Ask Perplexity how much a website costs and it answers in a paragraph, with three or four little source links at the end. Millions of pages discuss website pricing. It read a handful and repeated what they said. If you sell anything online, the question worth sitting with is: why those pages?
I spend a lot of time watching how assistants pick sources, because GEO is a core part of what I build for clients. The selection happens in two stages, and each stage filters differently.
Stage one: the assistant searches like a person would
When an assistant needs current information, it does not scan the whole internet. It runs searches against ordinary search engines and indexes, gets back a list of results, and reads the top ones. This is worth repeating because it kills a common myth: there is no separate “AI index” you can optimize for while ignoring Google. If your site does not surface in search results, the assistant never opens it, and nothing else about your content matters. Classic visibility work is the entry ticket; if you have doubts about yours, check whether Google can even see your site first.
Stage two: the assistant reads, and this is where it gets interesting
Once the assistant has a few pages open, it behaves less like a ranking algorithm and more like a careful, slightly paranoid researcher. It is about to state something as fact under its own name, so it favours pages that make that feel safe.
Pages that answer the question directly. An assistant asked “how long does a website project take” will quote the page that says “most projects take four to eight weeks” over the page that spends six paragraphs warming up to it. Direct answers, near the top, in complete sentences.
Claims a machine can verify elsewhere. Assistants cross reference. When your site says you are in Montréal and your Google profile, your LinkedIn and a directory all agree, each mention reinforces the others. When details conflict across sources, the safest move for the machine is to cite none of them.
A clear author and a clear reason to know. A named person with an identifiable background reads as more trustworthy than an anonymous blog. This is not sentiment; attribution is one of the few trust signals a machine can actually check.
Specifics over adjectives. “From 2,500 CAD, delivered in weeks” is usable material. “Affordable and fast” is not a fact, so it cannot be repeated as one.
A structure machines parse cleanly. Headings that describe what follows, short paragraphs, information embedded in a format machines read directly. I wrote a full plain-language piece on that machine-readable layer.
What this means for a business site
Most business websites fail stage two quietly. They rank fine, an assistant opens them, finds nothing quotable, and moves on to a competitor who wrote things down plainly. The fix is not clever. Write pages that state who you are, what you do, for whom, where and for how much, in sentences a stranger could lift and repeat without embarrassment. That is the entire game.
Building sites that pass both stages is part of every project I take on at Studio Squalli. If you want to know whether assistants can currently trust your site, the free audit tests exactly that.