An LLM SEO agency that measures it
Large language models pick sources in two completely different ways, and most advice in this category confuses them. We work the mechanism that is actually reachable, per model, and track your brand prompt by prompt so the work is checkable.
LLM SEO is optimising for how large language models choose what to say about a brand. There are two mechanisms and only one is reachable: what a model absorbed in training is fixed until the next model, while what it retrieves live at answer time is decided by crawlability, page structure and third party corroboration you can change this quarter.
- Models tracked
- ChatGPT, Claude
- Reachable mechanism
- Retrieval
- Reported as
- A rate
- From
- $99/mo
plus Gemini and Perplexity, per prompt
training data is fixed until the next model ships
not a screenshot, because answers vary run to run
month to month, no minimum term
The distinction most LLM SEO advice skips
Almost every guide in this category treats a language model as one system. It is two, they behave differently, and knowing which one you are talking to decides whether a tactic can possibly work.
Training data: fixed, and not for sale
What a model absorbed during training is set until the next model is trained. No agency can edit it, and nobody can tell you your brand was added. Work aimed here pays off on a schedule set by somebody else, if at all, which is why we do not price it as a deliverable.
Retrieval: live, and reachable
When a model searches at answer time it fetches pages then and there. That path runs through robots.txt, through whether the page renders without JavaScript, and through which third party sources the system already trusts. All three are things you can change this quarter.
Each model behaves differently
ChatGPT search leans on Bing’s index, which is why a site absent from Bing can be invisible in ChatGPT while ranking fine on Google. Claude, Gemini and Perplexity each fetch and cite differently again. One optimisation does not cover all four, and pretending it does is how a report ends up unreproducible.
What we do at the model level
Everything below targets retrieval, because retrieval is the half that answers to work. Where a tactic only plausibly touches training data, we say so rather than billing for it.
Per model access, checked bot by bot
GPTBot, OAI SearchBot, ClaudeBot, Claude SearchBot, PerplexityBot and Google Extended are separate agents with separate rules, and sites commonly allow one and refuse another without realising. We check each and open the ones you decide to open.
Bing index health, deliberately
Because ChatGPT search reads Bing, being absent or thin there is an LLM visibility problem rather than a footnote. Most SEO programmes have never looked at it. We verify indexation, submit through IndexNow and fix what is blocking it.
Retrieval ready page structure
Content in the server response rather than after hydration, answers stated in the opening, facts near the claim they support. A retrieved page gets read in fragments, so the fragment has to make sense alone.
Entity disambiguation
Organization schema and a sameAs set covering the profiles you own, so the model can resolve that the company in a third party source is you. This matters most for brands with a common name or a rebrand behind them, which is more of them than you would think.
Corroboration on sources models fetch
Directories, review platforms, community answers and press. A model retrieving live weighs a page it did not fetch from you far more heavily than one it did, which is why the fastest movement usually comes from somewhere other than your own site.
Prompt level tracking, monthly
Your real buying questions, asked of each model on a fixed cadence, recording whether you are named, who is named instead, and which sources the answer cited. Same questions every month, so the comparison holds.
How we run an LLM SEO engagement
The order is deliberate: the cheapest and most certain work first, so you find out early whether the problem was ever a content problem.
- 01
Establish the baseline per model
Prompt set run across ChatGPT, Claude, Gemini and Perplexity, plus the full audit on your pages. Results usually differ sharply by model, and that difference is the first thing worth knowing.
- 02
Fix retrieval, model by model
Crawler access, rendering, Bing indexation, sitemap and schema. Measured before and after on the same checks, which makes this the part of the engagement you can verify without trusting us.
- 03
Make the content survive being fragmented
Rewrite openings, attach sources to claims, name an author, keep dates honest. A retrieved page is read in pieces and a piece that only makes sense in context does not get quoted.
- 04
Build corroboration, then rerun
Listings, reviews, community presence and PR, then the same prompt set again. Reported as a rate across the set with the work of that period beside it, and where nothing moved, that is what the report says.
Why we report a rate and not a screenshot
Ask ChatGPT the same question three times and you can get three answers citing three different sets of sources. That is not a fault, it is how sampling works, and it means a single screenshot proves nothing in either direction. Every vendor screenshot you have been shown, including a flattering one about us, is one sample.
So the unit is the prompt set and the output is a rate: of N questions asked of this model this month, you were named in M. Run the same set next month and the two numbers are comparable. It is a duller artefact than a screenshot and it is the only one that holds up when somebody checks it.
The deterministic half is reported separately and plainly: crawler access per bot, render output, Bing indexation, schema validity, audit score. These do not vary between runs. If a month produced no citation movement but fixed four blockers, the report says exactly that rather than dressing it up.
If you want the vocabulary sorted out first, the short version is that LLM SEO, GEO and AEO are three overlapping names for related work: LLM SEO names the mechanism, GEO names the outcome, AEO names the format. Nobody is going to settle which term wins, and we sell the same engagement under all three.
Check what each model can actually reach
Retrieval is the reachable mechanism, and these three read exactly the things retrieval depends on. A minute each, on any URL.
Or run the same interrogation by hand: how LLM SEO works, ranking your business in ChatGPT, what ChatGPT search actually ranks. Free, no account, and the honest version of the do it yourself option.
What buyers ask about LLM SEO
What is LLM SEO?
Optimising for how large language models choose what to say about a brand. It splits into two mechanisms: training data, which is fixed until the next model is trained and which nobody can edit on request, and live retrieval, which runs through crawler access, page rendering and third party sources and can be changed now. Serious LLM SEO works the second one. Our full explainer covers both routes and why four assistants disagree about the same brand.
Is LLM SEO the same as GEO?
They describe the same engagement from different angles. LLM SEO names the mechanism, the model and how it selects sources. GEO names the outcome, being generated into the answer. AEO names the format, being the extracted answer. We sell one piece of work under all three names because buyers genuinely search for all three.
Can you get our brand into a model’s training data?
No, and be careful with anyone who says they can. Training happens on a schedule set by the model vendor, from a corpus nobody outside publishes, and there is no submission route. What broad public presence does is improve the odds for a future model, which is a reason to build corroboration anyway, not a deliverable with a date on it.
Why does Bing matter for ChatGPT?
Because ChatGPT search draws on Bing’s index. A site that ranks well on Google and is thin or absent in Bing can be invisible in ChatGPT for exactly that reason, and it is a blind spot in most SEO programmes because nobody has looked at Bing in years. Checking and fixing it is usually a fast win.
How is progress reported?
As a rate across a fixed prompt set per model, plus the deterministic audit numbers. Generative answers vary between runs, so a screenshot is a single sample and we do not treat one as evidence. Same questions, same models, same cadence, and where nothing moved the report says nothing moved.
Do you work with our existing SEO team?
Regularly, and it is the arrangement that tends to work best. They keep strategy and content; we take crawler access, rendering, entity work and the per model measurement, which is the part that needs an engine rather than an opinion. Everything we change is recorded against the engagement.
What does an LLM SEO agency charge?
Published entry retainers in this category run from $2,500 to $50,000 or more a month, and almost none of those figures come from the agency’s own site. Ours are $99, $299 and $799 a month, month to month. The named comparison with sources is on what GEO costs.
Find out which models can reach you
The crawler check and the render check together explain most cases of a brand being absent from one model and present in another. Both are free, both take about a minute, and neither needs an account for the first run.