LLM & AI Search Optimisation

LLM and AI search optimisation is the practice of publishing information so that AI assistants and answer engines can understand it, verify it and reference it accurately. It is also called generative engine optimisation (GEO), answer engine optimisation and AI search optimisation — the terminology is still settling. In practice it means answer-first writing, clear structure, unambiguous entity references, visible sourcing and information that stays consistent across a site.

What has actually changed

For twenty-five years, being found online meant appearing in a list of links and persuading someone to click one. The page was the destination and the ranking was the prize.

Increasingly, something reads the page on the visitor's behalf and gives them an answer. That happens in AI assistants, in generative search results, in conversational tools and in the growing number of applications with a language model behind them. The person may never see the page at all.

This changes what a page has to do. It is no longer competing only for a position in a list; it is competing to be the source a system draws on, and to be understood correctly when it is. A page that a language model cannot parse, or parses wrongly, is invisible in the new channel however well it ranks in the old one.

What this does not change

It is worth being clear about the continuity, because a great deal of what is currently sold as AI optimisation is search advice from 2015 with new labels on it.

The fundamentals have not moved. Useful information, clearly structured, accurately stated and properly sourced has always been what search engines were attempting to identify. AI systems are attempting to identify the same thing by different means. A site built to be genuinely good is already most of the way there — which is why Ferguson Media Group treats this as the latest stage in how discovery works rather than as a separate discipline requiring a separate budget.

The practices

These are the things that demonstrably help a machine understand a page, and every one of them also helps a person reading it. That coincidence is not accidental and is the reason none of this is a trick.

Answer-first writing

State the answer in the first paragraph, then explain it. A page that builds to its conclusion gives an extractive system nothing to extract from the top of the document, where it is most likely to look.

Standalone paragraphs

Write so that a paragraph still makes sense lifted out of the page. That means naming the subject rather than relying on "it", "this" or "the above".

Clear entities

Name organisations, people, products and places explicitly and consistently. A system building a picture of who you are cannot do it from pronouns.

Semantic HTML

Correct headings, lists, tables and landmarks. The structure of the markup is how a machine infers the structure of the argument.

Structured data

Accurate JSON-LD describing what is genuinely on the page. Schema that overstates or invents is worse than none, because it teaches a system to distrust the source.

Facts as facts

Specifications in tables, definitions as definitions, figures with units and dates. Information buried in prose has to be reconstructed; information in a table can be read.

Source attribution

Say where a claim comes from and link to it. Verifiability is one of the few quality signals that is genuinely difficult to fake at scale.

Author and publisher information

Who wrote this, who publishes it, and what they know about the subject — stated plainly and consistently across the site.

Visible dates

When something was published and when it was last reviewed. A system weighing two conflicting answers has little else to go on.

Internal consistency

The same fact stated the same way everywhere. Contradictions across a site do not average out; they make the whole source less reliable.

Genuine questions and answers

FAQ sections where people actually ask those questions, answered directly. Not a keyword list rewritten as interrogatives.

Original material

Research, data, testing or expertise that exists nowhere else. There is no substitute, and it is increasingly the only durable advantage.

Terminology

Several names are in circulation for roughly the same set of practices, and none has settled.

The distinctions are mostly marketing. What matters is whether the underlying practices are sound, and they are largely the same set in each case.

How this work is done

A typical engagement looks at what a site says, how it is structured, and whether the two agree.

1. Read the site as a machine would

   What is in the HTML before scripts run, what the heading structure implies, what the structured data claims, and where the two diverge.

2. Check the entity picture

   Is it consistently clear who publishes this, who wrote it, and what the organisation actually does? Inconsistency here undermines everything else.

3. Find the unanswerable questions

   The things a visitor or an assistant would want to know that the site does not state anywhere.

4. Restructure for extraction

   Answer-first openings, standalone paragraphs, facts moved into tables and lists, vague references replaced with names.

5. Add evidence

   Sources named, dates made visible, methodology stated where figures are published.

6. Fix the technical layer

   Server-rendered content, accurate schema, clean internal linking, and machine-readable discovery files where they are appropriate.

7. Review consistency

   The same facts, the same names, the same claims across every page — including the ones nobody has looked at for two years.

Machine-readable discovery files

A convention has emerged of publishing an `llms.txt` file: a plain-text summary at the root of a site describing what it covers and pointing to its most important pages. This site publishes one at /llms.txt.

It should be treated for what it is — a convention some publishers have adopted, not a standard that any AI system is obliged to read. It costs almost nothing to publish and may help. It guarantees nothing, and any claim that it ensures inclusion in AI results should be disregarded.

The same scepticism applies to the various other proposed files and protocols in this area. Ferguson Media Group implements the ones that are established and technically sound, and waits on the ones that are not.