The overlap and the gap between traditional SEO and LLM SEO

The most common question I get from clients about AI search is whether their existing SEO investment still matters. It is the right question. There is a lot of noise from agencies repackaging old work as GEO and there is also a lot of genuine change happening in how AI products select sources.

My short answer is that most of what was good SEO is still good GEO. Some of it does not transfer. A small number of things are genuinely new. This post is the practical map of what overlaps, what is new and what is worth less than it used to be, with a decision framework for where to put your next quarter of effort.

How AI search products actually pick sources

It helps to know what the AI is doing before deciding what to optimise. When a user asks ChatGPT search, Perplexity, Claude with web access or Google AI Mode a question, the system runs one or more searches against an underlying index. Google AI Mode uses its own index. ChatGPT search relies on a mix of its own crawl and Bing. Perplexity runs its own index of tens of billions of pages and tops it up with live crawling. Claude with web access uses Brave search.

The system pulls back a candidate set of pages, often ten or so and either reads the full pages or pulls passages from them. The model then assembles an answer from those passages, ranks the candidates by some internal scoring and decides which to cite. Citations tend to favour sources that match three things. They have to rank for the query in the underlying index. They have to give a clean, citable passage that answers the query. They have to read as authoritative enough that the model treats them as worth attribution.

That stack of decisions is what tells you which traditional SEO levers still pull and which ones do not.

What carries over from SEO to GEO

These are the practices that earn their keep on both fronts.

Topical authority and link equity

The candidate set the AI works from is mostly the same set Google or Bing would have produced for the query. If you do not rank for the topic, you are not in the candidate set and you cannot be cited. Domain authority and topical depth still feed the candidate set, so backlinks, on-site clusters of related content and the patient work of building site-level authority are still doing the job they used to do.

Structured data and schema

Schema is more useful now than it used to be. Traditional SEO used schema mainly for rich snippets. AI products read schema to understand what your content is, who wrote it, which entities it describes and how confident they should be in citing it. JSON-LD is the format every AI surface I have tested prefers. Author schema, organisation schema, product schema and article schema all pay back if you have not already done them.

Page speed, mobile, technical hygiene

The crawlers that AI products rely on still hit the page. If your page is slow, broken or returns a soft 404, you are out of the candidate set. None of the technical SEO basics have stopped mattering. Sitemaps, robots, canonical tags, internal linking, valid HTML. If your last technical audit was more than a year ago, schedule one before adding anything new.

On-page clarity and answer-first writing

Pages that answer the question in the first paragraph have always done well in search. They do even better now because the AI is pulling a single passage to cite and the cleanest passage tends to be the one that does not bury the answer under a 200-word SEO intro. Get to the answer fast. Put the supporting context after.

Freshness

AI products show a stronger preference for recent content on time-sensitive topics than traditional search did. Old pages on evergreen topics still rank. Old pages on a topic that has changed get displaced quickly. Annual review of cornerstone content was always a good idea and is now closer to required for anything where the underlying facts shift.

Brand search

Brand search has always been the cleanest signal of marketing health. It is more important now because branded queries are one of the few categories where AI products do not eat the click. A user who searches your name is still going to click your site. Build the brand and the search traffic comes through.

What is genuinely new

These are the practices that did not have an obvious analog in traditional SEO and that are worth attention now.

Citation-worthiness

AI products will skip a high-ranking page that gives a mushy answer and cite a lower-ranking page that gives a clean, specific one. A page can do everything right in SEO terms and still not be cited if its content is hedged, vague or padded. The practical move is to write pages with a clear thesis, named examples, specific numbers where they exist and an opinion where appropriate. The pages I see getting cited regularly tend to read like a confident developer or operator explaining a thing, not like a content marketing team trying not to upset anyone.

Distinctive viewpoint

There is no value in being the third article that paraphrases the same Wikipedia paragraph. The AI already has Wikipedia. What it does not have is your specific experience, your contrarian read, your numbers, your warning about a thing you tried that did not work. Distinctive content beats generic content on AI citation odds. This is the strongest argument I know for first-person writing on your blog if you have the standing to do it.

Entity consistency across the web

AI products build a sense of who and what an entity is by reading your site, your social profiles, your GitHub, your LinkedIn, your Companies House record and various directories. When those sources contradict each other, the AI gets less confident about the entity and is less likely to cite. When they line up, the entity becomes a stable thing the AI can talk about. Audit the consistency. Same name spelling, same description, same logo, same role across all the places you appear.

llms.txt where it makes sense

llms.txt is a plain-text file you place at the root of your domain to give AI systems a curated map of your most useful pages. Adoption sits in the low single-digit percentages of crawled domains as of early 2026 and the impact data is mixed at best, with several published analyses finding no measurable lift in AI citations. The case for setting it up is that it is cheap to do, not that it is proven. Worth setting up. Do not expect a measurable lift on its own.

Machine-readable proofs

Citing third-party validation in a way both humans and machines can read. Case studies with named clients and outcomes. Reviews and testimonials marked up with review schema. Awards and certifications with names and dates. Code repositories with usage counts. These signals influence whether the AI treats you as authoritative, more than a self-rating on a homepage does.

What is worth less than it used to be

A short list of tactics that have lost value in the new mix.

Long, SEO-padded intros. They reduce citation odds because the answer is too far down the page. Tighten them.

High-volume informational content built mainly for traffic. The traffic from these pages is the most exposed to the click decline. New content of this kind is a worse investment than it was two years ago.

Generic top-ten listicles. AI products write these themselves and do not need to cite a thin third-party version. If your listicle does not include a perspective the AI cannot synthesise from common knowledge, it is not in the citation set.

Pure keyword-density work. Old-school keyword stuffing was already a poor strategy. It is worse now because AI products tokenise and embed your content, so synonyms and concepts matter more than exact-match repetition.

A decision framework for the next quarter

For a UK business with limited time and budget, this is the order I would prioritise in.

Audit technical hygiene first. Fix anything that keeps you out of the candidate set. Schema, sitemaps, speed, internal links, broken pages. The cost is low and the cost of not doing it is large.

Sharpen your cornerstone pages second. Answer-first paragraphs, distinctive examples, clear thesis. These are the pages most likely to be cited if you can get them into the candidate set, so they deserve disproportionate attention.

Tighten entity consistency third. Same descriptions across LinkedIn, About page, Companies House, GitHub, directories. Add organisation and author schema if missing. This is the unglamorous work that quietly pays back for years.

Build distinctive content fourth. First-person, specific, opinionated where appropriate. This is what makes you citation-worthy when an AI is choosing between similar candidates.

Add llms.txt and refine schema fifth. Low-cost, low-but-positive-expected-value. Worth doing once the more impactful work is in place.

Deprioritise high-volume informational content unless it serves a specific purpose other than traffic. The math has changed and the unit economics on that content are worse than they were.

The shorter version

Most of what was good SEO is still good. The new work is about being chosen as a source once you are in the candidate set and about being a clearly described entity that AI systems can confidently cite. The work that has lost value is mostly the work that was always slightly suspect anyway.

If you are weighing where to put your next quarter of effort and want a more specific read on your site, I have written about the broader picture in generative engine optimisation explained and about the traffic shift in why search traffic is being quietly rerouted through LLMs. For client work, see AI integration services.