Most 'best software' answers from AI search trace to spam farms, a new study found. Verify before your next construction tech buy
Trellner Research ran 380 software queries through Perplexity and found 59.8% of citations pointed to low-authority sites, including three domains that published 215,128 machine-generated 'best of' pages built to be read by AI models, not people. That's the check to run before trusting an AI's pick on your next construction software buy.
A researcher ran 380 "best software" queries through Perplexity and found that 59.8% of the resulting citations pointed to websites ranked worse than the top 100,000 globally — and that three domains alone had published 215,128 machine-generated pages designed to be read by AI models, not people. If your project engineer, ops director, or estimator has ever typed "best submittal management software" or "top construction ERP for a mid-size GC" into an AI search tool, this is the reliability problem sitting underneath that answer.
What did the study actually find?
Trellner Research published the report in September 2026. The methodology: 380 software-category queries fed to Perplexity's sonar and sonar-pro models, with every resulting citation logged and checked against the Tranco list, a standard measure of domain popularity. Of the 7,534 citations collected, 59.8% went to sites ranked below #100,000 worldwide, and 23.4% went to domains that don't appear in the top million at all.
Three domains did outsized damage: wifitalents.com, worldmetrics.org, and gitnux.org. None existed before December 2023. All three were registered through the same registrar within a five-month window, share identical page templates, and collectively published 215,128 "best of" category pages. The researcher found fabricated-looking bylines — nine different "staff writer" names credited across the three sites for a single query — and unrendered template placeholders still visible on two of the three. Two of the sites titled their own homepage "Facts & Grounding Page" in the HTML, which reads less like a mistake and more like a label written for a model to find.
Why does this matter for a GC or sub buying software?
Software and platform recommendations were the single largest category in the study — bigger than consumer products or professional services. That's the exact category a construction company hits when it asks an AI tool to shortlist project management, submittal, estimating, or field-reporting software. If nearly 60% of the citations backing that shortlist trace to low-authority or purpose-built spam sites rather than real user experience or independent testing, the "top 5 construction PM tools" answer you got back may be marketing copy laundered through a template, not a researched comparison.
This isn't hypothetical for construction. A PM evaluating submittal software, an estimator comparing takeoff tools, or an ops director shortlisting a field-reporting app is a real, common use of AI search today — and it's precisely the query type this study shows is most exposed.
Is this just a Perplexity problem?
No — the study only tested Perplexity, but a separate cross-engine analysis points to the same weak spot elsewhere. Peec AI examined 232,000 citations across major AI search tools between December 2025 and February 2026 and found Google AI Mode (about 10.3%) and Perplexity (about 10.4%) cited self-promotional listicle content — vendor-written "best of" lists — at roughly three times the rate ChatGPT did (about 3.6%). Different metric, same underlying pattern: some AI search engines lean harder on content built to rank rather than to inform, and there's no way to tell from the answer alone which citations are which.
How to verify an AI's software recommendation before you buy
| Signal in the AI's answer | What to do instead |
|---|---|
| A "best of" list with no named reviewer or test methodology | Check the vendor's own site and demo the tool directly |
| Citation from a site you've never heard of | Search the domain by name — spam-farm sites rarely show up in trade press |
| Stats with no attribution ("saves teams 40% of their time") | Ask the AI tool for the source, then check whether that source is a real study |
| No mention of tradeoffs or what the tool doesn't do | Call a firm your size actually running it and ask what breaks |
| Product comparison table that reads templated across categories | Cross-reference against ENR, Construction Dive, or a peer's actual deployment |
None of this means stop using AI search to start a shortlist — it's a reasonable first pass. The mistake is letting that first pass become the purchase decision. Treat an AI-cited "best of" answer the way you'd treat an unvetted sub's self-submitted references: a starting point that still needs a phone call.
For a related reliability gap in how these tools handle your data, see our story on Google's AI Overviews cutting contractors' organic search traffic.
Forward this to whoever's shortlisting your next PM, estimating, or submittal software.
Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.
- Does Perplexity make up its sources when it recommends software?
- Not exactly. Trellner Research found the citations are real, live pages — the problem is that many of those pages were built specifically to be read by AI models, not by people, with templated content and fabricated-looking staff bylines, and Perplexity's system treats them as legitimate sources.
- What did the study actually measure?
- Trellner Research ran 380 software-category queries through Perplexity's sonar and sonar-pro models and logged 7,534 citations. 59.8% went to domains ranked worse than #100,000 globally on the Tranco popularity list, and 23.4% went to domains missing from the top million entirely.
- Which sites were identified as the biggest source of manufactured citations?
- wifitalents.com, worldmetrics.org, and gitnux.org — three domains that appear to be under common control, none of which existed before December 2023, that together published 215,128 machine-generated 'best of' category pages using identical templates.
- Is this only a Perplexity problem?
- No. A separate analysis by Peec AI of 232,000 citations across engines found Google AI Mode (about 10.3%) and Perplexity (about 10.4%) cited self-promotional listicle content at roughly three times the rate of ChatGPT (about 3.6%) between December 2025 and February 2026 — a related but distinct measure of the same underlying reliability problem.
- How should a GC or estimator vet an AI-recommended construction software tool?
- Treat the AI's citation as a lead, not a verdict. Confirm the recommendation independently against the vendor's own site, trade coverage from outlets like ENR or Construction Dive, and a reference call with a firm actually running the tool, before it goes anywhere near a purchase decision.