LLM SEO (SEO for AI): what changes, what stays the same, and what to do first

LLM SEO is the work of getting your website and your brand into the answers of ChatGPT, Claude, Gemini, Meta AI and Google’s AI Overviews, as well as into the classic results. People search for it as “LLM SEO”, “AI SEO”, “SEO for AI” or “SEO for ChatGPT”, and the industry also calls it GEO or AEO. They are different names for the same field; the nuances are in AEO vs SEO vs GEO.

The underlying question is usually whether traditional SEO is dead. It isn’t. Google says in its guide to generative AI features that, from Google Search’s perspective, optimizing for generative AI search is still SEO, and that the only requirement to appear in those features is that the page be indexed and eligible to be shown with a snippet. What changed is the outcome: measuring a ranking is no longer enough, because the answer is a paragraph that names some brands and cites some sources.

What stays the same is almost everything you were already doing well: indexing, internal links, clear titles, good page experience, original content. If your SEO has foundational problems, fix them first; no AI tactic makes up for a page that isn’t indexed.

The first step is knowing which crawler does what, because blocking one is not the same as blocking another. The main ones:

robots.txt isn’t the only door. A CDN or web application firewall (Cloudflare, Imperva and others) can block or challenge AI crawlers even when your robots.txt allows them. And a JavaScript challenge usually responds with a 200 status code, so your logs look fine while the bot receives a verification page instead of your content. No robots.txt check can see that: you have to request the page the way the bot does and compare it with what a browser gets. Keep in mind that some CDNs let the real, verified bot through by IP range and stop anyone who only copies its user agent, so the final confirmation is in your CDN logs.

Content has to be in the server’s HTML. Several AI crawlers read the HTML as it arrives and don’t run JavaScript. If your prices, products or descriptions load in the browser, the page is empty to them. Googlebot does render JavaScript, but it’s better not to rely on that for the others. The same goes for structured data: if a script or a client-side tag manager injects it, a bot that doesn’t run JavaScript won’t see it. Put it in the HTML, and make sure it matches what’s visible on the page.

Indexing in both indexes. Check Google Search Console and Bing Webmaster Tools. A page marked “Crawled - currently not indexed” doesn’t exist for AI Overviews, and a page Bing doesn’t have gives up one of the ways ChatGPT search can find it. IndexNow lets you tell Bing the moment you publish or change something.

Engines with search don’t run one query but several. Google calls this query fan-out: the model generates related, simultaneous queries to complete the answer. A question about car insurance in Guadalajara, Mexico can trigger separate searches for coverage, prices in Mexican pesos, requirements and reviews. Your content competes on each of those sub-questions, not just on the main phrase. What works:

• Answer first. The first sentence of each section states the point, with the fact. What follows explains it. A model summarizing your page takes what’s at the top.
• One page per group of real questions, not per variant. Google warns that creating content for every possible way of searching, in order to manipulate rankings, violates its scaled content abuse policy.
• Local facts, in writing. Country, cities, currency, timelines, coverage, requirements. If your page says “nationwide shipping” without saying which nation, the engine can’t use it for a question that names Mexico or Spain.
• Honest comparisons. A page that explains when your option makes sense and when another one is better is easier to cite for an “X or Y?” question than a brochure.
• First-hand experience. Google asks for original, useful content with first-hand experience. Your own data, cases and real prices: what isn’t on any other site.

What you don’t need, according to the same Google guide: breaking content into tiny chunks, writing in a special way for AI, or creating AI text files or special markup; Google says no llms.txt file is needed either.

This is the biggest change from classic SEO. In Refractia’s data, a brand’s own website accounts for between 0.5% and 2.8% of what AI cites when it talks about that brand. The rest is media, comparison pages, reviews, directories, forums, marketplaces and industry sites. For AI, an unlinked mention in an article the engine reads is still information about your brand; for classic SEO, it barely counted.

• Look at which domains the engines cite in your category and start there. Not every outlet carries the same weight for every question.
• Separate two things: whether the source mentions you and whether it links to you. An article about your category that doesn’t name you is a concrete opportunity.
• Sort out your entity: the same name, description and facts on your site, your profiles, your Google Business Profile and, where it applies, Wikidata. A knowledge panel that matches your brand helps the engine avoid confusing you with someone else.
• No bought mentions. Google says chasing inauthentic mentions doesn’t help as much as it may seem, and that its other systems block spam.

In Refractia, Influence Sources shows which domains the engines cite when they talk about your brand, separates your own site from earned media, and marks in a matrix whether each source mentions you and whether it links to you. Identity & DNA reviews entity signals such as the Wikidata item and the Google knowledge panel.

Search Console folds the traffic from Google’s AI features into the rest, and it doesn’t tell you what the answer says about your brand, whom it recommends instead of you, or how you do in ChatGPT or Meta AI. For that you have to ask the engines, under these conditions:

1. Category, comparison and purchase questions, written in the market’s language with the country named in full.
2. Several engines: ChatGPT, Claude, Gemini, Google’s AI Overviews (queried from each country) and Meta AI.
3. Every rate with its sample, “k of n”, and an engine that didn’t answer counted as “not measured”, not as zero.
4. Chat and AI search reported separately: what the model knows and what the web ranks move with different levers.
5. Beyond whether you’re named, how you’re described: sentiment, adjectives, attributes.

That is what Refractia does. On the site side, GEO Optimization calculates a GEO Score for every page from six pillars (understanding, retrieval, citation, recommendation, semantic coverage and coherence across engines), analyzes up to 500 pages in full-site mode and builds a content plan. On the answers side, it measures the five surfaces with the rules above. A person starts each scan, and we don’t measure Perplexity, Copilot or Grok.