When the grounded answer just echoes the affiliate listicle

Security is the category I most expected grounding to struggle with. "Best VPN" is a battlefield of paid affiliate listicles. Most of the top web results earn a commission on the apps they rank. Because this index runs every assistant with web search on, it inherits whatever that web says. So the test is simple: do grounded assistants cut through the affiliate consensus, or repeat it?

Mostly, they repeat it. ProtonVPN and NordVPN top the board, with Surfshark close behind, the commission-driven names the listicles have pushed for years. Google's AI Overview, asked for the best VPN, returned exactly that trio. The chat models add ExpressVPN to the headline answer, but it lands twelfth on the full board: the names that actually lead are the affiliate staples. On the headline question, the grounded answer and the affiliate answer mostly agree.

People already sense this. The most common thing they append to "best VPN" in a search box is the word "reddit", a request to route around the listicles the assistant is reading from.

That "reddit" tic, which showed up across the demand data for this category, is the tell. Users distrust the ranked listicles enough to ask a forum instead. The grounded assistant, left to its own retrieval, goes to the listicles anyway. If your category is one where the top of the web is bought, the assistant inherits the purchase.

It isn't pure echo

One result complicates the cynical read. Ask not for the best VPN but the most private, no-logs one, and the answer changes: Mullvad moves to the front. Mullvad runs no affiliate programme and barely advertises; it wins on reputation in privacy communities, not on commission. So when the question gets specific enough, the grounded answer does reach past the paid consensus to the thing the experts actually trust.

That's the more useful lesson than "AI just repeats SEO." The headline query echoes the affiliate web; the specific, intent-loaded query rewards genuine reputation. Which is also the whole argument for measuring by job, not just by category.

What this is

Thirty-three questions put to four AI assistants, plus Google's AI Overviews captured from the web. US English. One category spanning VPNs, password managers, antivirus and data-removal, with the per-job winner read from the intent slices. The prompt set was validated against real Google autocomplete demand before it was fixed. See the security & privacy index for the full board.