LLM SEO: What Actually Changes, and What You Cannot Measure Yet

Here is the sequence almost everybody goes through. You open ChatGPT, you ask it the buying question a customer would ask — best waterproof duffel for sailing, whatever your category is — and five brands come back, none of them yours. So you ask again in a fresh window to make sure, and this time six brands come back, four of them different from the first five. One of your competitors was in both. You have now got two contradictory readings and no idea which one describes reality.

That instability is not a bug in your test. It is the central fact of this whole discipline, and it is the thing the guides ranking for this term do not tell you. I read the three pages sitting above this one before writing; all three hand you a checklist, and not one of them explains why you cannot tell whether the checklist worked.

The symptom: you asked, you were not named, and the answer keeps changing

If your checks keep contradicting each other, nothing is broken — you are sampling a probabilistic system one draw at a time.

The usual reaction is to take a screenshot of the good answer and treat the bad ones as flukes, or to take a screenshot of the bad answer and conclude the site is invisible. Both are the same error in opposite directions. A single answer from a model tells you roughly what a single coin flip tells you about a coin.

The second symptom people bring is subtler: organic traffic soft, rankings unchanged, revenue flat or down, and no obvious cause in Search Console. That pattern is consistent with answers being read instead of clicked, and it is also consistent with four other things, which is why it needs measuring rather than assuming.

What is LLM SEO?

LLM SEO is the work of getting a brand named and cited inside generated answers — in ChatGPT, Gemini, Perplexity, Copilot and Google’s AI Overviews — rather than ranking a blue link for someone to click.

You will also see it called generative engine optimization, answer engine optimization, GEO, AEO and LLMO. They describe the same work. The proliferating acronyms are a marketing problem rather than a technical distinction, and the term you search for does not change what has to be done.

The most useful framing: this is a subniche inside SEO, not a replacement for it. These systems still pull heavily from Google and Bing results, so classic ranking work keeps paying, and it now pays twice. Anyone telling you that SEO is dead and a new discipline replaces it is selling the new discipline.

Why does the same question give a different answer every time?

Because the model is sampling, not looking up. A generated answer is constructed fresh each time, and the retrieval step that feeds it can pull a different set of sources on each run.

Three separate sources of movement stack on top of each other, and it is worth knowing which is which.

Put those together and the conclusion is unavoidable: one check proves nothing, and a favourable screenshot proves less than nothing because of who tends to send them. Any agency deck built on screenshots of good answers is showing you the tail of a distribution.

How do you measure LLM SEO without fooling yourself?

By sampling. Ask the same question a fixed number of times in fresh, logged‑out sessions, and record the share of runs you appear in rather than whether you appeared.

That single change — from a yes or no to a rate — is what turns this from theatre into measurement. A brand named in four runs out of ten has a real number attached to it, and a number can move. A brand with a screenshot has an anecdote.

1. Fix a question set. Eight to fifteen buying questions in your category, phrased as a customer would ask them, never including your brand name. Write them down and do not change them, because the wording is part of the measurement.
2. Fix a run count. Five to ten runs per question per engine. Fresh chat, logged out, memory off. Consistency matters more than the size of the number.
3. Record the rate, not the event. Share of runs in which you are named, per question, per engine. Competitors too, in the same pass — their rates are the only benchmark that means anything, and you are collecting them for free.
4. Record the cited sources. When an answer names its sources, those pages are the actual competition for the citation, and they are frequently not the pages outranking you in Google. This is the most actionable output of the whole exercise.
5. Repeat monthly, on the same set. Month‑over‑month movement in the rate is the signal. Anything measured once is noise wearing a suit.
6. Watch branded search and direct traffic alongside it. Given the attribution problem, aggregate branded demand is the closest thing to a business outcome you will get.

Do this for your top three competitors at the same time and you also get the answer to the only strategic question that matters here: is this category winnable for you, or is it locked up by brands the models already know?

What actually changes for a store

Less than the guides imply, and the biggest change is where effort goes rather than what the effort is.

Ahrefs ran the largest correlation study on this, across seventy‑five thousand brands, and published it with the caveat that every factor was weak and correlation is not causation. Read carefully, its ordering is still the most useful thing available: mentions of the brand elsewhere on the web, and on YouTube in particular, tracked far more closely with being named in AI answers than backlinks or domain authority did. Off‑site presence beat on‑site optimization, and it was not close.

The honest reading of that is uncomfortable for anyone selling a technical audit. Much of what those correlations measure is simply being a known brand, which a small store cannot decide to be. The usable conclusion is about allocation — where the next hour goes — not a promise that the hour converts.

For a store specifically, the first two rows and the comparison pages are where I would start, because they are cheap, they are entirely within your control, and they improve the page for people as well. The technical layer of this overlaps almost completely with ordinary Shopify SEO work, which is the practical argument for not treating it as a separate budget.

What has no evidence behind it

Two things are being sold hard right now with very little behind them, and one of them has been publicly contradicted by Google.

The llms.txt file. Google’s own AI optimization guidance states plainly that it is not needed for AI Overviews, AI Mode or any other generative search feature, and Google staff have compared it to the old keywords meta tag and confirmed they are not pursuing it. Server log analyses find the major AI crawlers rarely or never request the file. It costs ten minutes, so build one if it makes you feel better, but nobody should be charging you for it as an AI search deliverable.

Publishing volume. Page count is one of the weakest factors in the correlation data, which makes a programmatic content push a poor answer to this problem. It was already a poor answer to the old one.

The ten minute check you can run yourself

Run this before you pay anyone, including me. It takes ten minutes and it tells you whether there is a problem worth spending on.

1. Open a fresh, logged‑out chat and ask the buying question for your category — the product and the use case, never your brand name.
2. Write down every brand named, in order.
3. Do it twice more in new sessions. Compare the three lists. The overlap between them is the real result; the differences are the variance you now know to expect.
4. Ask the model why it named those. It will usually cite specific pages. Those pages, not the sites outranking you, are your actual competition for the citation.
5. Repeat the whole thing in Gemini and Perplexity. The overlap between engines is smaller than most people expect, which is itself useful to know.
6. Check your robots.txt for the AI crawler names above. If any are blocked, stop here and fix that first, because nothing else counts until it is true.

If you are named in most runs across most engines already, you do not have a problem and you should spend the money elsewhere. If you are named in none of them and a competitor is named in all of them, look at what the model cites when it names them — that is the brief, and it is usually a review site, a roundup or a forum thread rather than anything on their own domain. How I approach AI search visibility starts from that same check, because it is the one piece of evidence in this field you can gather yourself before spending anything.

Questions people ask