying roughly $400 to $500. The service reduced paid features mid-subscription and kept pushing the gated Computer option at me.
Google's AI Mode has been surprisingly capable and fast in my use, although its hallucination rate also seems higher. That leaves Perplexity with little to differentiate it.
My friend works at Profound, whose job is to get clients found by AI by spamming Reddit and other discussion sites. Traditional SEO firms were already doing versions of this before LLMs.
Profound has raised $155M and was reportedly valued at $1B, with investors including Lightspeed, Sequoia, Kleiner Perkins, Khosla, and NVIDIA. I hear it is shifting from AEO/GEO toward calling itself an "AI Marketer," but I doubt this spam strategy is defensible if platforms stop rewarding it.