SEO is not just titles and keywords anymore

Most SEO advice still describes a job that was mostly accurate a decade ago: write good titles, choose keywords, get links. None of that is wrong. It is just no longer where the interesting work is.

The interesting work is making a machine understand what a thing is.

Entities, not pages#

A search engine is not only matching a query to a page. It is trying to establish that an entity exists, what kind of thing it is, and how it relates to other entities.

That reframes the technical work. Instead of optimising pages in isolation, you are describing a subject consistently enough that a machine can build a confident model of it.

For a company that means: this organisation exists, does these things, is located here, and these pages are about it. For a person: this individual exists, has these roles, wrote these articles, and is the same individual referenced on these other sites.

Structured data only counts when it is connected#

This is the most common implementation mistake I see.

A site adds JSON-LD. There is an Organization block on the homepage. There is a WebPage block on each page. Neither references the other. A BreadcrumbList appears with no relationship to anything.

That is a set of isolated facts. It is not a description of an entity.

The version that works uses stable identifiers and references between nodes. A page node that says it is part of a website. A website that says who publishes it. Articles whose author points at the same person identifier every time rather than repeating a name.

On this site, everything references one Person identifier — #person — because that is the difference between one corroborated entity and thirty separate mentions that never consolidate.

Canonical architecture is unglamorous and decisive#

Duplicate titles across routes. Missing canonicals. Trailing-slash variants competing with each other. www and apex both resolving.

None of it is interesting and all of it dilutes. The fix is to generate metadata per route from a single source rather than authoring it per page, because anything typed into a field by hand eventually duplicates and the duplication is invisible until something diagnoses it.

Consistency across sources you do not control#

Here is the part that is not engineering.

If a person's name, role and website differ across LinkedIn, GitHub, X, a conference bio and their own site, no amount of structured data on the site compensates. Corroboration is what turns a claim into an entity, and corroboration happens across sources.

So the tedious work — making every external profile say the same thing and link back to the same place — does more for entity clarity than most on-site work.

GEO, and what it can honestly promise#

Generative engines cite sources. That has produced a category of advice about optimising for AI citation, some of which is reasonable and some of which is straightforwardly invented.

What is reasonable: content should be server-rendered and crawlable; claims should be attributable and specific; identity should be consistent; a page should be structured so a passage can be quoted usefully. All of that is good practice independent of AI.

What is not reasonable: promising citations. Nobody controls what a model surfaces, the systems change frequently, and anyone guaranteeing placement is guessing or selling. I would be suspicious of confident numbers in this area, including my own.

What I actually do#

• Generate metadata per route from one source.
• One canonical host, one canonical form per page.
• Structured data as a connected graph with stable identifiers.
• Breadcrumbs in markup that mirror the visible trail exactly.
• Server-rendered content — no identity or article text behind client rendering.
• The same name, role and link on every external profile.
• Write things worth citing, which remains the part with no shortcut.