AI SEO in 2026: what changed, what still works, and what to stop doing
Generative search did not kill SEO; it changed the endpoint. Traffic that used to land as a click now sometimes lands as a mention in a generated answer. Here is what to keep doing, what to add, and what to stop.
4 min readStallwart
The endpoint changed, not the game
For a long time SEO had one endpoint: a person clicked through to your page. Generative search added a second one. Some queries never leave the answer surface at all; the reader gets a composed response and moves on, and if your page was one of the sources, your brand was in front of them without a click ever landing. Traffic did not stop; it split into a channel you measure and a channel you do not.
That is the whole shift, and everything worth doing in AI SEO flows from it. The pages that work now are the ones that earn a click when the reader wants to go deep and earn a mention when the reader just wants a fast answer. They are different jobs on the same page, and most sites are still doing one and not the other.
What still works
The fundamentals did not move. High-intent queries still convert best. Deep, specific pages still beat broad, thin ones. Authoritative external coverage still compounds everything. Technically sound sites still rank; broken ones still do not. Anyone selling you a total reset for the AI era is charging for a rewrite you probably did not need.
- Answering the query the reader actually typed, not the topic the SEO tool suggested.
- Original writing with a point of view. Rehashed listicles were losing ground already; AI accelerated it.
- Internal links that concentrate authority on the pages you want to be found for.
- Fast, accessible, mobile-first pages. Core Web Vitals still rank; they just no longer decide.
- Coverage from trusted sites in the same space. Backlinks matter, brand mentions matter more.
What to add
The new work is not another SEO gimmick. It is engineering the page so a machine can quote it and citing sources cleanly enough that the machine trusts it. The specifics come out of the AEO and GEO playbooks: lead with a self-contained answer, publish an FAQ block with FAQPage schema, cite numbers and standards by name, keep canonical URLs stable.
- Lead every page with a single self-contained answer to the question the page is about.
- FAQPage and Article JSON-LD schema on every post.
- An llms.txt at the root, inventoried from the same data as the site so it never drifts.
- Plain-language variants of anything technical, so a summary can quote them cleanly.
- Measurement that captures citations and mentions in AI answers, not only clicks.
What to stop
Some habits that used to be neutral or mildly helpful are now actively damaging in a world where a language model is reading and composing.
- Publishing thin content on adjacent long-tail keywords. Language models collapse them into one answer and the padding stops helping.
- Splitting the same content across parallel URL namespaces (`/blog/x` and `/insights/x`). Ranking signal splits; citations become inconsistent.
- Moving cited URLs to freshen them. A moved URL is a broken citation somewhere else.
- Aggressive keyword stuffing. Answer engines read for meaning, and they punish pages that read as spam even when older ranking models ignored it.
- Blocking AI crawlers by default in robots.txt. That is a decision to opt out of the second endpoint; make it deliberately, not by copy-paste.
How to measure the new endpoint
The traffic that never becomes a click is real but harder to see. A workable stack: a light branded-search monitor to see if your name is trending in the wake of a topic, a periodic manual check across the major answer engines for the queries you sell against, and a review of referral traffic from AI surfaces where they do send clicks. None of it is a perfect metric. All of it is better than measuring only what analytics has always shown.
The point is not to obsess about attribution. The point is to notice, early, when you are being cited more or less often, and to have some idea why. That signal changes what you write next, which is the whole loop.
The short version
- Generative search added a second endpoint: a mention inside a composed answer. Traffic split; it did not stop.
- Fundamentals still hold: high-intent queries, deep pages, authoritative coverage, technically sound sites.
- Add: self-contained answers up front, FAQPage and Article schema, an llms.txt at the root, measurement that captures citations.
- Stop: thin adjacent-keyword pages, split URL namespaces, moving cited URLs, keyword stuffing, and blocking AI crawlers by copy-paste.
Questions this raises
- Did AI kill SEO?
- No. It added a second endpoint. Traffic that used to always end at a click now sometimes ends at a citation in a generated answer. Pages built for both endpoints do fine; pages built only for the click miss the second half.
- What still works in SEO in 2026?
- The fundamentals: answering the query the reader actually typed, deep and specific pages, coverage from trusted sites, sound technical hygiene, and internal links that concentrate authority. None of that moved with the shift to AI search.
- What should I stop doing for SEO in the AI era?
- Stop publishing thin adjacent-keyword pages, splitting content across parallel URL namespaces, moving cited URLs to freshen them, keyword stuffing, and blocking AI crawlers by default. Each of those actively hurts a site whose second endpoint is a language model composing an answer.
- How do you measure AI SEO visibility?
- Combine a branded-search monitor, periodic manual queries across the major answer engines for the questions you sell against, and referral analytics from AI surfaces that do send clicks. None of these is perfect; together they show whether citations and mentions are growing over time.
- Should I block AI crawlers in robots.txt?
- Only as a deliberate choice, not by default. Blocking AI crawlers is a decision to opt out of the second endpoint (citation inside a generated answer), which for most B2B sites is a visibility loss they did not intend.
Recognize this in your own operation?
Bring us the version of it happening in your business and we will tell you which part a system can take over.
Book a call