A language model answers a buyer's question from two places: what it learned during training, and what it fetches while answering. Your brand can be weak in either, and the fix is different for each. Most LLM SEO offers do not say which problem they are solving.
This page explains how the work is split, what we do on each part, what it costs, and what we cannot promise. It is written for founders and marketing leads who want to know what they are buying before they ask for a quote.
01What LLM SEO is, and what else it is called
LLM SEO is search optimization aimed at large language models. The same work goes by several names: generative engine optimization, answer engine optimization, AI SEO and LLM optimization. The names overlap heavily, and our guide on GEO versus AEO explains the small differences. What matters is the goal: your brand appears, accurately described, when someone asks a language model about your category.
It is not a separate discipline from SEO. Google says AI Overviews and AI Mode use its core ranking and quality systems, and that you need no new files or markup to appear in them. Language models that fetch pages live use search indexes and crawlers too. So the base is the same: pages that can be crawled, indexed and read. LLM SEO adds testing the answers, and working on what other sources say about you.
02How a model gets its answer: two paths
OpenAI documents separate crawlers for separate jobs. OAI-SearchBot is for surfacing sites in ChatGPT search. GPTBot collects content that may be used to train models. ChatGPT-User visits a page when a person asks ChatGPT to look at it. That split shows the two paths clearly.
| Path | What it uses | What it means for you |
|---|---|---|
| Live retrieval | Pages fetched or searched while the question is answered | Crawlable, indexed, clearly written pages can be quoted now |
| Learned memory | What the model absorbed from its training material | Slow to change; repeated, consistent mentions across many sources matter most |
| Both together | Retrieved pages combined with what the model already believes | A good page can be outweighed by old facts repeated elsewhere |
Based on OpenAI's published crawler documentation and Google's description of retrieval in AI features.
This gives you a quick diagnosis. If a fix to your own page changes the answer within days, you were fighting a retrieval problem. If the answer will not move after the page is right, other sources are probably repeating the old version, and the work is outside your own site.
03What our LLM SEO work covers
Question map
We list the questions buyers ask models in your category, in six types: branded, category, best-for, comparison, pricing and use-case.
Access and eligibility
We check that the right crawlers can reach your pages and that each page is indexed and eligible to show a snippet, which Google lists as the only technical requirement for its AI features.
Answer pages
Where a question has no page that answers it, we write or rewrite one: the answer first, then evidence, then detail.
Fact alignment
We list every place your price, product and location appear, find the conflicts and fix the ones that matter.
Outside sources
We find which third-party pages the models cite in your category and plan how you are described there honestly.
Re-test
The same questions are run again on three engines, with screenshots, and compared with the baseline as a range.
04Use fan-out to check coverage, not to make more pages
Google describes query fan-out as the set of related searches an AI feature runs at once to build one answer. Its guidance gives no tactic for it and warns that making separate pages for query variations to manipulate results breaches its scaled content abuse policy. We use fan-out for one thing: checking that every sub-question has a source somewhere, which might be a section of an existing page rather than a new page.
05A question map you can copy
Use this table to start your own. Fill one row per buyer question, then test each on three engines.
| Question type | Example wording | Page that should answer it | Named in test? |
|---|---|---|---|
| Branded | What is [brand] and who is it for? | Homepage or about page | Yes, no, wrong |
| Category | What is the best [category] for [type of buyer]? | Category or use-case page | Yes, no, wrong |
| Comparison | [Brand] vs [competitor]: which is better for [use]? | Comparison page | Yes, no, wrong |
| Pricing | How much does [category] cost? | Pricing page | Yes, no, wrong |
| Use-case | How do you [task] with [category]? | Guide or help page | Yes, no, wrong |
Keep the wording fixed once you start. If you change a question after seeing a bad answer, you are no longer measuring the same thing, and the before and after cannot be compared. Take the questions from real sources: sales calls, support emails and the People Also Ask boxes in Google. Questions invented at a desk tend to sound like your marketing, while buyers type something plainer and shorter.
06What we will not do
We do not claim that an llms.txt file gets you cited. Google states that it does not use such files for Search and that they neither help nor harm. We build one on request with the llms.txt generator, and treat it as optional, cheap and unproven. We do not buy mentions, which Google says are not helpful when they are inauthentic. We do not promise citations or a date. Google says no one can guarantee a number one ranking, and models vary between sessions.
What an LLM SEO offer often includes
- Prompt tracking with a score nobody explains
- A new page for every phrasing of a question
- Special files presented as the answer
- A promised citation window
What we do
- Named questions, repeated runs and screenshots
- One page per real question, or a section where that is enough
- Access, facts and outside sources first
- A range from repeated tests, never a promise
07How we know whether it is working
We do not use a single visibility score, because tools define scores differently and two dashboards can disagree about the same brand. We count four things on the question map, each run several times per engine: how often you are named, how often a page on your site is cited, how often the description is correct, and how much those counts change from run to run. The last one matters most early, because a gain smaller than the normal swing between runs is not a gain.
The measurement method is set out in how to measure AI visibility, and the case studies page explains how we report results and what a case can prove. We also tell you when a gap sits outside your own site, for example when a comparison site keeps being cited for your category and the only way to change it is to be described there.
08Who it suits, and the price
This suits you if
- Your buyers ask AI tools to recommend or compare products like yours
- You have pages but they never appear in answers
- You want a measured baseline and repeat tests
- You can approve site changes
Skip it if
- You want guaranteed citations
- You want hundreds of pages
- You sell a low-priced product where tools are enough
- You have not yet got basic pages indexed
The work is $2,000 a month, with the inclusions and exclusions on the pricing page. If you are unsure where to begin, the AI visibility audit shows which of the paths above your gap sits on. The broader programme is described under generative engine optimization services, and the AI visibility hub covers the whole topic. For ChatGPT alone, see ChatGPT SEO services.

