LLM SEO: a working guide, from first measurement to proof it worked

LLM SEO is the work of getting a brand named in the answers that AI assistants write, and of measuring whether that happens. The measurable part is a rate: how often ChatGPT, Gemini or Perplexity names you when asked a question your buyer would ask, with your brand name kept out of the question.

That second sentence is where this series puts its weight. Measurement is the part of this subject that can be checked, repeated and argued with. The advice about what to change afterwards is thinner than the volume of writing about it suggests, and a later section on this page says exactly how thin, in one place, without softening it.

This is the hub for Learn GEO, eight lessons written to be read in order. Each one ends with something you have in hand, and none of them needs our product: a spreadsheet does the whole method. What we sell, which is the same loop run weekly across three engines, is described on AI visibility tracking, and the method behind every number we publish is on our methodology page.

The loop

The whole series is one cycle with five steps. Lesson 1 is the vocabulary you need before the cycle starts; the rest of the lessons sit on a step.

Step	The question it answers	Where it is covered
Measure	Which questions, asked how many times, of which engines	Lessons 2 and 3
Read	What each number in the result does and does not answer	Lesson 4
Diagnose	Which explanations for an absence survive the evidence you actually have	Lesson 5
Change	What to edit on the page, and what the technical setup has to allow	Lessons 6 and 7
Re-measure	Whether the difference between two runs is larger than the movement you get from changing nothing	Lesson 8

Two properties of the loop are worth stating before you start it, because most published work on this topic assumes the opposite of both.

The first step is measurement, not content. A rewrite with no baseline in front of it cannot be evaluated afterwards, so the order is not a stylistic preference. The second property is that the cycle has no exit. An engine's answer to the same question moves between runs of its own accord, so there is no point at which a result is settled and the measuring can stop.

The eight lessons

Lesson	What you can do when you finish it
1. What GEO, AEO and LLM SEO actually mean	Say which AI surface any claim refers to, and reject a "visibility score" that does not name the engines behind it
2. How to build a question set that measures something	Write a question set that measures your market rather than your own typing
3. How to track brand mentions in ChatGPT	Run the whole measurement by hand and put a confidence interval on the result
4. Reading an AI visibility report	Read mention rate, average position, citation rate and share of voice without mixing up their denominators
5. Why the AI answer left your brand out	Turn a zero into a short list of testable explanations instead of a confident cause
6. How to make content citable	Edit a page so its passages still mean something after an assistant lifts them off it
7. AI crawlers, robots.txt and llms.txt	Tell a documented crawler rule from a repeated one, and find the robots.txt fault that returns 200
8. Did it work? Re-measuring after a change	Decide whether a difference between two periods is a result or an artefact of the sample size

1. The vocabulary, and which surface you mean

GEO, AEO and LLM SEO are three labels for one job. GEO came from a research paper, AEO from marketing vocabulary, and LLM SEO from the search industry, and in current usage they are interchangeable. The distinction that does matter is the surface: an AI assistant with an official API can be asked the same question every week and compared against last week, while Google's AI Overviews and AI Mode are a different surface and are outside what we measure. A supplier reporting "AI visibility" without naming the engines it queried has told you nothing you can act on. Read lesson 1.

2. The question set is the input with the most leverage

An assistant answers the sentence it was given, so the phrasing of your questions decides which brands appear next to you. The rules are mechanical rather than stylistic: each question combines a buyer, a job and a constraint; none of them contains your brand name, because a question that names you produces an answer that names you and a metric that can no longer move; and the set stays frozen between runs, because a reworded question is a new question with its own history. Lesson 2 publishes the gate our own pipeline enforces in code, including the length bounds, the duplicate threshold and the intent proportions, so you can apply the same checks in a spreadsheet column. Read lesson 2.

3. Measuring is arithmetic, not tooling

Ask each question in a fresh conversation, more than once, and record four things per answer: whether your brand was named in the prose, where in a list it sat, which URLs the answer cited, and which other brands shared the answer. Your headline number is mentions divided by valid answers, where a refusal is counted on its own line rather than folded in as a no. Then put a range on it. Lesson 3 works a 95% Wilson interval by hand on one of our own runs and lands on the result that sets the tone for everything after it: three mentions out of three carries an interval from 43.8% to 100%, which does not separate a brand that is always recommended from one recommended slightly more often than not. Read lesson 3.

4. Four numbers, four denominators

A report answers four separate questions, and collapsing them into one score destroys the information you needed. Mention rate is over valid answers. Average position is over the subset of answers where a list position was detected, so an answer that puts you in a comparison table contributes nothing to it in either direction. Citation rate is answer-level and citation share is URL-level. Share of voice counts your side at most once per answer while summing every competitor named, so it falls when an engine starts writing longer lists even though your mention rate has not moved. Lesson 4 reads one of our runs column by column, including four columns that came back empty because that measurement path never recorded them. A missing column is a finding; writing a plausible zero into it is not. Read lesson 4.

5. Diagnosis stops where the evidence does

When an engine names five vendors and not you, the run gives you two pieces of evidence: which competitors took the slots, and which sources the answer cited. Neither explains the selection, and no tool can, because the engines do not expose it. What the columns do is narrow the explanations to a small set that can be told apart by a second measurement designed to tell them apart. Lesson 5 works a real zero, checks it against a control run in the same category on the same day before reading anything into it, and marks the exact line where observation ends and hypothesis starts. Read lesson 5.

6. Editing for passages that survive being lifted

A citable passage means something after it is cut off the page: the subject is named inside it rather than inherited from a heading, the pronouns resolve within it, and every number carries a source and a date. Lesson 6 gives five edits that produce one, applies four of them to passages from this site rather than to invented examples, and then sorts each recommendation into tested and untested. That sorting is the part most articles on this subject leave out, and it is why the lesson is shorter on confidence than its competitors. Read lesson 6.

7. The technical layer is a gate, not a cause

Being fetchable is a precondition for being cited and nothing more, and the two get run together constantly. Lesson 7 covers only what the operators themselves document: the user-agent tokens OpenAI, Perplexity, Anthropic and Google publish, what blocking each one trades away, and the robots.txt rule that breaks quietly, which is that a group naming a specific bot replaces the wildcard group for that bot rather than adding to it. Our own file had that fault until the lesson was written, and the lesson says so instead of shipping the fix silently. It also sets out what anyone has published about llms.txt, which is less than you would guess. Read lesson 7.

8. Before you compare two periods, measure the noise

Re-measurement means asking the frozen question set again and comparing the two confidence intervals rather than the two percentages. It only means something once you know how far the number moves when nothing has changed. Lesson 8 establishes that floor by running the same measurement three times in fifteen minutes with nothing altered, then works out how much measurement a given size of effect requires before two intervals can separate at all. The answer is more than most programmes buy. It also makes the case for the conclusion nobody wants to write, which is that no difference was detectable at this sample size. Read lesson 8.

One category, measured three times

Four of the lessons share one demonstration category, and it is the only one we measured more than once, so it is where you can watch the same readings handled differently at each step. On 2026-09-20 we ran our free-check path against the ChatGPT API three times for the brand Miro in online whiteboard and visual collaboration tools, US market, English. The question generator produced the identical three questions every time, and nothing was changed between the runs.

Run	Recorded, UTC	Brand named	Point estimate	Wilson 95%
docs/content/runs/L2.json	2026-09-20 07:24	3 of 3	100%	43.8% to 100%
docs/content/runs/L4.json	2026-09-20 07:29	3 of 3	100%	43.8% to 100%
docs/content/runs/L8.json	2026-09-20 07:39	2 of 3	66.7%	20.8% to 93.9%

Both intervals are from the table on our methodology page, which is also where the arithmetic behind them lives.

Lesson 2 uses the first of those runs to show how a constraint inside a question changes who appears. Lesson 4 reads the second one column by column. Lesson 8 puts all three side by side and reads the disagreement between them, including the competitor names that came and went while nothing on our side changed, and that analysis is not repeated here. The point for a reader arriving at this page cold is narrower and comes before any of it: a third of the headline number moved without a single input changing. Any comparison built on three answers per period, in either direction, has shown you nothing.

What we can and cannot claim

This section exists so the honest version is in one place rather than distributed across eight pages as caveats you might skip.

What we stand behind. The measurement method: which engines we query and through which APIs, how questions are generated and gated, how each metric is computed and what sits in its denominator, and why a proportion from a small number of trials needs an interval rather than a decimal point. All of it is written out on our methodology page, with the limits listed rather than buried.

What is one study, not a rule. One peer-reviewed paper has tested content edits against a generative engine and published the result: Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735, KDD 2024), which reports that the method can "boost visibility by up to 40%" in generative engine responses (retrieved 2026-09-20). Read the setup before the number. That is an upper bound rather than a typical result, the engine it was measured against was reconstructed by the authors from the top five Google results and gpt-3.5-turbo, and the work dates from 2023. It is good evidence that what is written on a page changes whether a generative engine cites it. It is not a table of expected gains for your category, and the authors report that the effect varies by domain.

What is our hypothesis. Across lessons 6 and 7, the great majority of the tactical advice is reasoning about what the retrieval step has to do, not a result we measured. Definitions in the opening sentence, paragraphs that survive being lifted, tables instead of comparison prose, FAQ headings written as complete questions: none of it is tested, by us or by anyone whose data we have seen. Lesson 6 marks each recommendation individually, and most of them land on this side of the line.

What we do not have. A before-and-after measurement of our own. This site has not been running long enough for two comparable periods, and we would rather say so than publish a comparison we cannot support. So nothing in this series should be read as evidence that any particular edit moves a mention rate, including the edits we recommend.

What nobody can promise. That doing any of this gets you cited. The engines decide, they change without notice, and a page can be retrieved, read and summarised with the brand behind it never reaching the answer text. What the work removes is the failure modes you control. What happens after that is measured, not guaranteed.

What is outside our scope. Google AI Overviews and AI Mode are outside what we measure; our monitored surfaces are ChatGPT, Gemini and Perplexity. Personalisation is also out of scope, because these measurements go through APIs with no signed-in user and no chat history, so they describe what an engine says to an anonymous caller.

Questions

What is LLM SEO? The practice of getting a brand named in answers written by AI assistants, and of measuring how often that happens. The measurable half is a rate: mentions divided by valid answers, across a fixed set of buyer questions that do not contain your brand name, reported with a confidence interval rather than as a single percentage.

Is LLM SEO different from GEO or AEO? Not in current usage. GEO came from a 2023 research paper, AEO from marketing vocabulary, and LLM SEO from the search industry, and all three now describe the same job. Only AEO causes real trouble, because some writing still uses it for Google's answer boxes rather than for chat assistants. Lesson 1 sorts the terms out.

Where should I start if I have never measured this? With lesson 2 and lesson 3, in that order. Write a set of questions a buyer would ask with your brand name kept out of them, ask each one more than once in a fresh conversation, and record the four columns lesson 3 lists. That gives you a baseline, and nothing further in the loop is interpretable without one. On how many questions to write, the two lessons put the emphasis differently: lesson 3 offers a starting range, and lesson 2 declines to publish a recommended size on the grounds that we have no data supporting one. Read both and pick deliberately.

How many of the recommendations in this series have actually been tested? Few. The measurement method is implemented and documented, and one external paper supports part of the content advice under conditions that do not match a 2026 assistant. Everything else in lessons 6 and 7 is our reasoning, labelled as such at the point it appears. We have no before-and-after data of our own on any technique.

Do I need your product to follow any of this? No. Every lesson is written so the method runs in a spreadsheet, and lesson 3 works the arithmetic by hand for that reason. What a subscription buys is the cadence and the sample size: three engines rather than one, three repeats per question, and the intervals computed each week. That is described on AI visibility tracking.

Why does the same question return a different answer each time? Because these models sample rather than read from a fixed table, and small differences in framing retrieve different sources. It is normal behaviour rather than a fault, and it is the reason a single answer is an anecdote. Our three identical runs above differed from each other with nothing changed between them.

Do you measure Google AI Overviews? No. AI Overviews and Google's AI Mode are outside what we measure. Our monitored surfaces are ChatGPT, Gemini and Perplexity, through their official APIs, and no figure from us should be presented as covering Google's AI surfaces.

How long does it take to work through the series? The reading is an afternoon. The loop is not: a question needs at least four weekly runs before it enters a trend on our own reports, so the first comparison you can interpret is roughly a month after your first measurement. Lesson 8 explains why rushing that is the expensive mistake and waiting is the cheap one.

What to do next

Read lesson 1 if the vocabulary is still slippery, and lesson 2 if it is not. Then measure once before you change anything on your site, because a baseline taken after a rewrite is not a baseline. What the measurement covers when we run it, what it costs and what it deliberately leaves out is on AI visibility tracking.