that is the implementation that makes it work, and it is more doable than people fear. connect the generation to the fields each page already has rather than running one generic prompt over everything. the people who got burned on ai descriptions almost all used the one-prompt-for-everything setup and got interchangeable boilerplate. the ones who fed in real page data got varied descriptions that actually help. the difference is entirely in that plumbing.
the honest take on AI descriptions is that they are thin content if generated from a naked template and a real win if generated from actual page data, so the tool is neutral and the input is everything, which is the same lesson as every other AI-at-scale question. so it is the data you feed it, not the AI itself. so judge AI descriptions by what you feed the generation rather than by whether you use AI at all, because descriptions built from real per-page data are genuinely descriptive and ones from a generic prompt are duplicate filler that google reads as the thin content it is, so the operators who win with AI descriptions feed real data and curate while the ones who get penalized point a template at everything and publish the interchangeable output, which means the whole AI-descriptions question is really about the input discipline, and the honest answer is that AI descriptions are worth doing if and only if you feed them real per-page data, since that single factor determines whether they help your pages or become the thin-content liability that a naked-template approach reliably produces at scale.
the trick with AI descriptions is feeding it the real per-page data, the actual tags and attributes and specifics of each page. then you get descriptions grounded in what the page actually is. generate from nothing and you just get interchangeable filler that could sit on any page.
generic AI descriptions actively hurt you, grounded ones help, so the whole thing on the wordpress side is wiring the real page data into the generation. a plugin spitting out descriptions from nothing gives you filler across the entire site. feed it each page’s actual data and the output is worth having.
the AI-written-descriptions question comes down to grounding and review, because AI descriptions work when generated from real page data and spot-checked for quality, and fail when generated generically and trusted blindly, so the honest answer is AI descriptions are useful with the grounding and oversight and harmful without. so AI descriptions are useful with grounding and oversight, harmful without. so use AI descriptions with real-data grounding and quality oversight because they help that way and hurt generated blindly, because AI descriptions grounded in real page data and spot-checked for quality are useful while generic ones trusted blindly are harmful filler, so the operators who use AI descriptions well ground the generation and review the output rather than generating generically and trusting it, which means the grounding and oversight determine whether AI descriptions help, since grounded reviewed descriptions are useful and generic unreviewed ones are harmful, and using AI descriptions with the real-data grounding and quality oversight rather than blind generic generation is the honest approach, making the grounding-plus-oversight the answer, because AI descriptions are genuinely useful when generated from real per-page data and spot-checked for quality and genuinely harmful when generated generically and trusted blindly at scale, and the operators who benefit ground the generation in real data and maintain quality oversight rather than treating AI description generation as a blind set-and-forget that produces the generic filler which hurts the site.