What actually changed
For twenty years the transaction was simple: a person typed a question, saw ten links, and clicked one. Now a large share of questions are answered in place — by an AI overview above the results, or by an assistant the person never leaves. The click is optional.
The instinctive reading is that SEO is finished. The accurate reading is narrower: informational queries with generic answers have been absorbed, while queries that require judgement, local specificity, credentials, current pricing or a decision still send people to sites. Losing traffic to "what is technical SEO" matters far less than owning "who can certify lead paint clearance in Brooklyn".
Being the source rather than the tenth link
Assistants synthesise from pages they can parse and trust, and they name sources. That makes citation the new ranking. Pages that get cited tend to share unglamorous traits: one clear subject, claims stated plainly in the opening lines, concrete numbers, and structure an extractor can follow — real headings, short paragraphs, lists where a list is honest.
Hedged, padded writing performs badly here for the same reason it performs badly with readers: there is no sentence a model can lift as an answer. If your page never states a position, nothing can quote it.
Specificity is the whole strategy
The pages that survive are the ones a general model cannot generate from its own training data. Local reality, prices you actually charge, the ordinance that changed this year, the shape of the report you deliver, the failure modes you have seen thirty times — none of that can be synthesised, so it must be cited.
This is why programmatic city-page spinning has aged so badly. Fifty near-identical pages with a town name swapped in were always weak; now they compete against a system that can produce that text instantly and for free. One genuinely local page that names what differs in that jurisdiction beats fifty templates.
Structured data and machine legibility
Schema markup does not make a page rank, but it makes a page unambiguous: this is a person, this is the service, this is the area served, this is a review, this is the publication date. Ambiguity is expensive when a machine is deciding whether you are the right source.
The basics cover most businesses: an entity for the person or organisation, one for the site, service markup on service pages, breadcrumbs on deep pages, and article markup where you publish. Keep it accurate and keep it consistent with the visible page — invented markup is a liability, not an edge.
Measure enquiries, not impressions
If assistants answer more and click less, sessions can fall while qualified enquiries rise. Judging performance on traffic alone will make you optimise for the wrong thing, and often make you abandon pages that are quietly doing the work.
Track the small set of numbers tied to money: enquiries by source, calls, the queries where you appear as a cited source, and the pages last touched before contact. That is enough to steer with, and it survives every future change to how results are displayed.