215,128 Fake 'Best Software' Pages Are Feeding AI Answers. That Gap Is a New Service: Auditing and Fixing Who Your Category's AI Recommendations Actually Cite.

A report went looking for where AI answer engines actually get their "best software" recommendations from.

The answer is not reassuring for most software companies.

Across 380 software categories and 1,807 distinct products tracked, the report found 7,534 total citations spread across 2,055 distinct domains. Of those sources, 59.8% rank worse than #100,000 on Tranco, and 23.4% fall outside the top million websites entirely.

That means most of what an AI answer engine cites as a "source" for a best-of recommendation is not a high-authority review site. It's something far smaller.

The three sites doing this at scale

Three domains — worldmetrics.org, gitnux.org, and wifitalents.com — were all registered "between December 2023 and May 2024." Between them they built 215,128 machine-generated best-software pages:

• worldmetrics.org: 103,578 total URLs, 70,731 of them best-software pages
• gitnux.org: 107,083 total URLs, 71,684 of them best-software pages
• wifitalents.com: 105,541 total URLs, 72,713 of them best-software pages

Both worldmetrics.org and gitnux.org use HTML page titles that self-describe as a "Facts & Grounding Page" — language aimed squarely at how AI retrieval systems select and trust sources.

The report tested this against perplexity/sonar and perplexity/sonar-pro through OpenRouter, and "both models report the URLs they retrieved" — meaning the citation behavior is directly observable, not inferred.

Who's actually winning citations

Legitimate high-authority sources still show up — g2.com had 291 citations (3.86%), reddit.com had 261 (3.46%), gartner.com had 158 (2.10%). One source, guideflow.com, had 194 citations (2.57%) and ranked #3 overall. But those numbers sit inside a pool where the majority of citations go to low-authority sources most brands have never heard of, let alone tracked.

The service this creates

Every software company that cares whether it gets recommended by AI now has a blind spot: they don't know which domains are shaping their category's AI answers, or whether those domains are legitimate or mass-produced.

That's an audit, and audits are sellable:

1. Run the client's category through the same kind of test the report ran — ask AI answer engines "best [category] software" and record every cited domain.

2. Rank each cited domain by authority so the client can see the split between real sources (G2, Gartner, Reddit threads) and manufactured ones.

3. Show the client exactly which manufactured pages are shaping their category's narrative, and whether their own product appears at all.

4. Sell the fix: structured, fact-dense comparison content built to be citation-worthy, plus outreach to get listed correctly on the legitimate high-citation sources.

5. Re-run the audit on a cadence, because AI retrieval and trust signals shift — this is exactly the kind of finding that stops being true if nobody rechecks it.

Bottom line

The report didn't just expose a manipulation tactic. It handed away a measurement methodology. Anyone can run the same test for a client's category this week and show them, in hard numbers, who is currently winning the AI recommendation game in their market — and it's often not who they'd guess.

Sources:

https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/