Plenty of pages will tell you AI SEO is a new discipline, and plenty will tell you nothing has changed. Both are selling something. The practical question for anyone doing the work is narrower: what lands differently in the week once AI search matters to your category?
This page answers that and nothing else. For the conceptual comparison between ranking and being cited, see SEO vs GEO, which covers it properly.
What carries over untouched
The foundation is the foundation. Crawl access, clean indexation, a coherent site architecture, fast templates, internal links that concentrate authority where it converts, and real credibility earned from elsewhere. None of this became optional, and no amount of AI-era tactics substitutes for it.
If anything, one item got stricter. Most AI crawlers do not execute JavaScript, so a client-rendered marketing site that Googlebot handles tolerably can be effectively invisible to the systems generating answers. Server-side rendering moved from a recommendation to a prerequisite without most teams noticing.
What changes in the writing
The format of a page now affects whether it gets used, independently of whether it ranks. Research presented at ACM SIGKDD in 2024 tested nine content strategies across 10,000 queries and found adding quotations was the strongest, at roughly 41%, with citing sources and adding relevant statistics close behind.
In practice that means three habits: answer the question in the first forty to sixty words rather than circling it, attribute claims to a named source instead of asserting them, and include specific figures where they are genuinely relevant. These are not AI tricks. They are what good technical writing has always looked like, now with a measurable payoff.
The constraint is honesty. A model quoting you is lending you its credibility, and an invented figure is the fastest way to lose it. Padding a page with decorative statistics fails for the same reason it fails with human readers.
What stops paying, and what now costs you
Keyword density was the one tactic carried over wholesale from an older era, and the research is unusually clear about it: on a live engine it performed about 10% worse than making no changes at all. It is the only tested method that left content less visible than doing nothing.
Volume for its own sake stops working too. Thin pages published to cover a keyword were already weak; now they are also unlikely to be retrieved, because there is nothing in them worth quoting.
Chasing clicks on purely informational queries pays less than it did. Ahrefs found a 34.5% lower position-one click-through rate where an AI Overview appears, with informational queries absorbing most of it. That is an argument for weighting effort towards commercial pages rather than for abandoning content.
The genuinely new work
Measurement is the real addition. There is no Search Console for assistants, so you build the instrument: a fixed set of prompts, run on a schedule across each engine, recording which brands are named and which sources are cited. Without it, losses are invisible, because no analytics event fires when a model recommends a competitor.
Corroboration becomes a workstream rather than a byproduct. What is written about you elsewhere shapes the picture a model has of your category, which makes comparison posts, community discussion and industry coverage worth deliberate attention rather than being treated as link building with a different name.
And one thing that is not new work, despite being sold as such: there is no AI-specific markup to add. Google states plainly that no additional requirements or special optimisations are necessary for its AI features. Anyone charging for an AI schema implementation is charging for a file.