A newly published data study set out to answer a simple question: when you ask an AI search engine for the “best” software in a category, where is it actually getting that answer from? The findings are a useful reality check for anyone treating AI-grounded answers as reliable buying research — or building a marketing strategy around getting cited by one.
The setup
The researchers queried Perplexity’s two grounded models (Sonar and Sonar Pro) across 380 software categories — everything from CRM to museum collection management — and logged every citation the models retrieved to support their recommendations. That produced 7,534 citations across 2,055 distinct domains, which the study then cross-referenced against site popularity rankings (Tranco) and historical web archives (Wayback Machine).
The headline finding
Nearly 60% of citations pointed to domains ranked worse than #100,000 globally, and almost a quarter didn’t appear in the top million sites on the internet at all. The median citation sat around rank 71,000 — far from the household names most people would assume are shaping a purchase recommendation. For comparison, Wikipedia was cited three times in the entire dataset.
A vendor’s own blog turned out to be a top source
One standout: Guideflow, a company that sells interactive product demos and doesn’t compete in any of the 380 categories tested, still had its marketing blog cited 194 times — more than Gartner. It showed up as the grounding source for completely unrelated categories like 3D rendering software and architecture practice software. The report is careful to note nothing deceptive is happening here; it’s an ordinary company blog. The interesting part is simply what the retrieval layer did with it.
Sites built for machines, not people
The report’s most striking discovery is a cluster of three domains, all registered within months of each other in late 2023/early 2024, sharing infrastructure and an identical page template, that together published 215,128 auto-generated “best software” pages — despite there not being anywhere near that many actual software categories.
Two of the three sites literally title their homepage “Facts & Grounding Page,” and describe themselves in their meta descriptions as a “machine-readable record” of company and compliance details — language clearly aimed at retrieval systems, not human shoppers. One of the sites also sells “custom market research” and “vendor selection” reports starting in the thousands of euros, positioned directly above the same auto-generated rankings the AI is citing as evidence.
When the researchers pulled the same category from all three sites, the rankings barely agreed with each other, each cited different fictional-sounding “expert” staff, and all three carried the same unrendered template bug in their byline — a strong signal of machine-assembled pages dressed up as editorial content.
When “best” points somewhere it really shouldn’t
Of the 1,502 vendor homepages surfaced across the study, a couple of recommendations resolved to genuinely alarming destinations: one model’s pick for a “research data management platform” redirected to an online gambling site, and one pick for “data quality tools” redirected to a Monaco casino group’s website. Neither was flagged before being served up as a confident answer.
What the study is careful not to claim
To its credit, the report is explicit about its limits. It only measured Perplexity — not ChatGPT, Gemini, Copilot, or Google’s AI Mode — and the two Perplexity tiers tested share most of their retrieval layer, so this is really one search stack sampled twice rather than two independent checks. It also didn’t test whether removing these sources would actually change the final recommendations; it only measured what the evidence base is made of, not whether the answers themselves are wrong.
Why it’s worth paying attention to
As AI-grounded search becomes a default research tool for software buyers — and a growing target for marketers chasing “AI visibility” — this study is early evidence that the citation layer behind these answers is already being gamed at scale, by content built explicitly to be machine-readable rather than human-useful. Worth watching as more buyers start trusting a well-formatted AI answer over their own research.
The full dataset and methodology behind this report are published under CC BY 4.0 for anyone who wants to verify the findings.
Hashtags
#AIGrounding #GenerativeAI #AEO #AnswerEngineOptimization #AISearch #Perplexity #ContentMarketing #SEO #B2BSoftware #MarTech #TrustAndSafety #AIResearch #DigitalMarketing #DataIntegrity
Tags / Topics
AI Search & Grounding · Answer Engine Optimization (AEO) · Perplexity AI · Content Farms & Programmatic SEO · B2B Software Buying · AI Trust & Safety · Marketing Ethics · Search & Retrieval Systems
Suggested Links to Reference This Week
- Tranco List (tranco-list.eu) — the domain-ranking methodology referenced in the study
- Wayback Machine / Internet Archive (archive.org/web) — used to date when the flagged domains first appeared
- Perplexity AI’s documentation on Sonar/Sonar Pro — for context on how citation/grounding works
- OpenRouter (openrouter.ai) — the API layer the study used to query the models
- Background reading on programmatic SEO and AI content farms, for readers unfamiliar with the tactic
- Recent industry writing on Answer Engine Optimization (AEO) / Generative Engine Optimization (GEO) — the discipline this study is effectively a case study in