Short answer. LLM SEO is the practice of making your content easy for AI assistants like ChatGPT, Perplexity, and Google’s AI Overviews to find, quote, and cite. It shares its foundations with classic SEO (crawlable, useful, well-structured pages) but is judged on citations and brand mentions, not rankings or clicks.
Search for this term and most of what comes back is one of two things: a Reddit thread asking what LLM SEO even means, or a checklist that promises AI visibility if you add schema and write in short paragraphs. Neither tells you the two things a marketer with a real budget needs to know. What is actually different from the SEO you already do, and how would you know if any of it worked?
The second question is the one nobody answers, and it is the one this post spends the most time on. Vercel put it plainly when it wrote about its own approach: “Measuring visibility in AI systems is still an evolving challenge. There’s no reliable dashboard showing if your content appears in answers or is embedded in training data.” That is the honest state of the field. Anyone selling certainty is selling something else.
What LLM SEO is, and what it is not
The name is a bit of a mess. People also say GEO (generative engine optimization), AEO (answer engine optimization), and AI SEO, and they mostly mean the same thing: work that increases the odds an AI system names your brand or cites your page when someone asks a relevant question.
It helps to separate three things that get lumped together:
| Goal | What you are trying to influence | How you see it |
|---|---|---|
| Ranking | Your position in a list of blue links | Rank trackers, Search Console |
| Citation | Whether an AI answer links to or quotes your page | Prompt sampling, referral traffic |
| Mention | Whether an AI answer names your brand, with or without a link | Prompt sampling, branded search lift |
Classic SEO optimizes the first. LLM SEO optimizes the second and third. They overlap, because assistants that search the web lean on the same indexes and the same pages, but they are not the same scoreboard. A page can rank fifth and get cited constantly, or rank first and never appear in an answer.
It is also not a replacement for SEO. If your pages cannot be crawled, indexed, and understood by a machine, nothing else on this list matters. LLM SEO sits on top of that foundation.
What actually moves citations
Nobody outside the labs knows the exact retrieval and citation logic, and anyone claiming a formula is guessing. What follows is the set of practices that are consistent with how these systems are described to work, and that cost you nothing if they turn out to matter less than advertised. They are also just good content practice.
Answer the question first. Assistants pull passages, not pages. A section that opens with a direct, self-contained answer is easy to lift. A section that builds to its point over four paragraphs is not. Put the answer in the first one or two sentences under the heading, then explain.
Make headings match real questions. People prompt assistants in full sentences. Headings phrased as the question a buyer would ask (“Does server-side tracking work with Safari?”) map neatly onto those prompts and give a retrieval system a clean label for the passage below.
Be specific and quotable. Vague claims are not worth citing. A precise, sourced statement is: a number, a named mechanism, a date. Original data is the strongest version, because it exists nowhere else and an assistant has no substitute to cite instead.
Keep the page readable by a machine. Serve the content as HTML, not behind a script that only renders in a browser. Use real headings, real tables, and descriptive link text. Structured data such as FAQPage markup is cheap to add and describes your content explicitly, though nobody can promise it changes a citation outcome.
Let the crawlers in on purpose. Check your robots rules, and know that the AI companies run separate bots for search and for training. OpenAI’s OAI-SearchBot, Anthropic’s Claude-SearchBot and PerplexityBot fetch pages for answers, so blocking them opts you out of being cited. GPTBot and ClaudeBot collect training data, and blocking those is a legitimate choice as long as you made it deliberately. Some teams block all of them by accident through a blanket bot rule.
Keep pages fresh and consistent. Vercel recommends a set refresh cadence, reviewing content at 30, 90 and 180 days. The related point is consistency: if your pricing page, your docs, and a third-party review all describe your product differently, an assistant synthesizing an answer has to pick a version. Make the true one the loudest.
Earn mentions off your own site. Assistants form an impression of a brand from many sources, including comparison articles, forum threads, and review sites. Being described accurately in the places buyers already read is part of LLM SEO, and it is the part you control least. Treat it as PR and community work, not markup.
None of this is exotic. If that feels underwhelming, good. The teams that struggle here are usually the ones hoping for a trick.
Why this is worth doing at all
The case rests on evidence that is real but thinner than the headlines suggest, so here it is at its true size.
In June 2025, Vercel reported that ChatGPT referred around 10 percent of its new signups, up from 4.8 percent the previous month and 1 percent six months earlier. Vercel also cited Tally, which reported AI search becoming its biggest acquisition channel. Those are striking numbers, and they come from developer-tool companies whose buyers already live inside AI assistants. Your audience may look different.
The other half of the argument is defensive. Vercel’s write-up cites an Ahrefs study that found AI Overviews correlated with a 34.5 percent lower click-through rate for the top-ranking page. Ahrefs re-ran the study in February 2026 and put the drop at 58 percent. If anything like that holds for your queries, then ranking without being cited in the Overview is a shrinking prize.
This very term shows the effect. In Ahrefs (US, September 2026), “llm seo” has about 1,600 searches a month, and the estimated traffic potential of the top-ranking page is roughly 10 visits. The volume is real and the clicks are not, because the AI Overview absorbs the query. If you are writing about this topic for traffic, the numbers say do not. If you are writing to be the source the Overview cites, that is a different bet, and it is the one this post is making.
The measurement problem
Here is where most LLM SEO advice quietly stops. You have made your pages easier to cite. Now what do you count?
There are three signals, and each has a hole in it.
Prompt sampling. You run a fixed list of buyer questions through the assistants on a schedule and record whether you are cited or mentioned. This is the closest thing to a rank tracker, and it is useful. It is also a sample of a system that answers differently from run to run, so treat single results as noise and look at trends over weeks.
Referral traffic. When a human clicks a citation, you can sometimes see it. ChatGPT has tagged citation links with utm_source=chatgpt.com since its search feature launched in late 2024, and extended the tag to its “More” source links in June 2025. Perplexity and Gemini generally send a referrer on the web. Mobile apps often strip the referrer, so clicks without a UTM tag land in Direct. What you see is a floor, and the gap between the floor and the truth is large. The LLM traffic attribution breakdown explains the three layers involved (crawlers, organic citations, paid clicks) and why mixing them corrupts every number. If you want the setup steps, the AI traffic attribution guide covers the ways to segment it, and the track ChatGPT referral traffic walkthrough shows the GA4 configuration.
Brand demand. The largest effect may be invisible. An assistant names you, the user clicks nothing, and searches for your brand an hour later. That visit files under branded search. Watching branded search and direct traffic move alongside your prompt-sampling results is the only way to see it, and it is correlation, not proof.
The uncomfortable conclusion is that you will be making decisions on partial data. That is fine. It just means you should choose your metrics before you start, write down what you expect, and resist declaring victory on a single good week.
What our own numbers say
Since this post is about measurement, here is ours, unpolished.
Convultra tracks AI referrals on its own site. Over the 90 days to September 29, 2026, our project recorded 11 sessions from AI assistants: 10 from ChatGPT and 1 from Claude. Together they produced 16 pageviews and zero conversions.
That is a small number, and we are not going to dress it up. Two honest readings follow. First, at this stage of a young domain, AI assistants are not a meaningful acquisition channel for us, whatever the industry headlines say. Second, and just as important, it is a floor: app clicks that arrive with no referrer and no tag are not in it, and neither is the visitor who was told our name and typed it into a search box later. Eleven is the count we can prove, not the count that happened.
This is the reason we built the AI referral view into the product the way we did. It picks out the humans who clicked through from an AI assistant, keeps them inside your traffic, conversion and revenue figures rather than in a side report, and shows which assistant sent them. It cannot credit a click that arrives with nothing to say where it came from, and neither can anything else. If a vendor tells you it can, ask how.
If you want to understand where your own numbers sit, the AI attribution page shows what Convultra reports and what it deliberately does not claim.
A practical order of operations
If you are starting from zero, do these in sequence. Skipping to the fun part is how teams end up with a schema project and no idea whether it helped.
- Baseline first. Write 20 to 30 real questions a buyer would ask. Run them through ChatGPT, Perplexity, and Google today and record who gets cited. This is your before picture.
- Fix access. Confirm the search crawlers are allowed, key pages render as plain HTML, and your sitemap is current.
- Rewrite your top ten pages for extraction. Lead each section with the answer, use question-shaped headings, and replace vague claims with specific ones.
- Publish something citable. One original data point, benchmark, or definition that other pages will want to reference is worth more than ten reworded explainers.
- Instrument referrals. Segment AI referral traffic so you can see the floor, and keep the conversion tracking behind it accurate, because a visitor who arrives from an assistant and then converts through a blocked pixel is invisible twice.
- Re-run the baseline monthly. Compare trends, not single runs. Give it a full quarter before you judge.
Step five is where LLM SEO meets the rest of a paid-acquisition stack. AI referral visitors are still visitors who can be lost to ad blockers and Safari’s tracking limits, and undercounted conversions from any channel skew the case for the next budget decision.
FAQ
Is LLM SEO different from GEO?
In practice the terms are used interchangeably. GEO stands for generative engine optimization, AEO for answer engine optimization. All describe the work of being cited or mentioned in AI-generated answers.
Does LLM SEO replace traditional SEO?
No. Assistants that search the web draw on the same crawlable, indexed pages, so the technical and content fundamentals still apply. LLM SEO adds a second goal (citations and mentions) on top of rankings.
Does schema markup make ChatGPT cite me?
Nobody can promise that. Structured data such as FAQPage markup states your content explicitly and costs little to add, but there is no published evidence that it guarantees a citation.
Can I see how much traffic AI assistants send me?
Partly. ChatGPT adds utm_source=chatgpt.com to the links it cites and some assistants send a referrer, but mobile app clicks often arrive with no referrer. Treat the number you see as a floor.
How long does LLM SEO take to work?
There is no reliable timeline. Assistants refresh their sources at different rates, and results vary run to run. Measure a baseline, change one thing at a time, and judge over a quarter rather than a week.
Know what your AI traffic actually does
Being cited is only half the job. You also need to see which assistants send visitors, and whether those visitors convert. Convultra shows you both with one line of JavaScript. Start your free trial.