Agentic SEO

Search engine optimisation used to be bought after launch, as a piece of work to find what the build had missed. Our agentic product lifecycle now does most of it while the site is being made, with nineteen years of practice built into every stage. The technical work and the structure of every page are done as each page is built, checked by the industry's own tools and gated on every pull request. What hasn't changed is the part that needs a person: what each page is for, the words on it, and whether anyone links to it.

The three kinds of SEO

The trade splits SEO into technical, on-page and off-page work. What a machine can verify is agentic, and what a person has to judge stays human. Technical SEO is all agentic, off-page is all human, and on-page splits down the middle.

What became agentic

Technical SEO and the structure of each page are now a floor the skills apply as a page is made, proved by Lighthouse and a request-level test sweep on a sample of pages, using the industry's own tools and no checks of ours. A gate on every pull request stops a fixed defect coming back. The engines are told where the site is, through Search Console and Bing Webmaster Tools, and their reports are read by a person rather than by code.

The piece with the best evidence behind it is the least glamorous. A rebuild that changes URLs without a redirect map loses what those URLs earned, and the loss shows in Search Console within weeks. So the map is built from the old sitemap before launch, lives in the framework's own config, and is asserted by a test that fails if an entry goes missing. The gate before production reads that test, and a person decides what a red one means.

What stays human

A page that nobody searches for ranks perfectly for nothing. Deciding which question a page is the answer to, in the words people use, is the half of keyword research the evidence supports, and it belongs to whoever owns the words, with the site's own query data in front of them once it is live. The same goes for whether the content is worth reading and citing. An agent proposes the line and reports what needs rewriting; a person decides.

Links from sites people trust remain the strongest signal any engine has, and they are earned off the site, by the work being worth citing and by someone telling people about it. That is marketing, and no build skill substitutes for it. Our audit names it every time as the lever it did not touch, so a clean technical score is never read as a promise of traffic.

What we ask of a client

On day one we send one list, so the access arrives while the automated audit runs, and anything missing becomes a stated gap in the report, not a silent one:

• Google Search Console for every property: indexing, the canonical Google chose, sixteen months of queries and AI Overview impressions.
• Bing Webmaster Tools: Copilot and Bing AI citations per page, with the queries behind them, the only query-level AI data any provider publishes.
• Analytics, for visits from AI assistants and the pages they land on.
• The CDN or firewall dashboard, because blocking AI crawlers by network address overrides robots.txt and can't be seen from outside.
• Thirty days of server or CDN logs, for which bots fetched which pages and where they hit errors.
• The previous sitemap, if URLs have changed, so the redirect map covers every old address.

And four decisions only the client can make: which AI crawlers are allowed, for training, search and user-initiated fetches separately; which markets and languages matter; the ten to twenty pages that earn the money; and what they want to be found for, in their own words.

The same work serves the answer engines

GEO and LLM attribution reaches the same conclusion from the answer-engine side. A site that is well made in the plain sense, with its content in the HTML, every page at a stable address, one heading structure and provenance on the page, is found by search engines, cited by assistants and usable with assistive technology from one effort. There is no separate SEO workstream and no separate AI-readiness workstream. There is building it properly, and then the two human levers.

What we do not cover

Rank tracking as a measure of success, since results are personalised and move daily. Authority scores, which are one vendor's model of another company's index. Keyword difficulty and volume figures, which are estimates from bought clickstream panels sold with more precision than the method allows. Readiness scores and llms.txt as visibility levers, for the reasons in GEO and LLM attribution. A monthly retainer of optimisation with no hypothesis about what will change.

None of this makes the field worthless. The parts with evidence are boring: redirects, indexing, real URLs, one intent per page, and links. Those are the parts the skills build in, and a client can check them in the engines' own reports.