Perplexity just recommended software based on a page that doesn't exist.
Not a page that was outdated. Not a page with a biased review. A page that was manufactured from thin air — auto-generated by a content farm designed specifically to fool AI search tools into treating it as a credible source.
Researchers recently uncovered three websites that collectively produced 215,128 fake "best software" review pages. These weren't hand-written articles with a slant. They were templated, auto-generated at scale, and seeded across the web with one goal: get cited by AI assistants like Perplexity, ChatGPT search, and others that pull from live web sources. It worked. These pages are being cited right now, in answers that real business owners are reading and acting on.
If you're using any AI assistant to research vendors, compare tools, or validate suppliers — and most SMB operators are — there's a real chance the answer you just trusted was built on poisoned data.
How AI Gets Fooled (And Why It Matters to Your Business)
Here's the thing about AI research tools: they're wired to sound confident. That's a feature, not a bug — until the source they're drawing from is garbage.
Most AI assistants that search the web don't read sources the way a human researcher would. They retrieve content that appears authoritative — high word count, structured layout, keyword-rich headings — and synthesize it into an answer. A well-formatted fake page looks identical to a well-formatted real page. The AI doesn't know the difference. It just reports back.
The three sites researchers identified had this dialed in. They generated pages like "Best Project Management Software for Restaurants" or "Top 10 Accounting Tools for Small Businesses" — exactly the kind of queries an operator types when they're making a buying decision. The pages looked like reviews. They named real products. They used comparison tables. But the underlying content was fabricated, and in some cases the "recommended" tools either paid for placement or were simply invented.
I ran into a version of this myself. An AI assistant I was using cited a "review" to support a vendor recommendation I was evaluating for my import operation. Something felt off — the writing was a little too generic, the claims a little too clean. I clicked through. The page was clearly templated: same sentence structures repeated across dozens of product categories, no author, no date, no real detail. Confident AI answer. Worthless source.
The Real Cost of Fast-But-Wrong Research
Speed is the whole point of using AI for research. You can evaluate five vendors in the time it used to take to read one review thread. That's a genuine superpower for a small operation without a dedicated procurement team.
But fast plus wrong is expensive. Consider what's actually at stake when you rely on AI research without verification:
- Vendor selection: You onboard a software tool based on "top-rated" status that was manufactured. You spend weeks integrating it before realizing it doesn't do what you need.
- Supplier vetting: You shortlist a supplier an AI flagged as reputable. The "review" was planted content. You wire a deposit.
- Competitive intelligence: You benchmark your pricing or product against a competitor analysis that's built on fake comparisons. Your strategy drifts in the wrong direction.
- Compliance and tools: You adopt a platform flagged as "compliant" with a specific regulation. The claim was never verified by anyone.
None of these are hypotheticals. They're the exact use cases SMB operators bring AI into every day. The risk isn't abstract — it's baked into the workflow most people already have.
The Fix: One Extra Step That Changes Everything
The answer here isn't to stop using AI for research. That would be like throwing out your calculator because someone once typed in the wrong number. The tool isn't the problem. The missing layer is.
Add one verification step to every AI research task before a decision gets made. That's it. Here's how we do it in practice:
- Use AI to generate the shortlist. Let it be fast. Let it cast a wide net. That's what it's good at.
- Cross-reference against a primary source. That means the vendor's own website, a known industry database (G2, Capterra, and similar have editorial standards — not perfect, but real), or a direct conversation with the vendor.
- Check the AI's cited sources directly. Click through. Read the page. Ask yourself: does this look like something a real person wrote with real knowledge? Is there an author? A date? Specific detail that couldn't be templated?
- Flag anything that sounds perfectly polished. Ironically, fake review pages are often too clean. Real reviews have friction — complaints, caveats, context. If everything sounds like a press release, treat it like one.
This adds maybe five minutes to a research task. It's the difference between AI as a liability and AI as a genuine competitive advantage.
The operators who get burned aren't the ones using AI. They're the ones using AI without a verification layer — and right now, most workflows don't have one built in.
Build the Workflow Before the Next Decision
215,128 fake pages is not a rounding error. It's a coordinated operation targeting exactly the kind of queries your business runs. The tools you use to research vendors, evaluate software, and validate suppliers are pulling from a web that now includes a meaningful volume of deliberately deceptive content — and your AI assistant cannot always tell the difference.
What you can do is build a research workflow that uses AI's speed while adding a human checkpoint at the moment of consequence. We've done this inside a real import and e-commerce operation — not as a theoretical framework, but as a working system with n8n automations, structured prompts, and clear verification rules that the whole team follows.
If you want to see how that works and build something like it for your own operation, book a call with Maqia at maqia.co. We'll look at your current AI research workflow, identify where the exposure is, and show you exactly how to add the verification layer without slowing down the speed you're getting. One conversation could save you from an expensive decision made on data that was never real.