LLM SEO is the work of getting your brand, products and pages cited, mentioned or recommended by assistants such as ChatGPT, Perplexity, Gemini and Google's AI features. Most of it is classic SEO, however much the label tries to make it sound new. Google's own documentation says AI Overviews and AI Mode have no additional technical requirements beyond a page being indexed and eligible to show with a snippet. A smaller slice is genuinely new, and it sits around passages, entities, crawler access and measurement. This post separates the two, gives you a table you can drop into a planning doc, and ends with a 30-day starting plan. Nothing here guarantees you appear in an answer. Nobody can promise that.
What LLM SEO actually means
The cleanest definition I can offer is this: optimising to be retrieved, cited or mentioned by LLM-powered assistants and AI search features. Those three verbs are not interchangeable. Retrieved means your page gets pulled in while an answer is being built. Cited means you are linked as a source. Mentioned means the assistant names your brand, with or without a link.
LLM Pulse defines it as optimising your brand, products and content to be cited and recommended by large language models, and calls it the everyday name for the discipline also known as GEO and AEO. The same piece says GEO was popularised by a 2023 Princeton paper, and that AEO was used heavily in 2024 and 2025. Newsdata.io goes further and puts LLMO, GEO, AEO, LLM SEO and AI SEO into one bucket. The taxonomy is already untidy.
So the labels are interchangeable. Fine. Pick one and stop arguing about it. If you want the keyword-research angle on the naming mess, we compared them in AEO vs GEO vs SEO keyword research.
One caution before going further. Google's guidance on third-party SEO advice mentions AEO and GEO by name, and says some of that advice misinterprets or makes claims about what 'Google says'. Its recommendation is to check any advice against official guidance. That includes this post, so I'll be explicit about where each claim comes from.
How retrieval works in practice
An assistant can know about you in two broad ways. It can look you up while answering, or it can remember you from training. And those need different work, which is where a lot of vague LLM SEO advice starts to blur.
Search-grounded answers
Google describes retrieval-augmented generation (RAG), also called grounding, as a technique that relies on its core Search ranking systems to retrieve relevant, up-to-date pages from the Search index. The model then reads those pages and writes the answer, with links to supporting pages. Google's optimisation guide says its generative features are rooted in core Search ranking and quality systems. Its answer to 'is SEO still relevant?' is yes.
Then there is query fan-out. Google defines it as a set of concurrent, related queries generated by the model. Its example: a user asks 'how to fix a lawn that's full of weeds', and the system may also run searches like 'best herbicides for lawns', 'remove weeds without chemicals' and 'how to prevent weeds in lawn'. One question becomes several retrievals, not just one neater search.
Two more facts are worth keeping in mind. AI Overviews only show when Google's systems decide they add something beyond classic Search, so they often don't trigger. And OptimizeGEO describes ChatGPT-style real-time search as the same retrieve-then-answer layer, and claims it runs through Bing's index. That is the vendor's claim. I can't point you to an OpenAI statement confirming it, so treat Bing Webmaster Tools as cheap insurance rather than a proven lever.
Training-data recall
OptimizeGEO lays out three layers: training data (parametric memory), real-time web search via RAG, and agentic scraping. The first is the slow one. If a model has absorbed lots of consistent, positive references to your brand, it may name you without searching anything. You can't edit that after the fact. You can only influence what gets written about you over time, which is less glamorous but more honest.
Why one screenshot proves nothing
OptimizeGEO cites Ahrefs (November 2025) saying 45.5% of AI citations get replaced when the same AI answers the same question again. Secondhand, yes. But it matches what anyone sees who reruns a prompt. Answers wobble. So any measurement plan needs repeated sampling, not a single lucky screenshot.
The query shape changes too. LLM Pulse's comparison says classic search runs on short strings of 3 to 5 words, while LLM prompts are long natural-language questions of 15 to 50 words, answered with one synthesised response that sometimes cites 3 to 8 sources. That changes the unit of work from a keyword to a fuller question.
What is genuinely new
Strip away the hype and four jobs are actually different from what you did in 2022.
Passage-level citability
If an assistant fans a question out into sub-queries, it is hunting for pieces of pages that answer specific sub-questions. A page that only works as a whole, with each paragraph leaning on the one before, is harder to lift from. That 'passage-level' framing is my inference from how fan-out works, not a Google instruction. Google's guide does tell you to create unique, compelling, useful content, and says that will likely influence presence in generative search more than any other suggestion in the guide.
In practice that means:
- Open each section with the direct answer, then support it.
- Name the subject in the passage itself. 'It' and 'this approach' die when the passage is pulled out of context.
- Give each sub-question its own heading.
- Put important facts in text, not only in images or scripts.
None of that is exotic. It is just editing with extraction in mind.
Entity consistency across the web
Index Lab says five factors decide which brands ChatGPT, Gemini and Perplexity recommend: mention frequency, source authority, review sentiment, query fit and structured data. That is a vendor's framework, not a published ranking formula. Still, the first three all depend on what the wider web says about you, in your words and other people's. That makes brand consistency operational, not cosmetic.
So the job is an audit. Does your brand name, one-line description and category read the same on your site, your social profiles, directories, review sites and press mentions? Google's own list of recommended practices includes keeping Merchant Center and Business Profile information up to date, which is the Google-side version of the same hygiene.
AI crawler access decisions
This one is new because it is a decision, not a default. Classic SEO had one question: can Googlebot get in? Now you choose, bot by bot, who may fetch what. That is a policy choice as much as a technical one.
Google says robots.txt directives for Googlebot are the control for crawling for Search, and that nosnippet , data-nosnippet , max-snippet or noindex limit what is shown. Google-Extended is the control for limiting AI training and grounding in some of Google's other systems. Other vendors publish their own crawlers. OpenAI documents its on the OpenAI bots page, and you should read the current version there rather than copy a user-agent list from a blog post.
Also check your CDN and host. Google explicitly lists allowing crawling 'by any CDN or hosting infrastructure', and bot-protection rules are a common way to block crawlers by accident. If you are weighing an llms.txt file, OptimizeGEO lists it among its eight core strategies, and our post on llms.txt vs robots.txt covers what each file does and doesn't do. The boring access layer is where a lot of avoidable damage starts.
Mention tracking
Search Console won't tell you how often an assistant names you. Google says sites in AI features are included in overall Search traffic, reported under the 'Web' search type in the Performance report. That is Google's surfaces only, and it is not split out. So you get visibility, but not the slice most people now want.
LLM Pulse says measurement shifts to mention rate, citation rate, share of voice and sentiment. To get those you need a fixed set of prompts, rerun on a schedule, with answers logged. An early founder interview on Stage 2 Capital describes a team moving from ranking for 'What is procurement?' to prompts like 'What's the best procurement tool for charter schools?', and recommends prompt research alongside keyword research. Sensible. Our AI citation checker is one place to start sampling.
What has not changed
Most of the work, frankly. Here is Google's own list of what still matters for AI features, from the AI features documentation:
- Crawling allowed in robots.txt and by any CDN or hosting infrastructure.
- Internal links that make content easy to find.
- A good page experience.
- Important content available in text.
- High-quality images and videos where relevant.
- Structured data that matches the visible text.
- Up-to-date Merchant Center and Business Profile information.
Google's wording is that specific optimisation isn't required for AI Overviews and AI Mode, and that all existing SEO fundamentals continue to be worthwhile. To be a supporting link, a page must be indexed and snippet-eligible. That is the whole technical gate, and it is a narrower gate than the industry noise suggests.
Crawlability and helpful content
If the page can't be crawled, indexed and shown with a snippet, the rest is academic. And the content bar is the old one: unique, useful, written for people. Commodity content that repeats what ten other pages say gives a model no reason to pick yours.
Authority
The retrieval layer rides on ranking systems, and the recall layer rides on who mentions you. Both reward the same slow work: being referenced by sources people already trust. LLM SEO guides dress this up as 'source authority' and 'mention frequency'. It is links, PR and reputation under newer names. The label changed faster than the work did.
Structured data, where the sources disagree
This is the messiest part. Vendor guides tell you to deploy schema markup. OptimizeGEO puts it second on its list. Google says only that structured data should match the visible text. And Index Lab's July 2026 update reports that Google's guidance says no special schema.org markup is needed for AI features, and that a controlled Ahrefs study of 1,885 pages that added JSON-LD, matched against 4,000 control pages, found no meaningful citation uplift on any platform. A separate experiment, it says, showed AI systems read JSON-LD as plain text rather than parsing it.
My read: keep your schema accurate because it helps classic Search features and Google asks for it to match the page. Don't expect it to buy you citations. But don't rip it out either, because accuracy is still part of the job.
Which tactics are new and which are relabelled
This is my own sorting, built from the sources above. It is a judgement call, not a test result, and the rows on each side are not meant to pair up. Treat it as a planning frame, not a law of nature.
| Genuinely new for LLM answers | Classic SEO work with a new label |
|---|---|
| Writing self-contained passages that answer fan-out sub-questions | Clear headings and answer-first writing |
| Keeping brand name, description and category consistent across third-party sites | Keeping business details accurate (Business Profile, Merchant Center) |
| Deciding bot by bot which AI crawlers may fetch your pages | Allowing crawling in robots.txt and through your CDN |
| Tracking mention rate, citation rate, share of voice and sentiment on a prompt set | Rank tracking and Search Console reporting |
| Researching full-sentence prompts next to keywords | Keyword and search-intent research |
| Resampling answers repeatedly because citations churn | Watching rankings move over time |
| (nothing here) | Fast, crawlable, indexable HTML with text content |
| (nothing here) | Unique, helpful content and earned authority |
| (nothing here) | Structured data that matches visible text |
Count the empty cells. That tells you where the weight sits.
A 30-day starting plan
Thirty days gets you a baseline and a first round of fixes. It will not settle anything. Google notes that crawling can take anywhere from several days to several months before changes to preview controls are processed, so judge direction, not verdicts. That restraint matters when early graphs start looking persuasive.
- Days 1 to 7: check the gate. Confirm your priority pages are indexed and snippet-eligible. Read your robots.txt and your CDN or firewall rules. Use the URL Inspection tool to see what Googlebot actually receives. Decide, in writing, which AI crawlers you allow, checking each vendor's own documentation, including the OpenAI bots page.
- Days 8 to 14: build the prompt set. Write full-sentence prompts the way a buyer would ask them, including comparison and 'best for' questions. Run each more than once, because answers change between runs. Log whether you were mentioned, cited or absent. Record sentiment too.
- Days 15 to 22: fix passages and entities. Take your five or so most valuable pages. Add direct answers under each heading, name the subject inside the passage, and cut anything that only makes sense in context. In parallel, audit how your brand is described on external profiles and align them.
- Days 23 to 30: resample and decide. Rerun the prompt set. Compare against the baseline, remembering the churn. Pick two or three changes worth continuing and drop the rest.
If you want help running this, here is how we approach AI search work. For the ChatGPT side of it, our notes on getting cited by ChatGPT search go deeper. Start there if the immediate problem is citations rather than the whole operating model.
What I would not do in these 30 days: chase every vendor's eight-point list, or rebuild your schema because a blog said so. Triage first, because activity is cheap and clarity is not.
FAQ
Will AI Overview traffic show up separately in Search Console?
No. Google says sites appearing in AI features are included in overall Search traffic and reported in the Performance report under the 'Web' search type. You can't isolate it with a filter. Google also says it has seen clicks from results pages with AI Overviews are higher quality, meaning users are more likely to spend more time on the site, so look at engagement and conversions in your analytics as well as clicks.
Does blocking Google-Extended take me out of AI Overviews?
Google's documentation draws the line differently. It says robots.txt directives for Googlebot are the control for crawling for Search, and that nosnippet , data-nosnippet , max-snippet or noindex limit what is shown. Google-Extended is described as the control for limiting AI training and grounding in some of Google's other systems. Test any change rather than assuming, and allow time for recrawling. But don't pretend one directive governs every use case.
Should I worry about the zero-click statistics going around?
Treat them with suspicion. TripleDart cites 60% of searches ending without a click and a Gartner forecast of a 25% decline in traditional search by 2026. Those are vendor-cited, secondhand figures, and I haven't traced them to an original study. Use them as colour, not as a basis for budget.
How long before changes show up?
For Google, slower than you want. The company says recrawling can take anywhere from several days to several months, depending on how often its systems decide a page needs refreshing. Other assistants publish no such timeline in the sources here, so measure on a schedule and look for direction over weeks. One prompt run is not a trend.
The sharpest fact in all of this is Google's own: no additional technical requirements for AI Overviews and AI Mode beyond being indexed and snippet-eligible, and no tactic here guarantees you a place in any answer. Do the old work properly first, then add the few new jobs: passages, entities, crawler decisions and prompt-level tracking. That is less exciting than a new playbook, but it is closer to the truth.