LLM SEO: what actually changes, and what doesn’t

Is LLM SEO the same as GEO, AEO and AI search optimization?

Effectively yes. SEO for LLMs, AI search optimization, ChatGPT SEO and generative engine optimization all describe roughly this activity, with GEO being the term the academic literature settled on. There is no meaningful technical distinction between them, only differences in which term a given audience recognizes.

How does getting into an AI answer differ from ranking a page?

Ranking is a sorting problem. Retrieval into a generated answer is a selection and composition problem, and the difference matters more than it first appears.

A ranked list shows your page and lets the user decide. A language model reads a set of candidate documents, extracts the fragments that support the answer it is assembling, and attributes some of them. Your page can be retrieved and still contribute nothing citable if no passage in it stands alone.

This is why the practical question changes from "does this page deserve position three" to "is there a paragraph here that a model can lift, attribute, and be confident is correct."

What does Google itself say you need to do differently?

Very little, and this is the most under-quoted fact in the entire field. Google’s documentation on AI features states plainly: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

The same page closes off a popular upsell, saying you do not need to create new machine readable files, AI text files, or markup, and adding that "there’s also no special schema.org structured data that you need to add."

What actually transfers from classic SEO, and what doesn’t?

Most of the foundation transfers untouched. Crawlability, indexation, server-rendered text, matching intent, and genuine subject knowledge are prerequisites in both worlds. A page a crawler cannot read is a page a model cannot cite.

The row that trips people up is the last one. LLM SEO is not a replacement discipline. It is a layer that only pays off once the classic foundation is already in place, which is the crux of how GEO and SEO differ in practice.

Why is the passage, not the page, the unit of optimization?

Because that is the granularity at which both retrieval systems and Google’s own ranking operate. Google describes passage ranking as "an AI system we use to identify individual sections or ‘passages’ of a web page", used to judge how relevant that page is to a search.

Retrieval pipelines behind AI answers work the same way. Documents are split into chunks, chunks are embedded and matched against the query, and the model sees the winning chunks, not your page. Any paragraph that depends on the three paragraphs above it to make sense will arrive at the model stripped of that context.

The test is mechanical: cut any paragraph out of the page, show it to someone cold, and see whether it still answers a question completely and attributes its claims.

Ranking well is still important in the age of AI search. Many people
wonder whether their rankings still matter now that AI answers are
everywhere. The truth is that there is definitely still a relationship,
and you should not neglect your rankings, but things have shifted quite
a bit over the past year and it is worth understanding why.

## Do you still need to rank in the top 10 to be cited?

No. A top 10 ranking helps but is no longer a precondition. In an Ahrefs
analysis of 863,000 SERPs and 4 million AI Overview URLs published in
March 2026, 37.9% of URLs cited in AI Overviews also appeared within the
first 10 results, down from roughly 76% in July 2025. A further 31.0%
did not rank in the top 100 at all.

What is query fan-out, and does it really change keyword strategy?

Query fan-out is real and documented by Google. Describing AI Mode, Google says it "uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf."

The mechanism is confirmed. The popular inference drawn from it is not, and this is where honesty is worth more than a tidy narrative.

The inference goes: because engines fan out into sub-queries, pages that rank for those sub-queries get cited disproportionately. It is plausible, it matches the observed drop in top 10 overlap, and it is repeated constantly. But when Ahrefs tested it directly in July 2025, the data pointed the other way: pages cited from beyond the top 10 ranked for fewer keywords on average (887 versus 1,020) and for shorter queries (7.7 versus 8.5 words) than pages cited from the top 10. Their own conclusion is that the theory "doesn’t neatly fit" and remains unconfirmed.

You should also know that studies measuring adjacent things produce different headlines. Originality.AI reports that 52% of organic citations come from top 10 results, because they measure only citations that appear somewhere in the top 100 at all. Both numbers can be true. They answer different questions.

Do brand mentions matter more than backlinks?

The evidence leans that way for AI citation specifically, and the caveats are heavy enough that you should carry them with the claim.

In a study of 75,000 brands published in December 2025, the strongest correlation with AI mentions was YouTube mentions at roughly 0.737 (Spearman). Branded web mentions ranged from 0.656 to 0.709 across ChatGPT, AI Mode and AI Overviews. Domain Rating, the classic authority proxy, sat far lower at 0.266 to 0.326.

Now the caveats, all of which matter. This is correlational, not causal. It was produced by a company that sells SEO software. It is confounded by brand size, since large brands accumulate both mentions and citations for the same underlying reason. And the sample was pre-filtered to domains with Domain Rating above 40, which removes much of the variance in the link metrics being compared. The study’s own authors write that "correlation isn’t causation."

The reasonable takeaway is directional: unlinked brand presence appears to matter more for AI citation than it ever did for rankings. That is a reason to invest in being talked about, not a reason to stop earning links, which still drive the Google rankings that feed a large share of citations.

What formatting and structure actually help extraction?

Here the strongest evidence is a peer-reviewed paper rather than a vendor study. Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024), tested content edits and measured how visible each source became in generated answers.

The methods that won were adding statistics, adding direct quotations from credible sources, and citing sources explicitly, along with straightforward fluency improvements. Against an unoptimized baseline scoring 19.3 on their position-adjusted word count metric, quotation addition reached 27.2, statistics addition 25.2, and citing sources 24.6.

Read that with the methodology in view. The main experiments ran against a generative engine the authors built themselves using GPT-3.5-turbo, not a live commercial product. Their real-engine check on Perplexity covered only 200 samples, where quotation addition improved by 22%. Whether 2023-era numbers transfer to current multi-stage retrieval is genuinely unknown, and the paper also finds that effectiveness "varies across domains."

The direction is trustworthy even where the magnitudes are not. Concrete numbers, named sources, real quotations and clear prose make a passage more extractable. That is also just good writing, which is a useful sign you are not chasing an artifact.

What does not work?

• Keyword stuffing. It was the worst performer in the GEO paper, scoring 17.7 against a 19.3 baseline, meaning it left content less visible than doing nothing. On Perplexity it performed 10% worse than baseline. The oldest reflex in SEO is now actively counterproductive here.
• Thin, mass-produced AI content. Google’s spam policies name it directly, defining scaled content abuse to include "using generative AI tools or other similar tools to generate many pages without adding value for users". The policy applies "no matter how it’s created", so the question is never whether AI touched the page. It is whether the page adds anything.
• Schema as a citation lever. Structured data earns rich results and removes ambiguity for machines, both good reasons to use it. But Google explicitly says no special schema is required for AI features, and there is no solid public evidence that it lifts citation rates. Treat it as hygiene, not as a growth tactic.

How do you measure any of this?

Awkwardly, which is the honest answer. Google folds AI feature traffic into existing reports: sites appearing in AI features are included in the overall search traffic in Search Console under the Web search type, with no separate breakdown. ChatGPT, Perplexity, Gemini and Claude give publishers no analytics at all.

That leaves three things you can actually track. Referral traffic from AI domains in your analytics, which captures clicks but misses the majority of impressions where the user never leaves the answer. Citation share, measured by repeatedly asking each engine the queries that matter to you and recording who gets named. And classic rankings, which still correlate meaningfully with citation even though they no longer determine it.

Citation share is the only one that measures the thing you actually care about, and it has to be sampled over time, because answers vary between runs for the same prompt. A single spot check tells you almost nothing.

This is the loop Optimization for AI is built to close: it tracks how often a domain is cited across ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude alongside Google organic rank, identifies the queries where competitors are named and you are not, and re-measures after you publish. Plans start free, and you can see aggregate movement on the public AI visibility leaderboard before committing to anything.

Where to start

1. Fix indexability first. A page a crawler cannot read is a page a model cannot cite.
2. Rewrite your highest-intent pages passage by passage, so each section answers one question completely and attributes its claims.
3. Expand coverage into the sub-questions around that intent.
4. Measure citation share before and after, so you know whether any of it worked.

None of this is exotic. The uncomfortable part of LLM SEO is that the tactics with real evidence behind them are unglamorous, and the exciting mechanics are mostly unproven.