SEO is splitting into two disciplines with different rules. The classic half, ranking in the search index, still works, still responds to on-page structure, and still delivers visibility in days. The new half, being cited by an answer engine, is won mostly off your page, through references and mentions, and the tactics vendors sell for it barely move it. The optimisation target is shifting from "rank my page" to "be the source the model quotes", and those require different work.
Ross Jones, Founder, The Hopium Lab. Last modified 22 July 2026.
A forward read, honest about what is proven and what is still a bet. Companion to the field report on getting a new site showing up, which covers the classic half with data.
What is the target that actually changed?
The target changed from a ranked list of links to a synthesised answer, and that changes what winning means. For twenty-five years, SEO optimised for a position in a list, get to the top few and you get the click. When the interface becomes an assistant that reads sources and composes one answer, there is no list to top. There is a set of sources the model drew from, and either you are in it or you are invisible, with no page two to console you.
So the question stopped being "what is my rank" and became "am I one of the handful of things the model pulled from to build its answer". That is a different game, and the uncomfortable part is that being in the model's answer is not something you can fully control from your own page, the way you could control a title tag.
Ranking got you into a list a human scrolled. Citation gets you into an answer a machine composed. The first was won on your page. The second is mostly won off it.
Does the old on-page playbook still work?
Yes, for the classic half, and it works fast, which is exactly why it gets over-applied to the new half where it does not. Question-shaped titles, standalone first-sentence answers, clean server-rendered HTML, one page per query: that playbook fills a search-console impressions graph in days, and I have watched it do so from zero on more than one site. It is real and it is not going away, because the classic index is not going away.
It also lays the foundation the answer engines need, because a page a crawler cannot parse is a page a model cannot quote. So the on-page work is necessary. The mistake is believing it is sufficient for citation, that the same tricks that won you rank will win you the model's answer. They will not, and the evidence on that is unusually clear.
What actually moves the answer-engine needle?
Off-site signals move it, references, mentions, and presence in the places models weight, far more than anything you do to your own markup. A study of Product Hunt startups found on-page optimisation had roughly no correlation with whether an assistant surfaced a brand, while referring domains and community presence did. A large brand-mentions analysis put branded web mentions and links well ahead of on-page factors for AI visibility.
The tactics being sold for this are mostly the ones the evidence says do not work. Adding schema markup produced a small decline on one AI-answer surface in a controlled test. The serving-file convention pitched as the way to talk to AI crawlers gets fetched a negligible number of times against hundreds of millions of bot visits, the crawlers read your HTML directly. There is also a recognition-versus-discovery gap worth knowing: assistants will recall a brand they are told about far more readily than they will volunteer it in an open query, which means being nameable is not the same as being discovered, and only the second one brings you anything.
| Tactic | Moves classic rank | Moves AI citation | Verdict |
|---|---|---|---|
| Question-shaped, well-structured pages | Yes, fast | Sets the foundation | Necessary, not sufficient |
| Clean server-rendered HTML | Yes | Prerequisite to being quoted | Do it |
| Schema markup | Helps rich results | Slight negative in one test | For rich results only |
| Serving-file conventions | No | Barely fetched | Skip as a citation lever |
| Referring domains and mentions | Yes | The main lever | The actual work, and the slow bit |
| Being the primary source others cite | Indirectly | Strongly | The durable strategy |
What does "feed the model layer" actually mean?
It means becoming the thing a model reaches for, a primary source that other content cites, rather than one more page competing to be reached. The durable move in an answer-engine world is not to out-optimise everyone for a query. It is to be the original, checkable source that the assistant, and the humans writing the content the assistant reads, keep pointing back to.
That is why the whole content strategy on this site is primary-source-shaped: real numbers, published specs, dated first-person incidents, forkable artefacts. Not because it ranks better in the classic index, though it does, but because a primary source is what gets referenced, and references are what feed the model layer. You do not optimise a page to be quoted. You become the kind of source that is worth quoting, and you accept that the payoff is slower and mostly earned off your own property.
You cannot markup your way into an AI answer. You earn your way in by being the source worth citing, which is slower, off-site, and the one thing no plugin sells you.
What are the failure modes?
Four, and they cluster around treating the two halves as one.
Buying GEO snake oil. Paying for schema packages and serving files sold as AI-citation levers, when the controlled evidence says one slightly hurt and the other is barely read. Money spent on the tactics that measurably do not work.
Abandoning classic SEO as dead. Declaring search over because of AI answers, and stopping the on-page work that still delivers fast, measurable visibility and lays the citation foundation. The classic index still sends enormous traffic; it did not vanish because assistants arrived.
Optimising the page for citation. Pouring effort into on-page changes hoping to be quoted, when citation is won off-page. Right effort, wrong surface.
Confusing nameable with discovered. Getting a model to recognise your brand when prompted and mistaking it for the model surfacing you unprompted. The first is a party trick; only the second is distribution.
How do you split your effort?
Run both disciplines, and put your effort where each one actually responds.
SEO IN TWO HALVES
The Hopium Lab · v1.0 · 22 July 2026 · take it, fork it, argue with it
THE CLASSIC HALF (fast, on-page, proven)
[ ] One page per real question, question-shaped title
[ ] Standalone first-sentence answer, clean server-rendered HTML
[ ] Read the graph right: impressions before rank before clicks
[ ] This still works. Keep doing it
THE ANSWER-ENGINE HALF (slow, off-page, the new frontier)
[ ] Be a primary source: real data, specs, dated incidents, artefacts
[ ] Earn referring domains and mentions, the main lever
[ ] Be present where models weight discussion, not just on your site
[ ] Do NOT buy schema or serving files as citation levers
THE OVERLAP (do once, serves both)
[ ] Parseable HTML and clear structure: foundation for rank AND citation
THE TEST: for any tactic, ask "does the evidence say this moves rank,
citation, both, or neither?" If neither, someone is selling it to you
because it is easy to sell, not because it works.
The honest forward read is that SEO did not die and did not stay the same. Its fast, controllable, on-page half is intact and still worth doing on day one. Its new half, being the source a machine composes into an answer, is real, is coming, and is won by the oldest thing in the business: being genuinely worth citing, and getting other people to point at you. The tools changed. The thing that earns you a place in the answer did not.
Ross Jones, Founder, The Hopium Lab. Last modified 22 July 2026.