Getting cited by LLMs: a field with more consultants than evidence
GEO, answer-engine optimisation, LLM visibility — the offers are arriving faster than the knowledge. What plausibly helps, what is genuinely unknowable right now, and how to spend nothing to find out.
Photo: free stock photography (Unsplash licence) — see imprint
The referrals are real, the methodology is not
Something changed this year that is worth taking seriously: in a decent number of shops we look after, there is now a visible trickle of sessions arriving from chat assistants. Small, but real, and the people arriving that way tend to be unusually far along — they did not come to browse, they came because something told them you sell the part. That is a genuinely new channel and pretending otherwise is silly.
What is not real is the confidence with which people are now selling you optimisation for it. There is no rank tracker for a model's opinion. Ask the same question twice and you may get different sources. Ask from a different account, a different country, a different day after a silent model update, and the answer changes again. Nobody outside those companies has a reliable, reproducible measurement — and you cannot optimise what you cannot measure twice.
Most GEO services being sold right now are guessing in a suit
It costs us money to write this, because GEO would be an easy thing to sell. The pattern is familiar to anyone who was around for early SEO: a new opaque system appears, nobody knows the rules, and the gap between demand for certainty and supply of knowledge fills with confident people. The tell is the promise. If a proposal guarantees mentions in an assistant's answers, ask what mechanism produces the guarantee. There is no admin panel. There is no submission form. There is no rep to call.
The second tell is the dashboard. Several tools now sell you a share-of-voice number for LLM mentions, built by asking a model a list of questions on a schedule and counting names. That is not a measurement of your visibility — it is a measurement of what a sampled model said to a robot with no history, in one locale, on one day. It is directionally interesting and it is not a KPI. Do not let it into a board pack, because once a number is in a board pack somebody starts optimising it.
What plausibly helps is embarrassingly unglamorous
Strip away the vocabulary and the plausible advice reduces to this: be a clear, factual, machine-readable, genuinely authoritative source on a narrow thing. Not because we have proof it moves a model — we do not, and neither does anyone selling you a package — but because every one of these actions pays for itself through other channels regardless. That is the only responsible way to spend money on an unknowable: make sure the work is worth doing even if the theory is wrong.
- State facts as facts — a sentence a model can lift without reconstructing your argument.
- Keep product schema honest and current — price, availability, identifiers, no wishful markup.
- Own one narrow topic properly rather than forty topics thinly.
- Put the answer near the top — buried conclusions are hard to quote and always were.
- Be referenced elsewhere by people who are not you — that has always been the hard part.
Being quotable is a writing problem, not a technical one
The one lever we are reasonably confident about is not schema, it is prose. A page that says 'the M12 variant fits, the M10 does not, and here is the failure mode when people force it' can be quoted in one line. A page that says 'our comprehensive range offers solutions for every requirement' cannot be quoted at all, because there is nothing in it. This was already true for featured snippets, for human readers and for your own sales team. Nothing about it is new — the audience just got one more member.
This is also why the marketing-copy layer of your site is the least likely thing to ever be cited. Nobody quotes your hero headline. They quote the paragraph where you admitted a limitation, named a number, or explained why the cheap option fails in a specific condition. Write more of the second kind, and the GEO question mostly answers itself without a retainer.
What we would actually do this quarter
Spend an hour, not a budget. Segment your referrers so assistant traffic is visible as its own line — you cannot argue about a channel you cannot see. Then ask two or three assistants the questions your customers actually ask, in German, and read what comes back about your category. You are not measuring; you are looking. If a competitor is named and you are not, that is a hypothesis worth an afternoon. If the answer is wrong about your product, that is a factual page you should have anyway.
And then be patient in a way the market currently is not. This field is roughly where SEO was when people were still stuffing keywords into white text — the mechanisms are unpublished, the incentives are unclear, and the eventual practice will look nothing like what is being sold in November 2025. Companies that spend the next year writing genuinely useful, specific, honest pages will be fine under whatever rules emerge. Companies that buy a GEO package will own a report.
| Claim you will hear | Status | What we would do |
|---|---|---|
| 'We guarantee mentions in ChatGPT answers' | No mechanism exists to guarantee it | End the meeting politely |
| 'Here is your LLM share of voice, tracked weekly' | A sample of one robot's session, not a metric | Read it as a hint; never target it |
| 'Schema markup gets you cited' | Unproven — but the markup is cheap and useful anyway | Do it, for the other reasons |
| 'Rewrite everything in Q&A format' | Plausible, unproven, and easy to overdo | Answer first, then argue — as good writing already does |
| 'Be the authority on a narrow topic' | Boring, slow, and the only thing that has ever aged well | Start now; it takes a year either way |
- Assistant referrals are a real, small, high-intent channel — segment them before you argue about them.
- There is no reproducible rank measurement for a model's opinion, so there is no honest GEO KPI.
- Only spend on GEO work that pays for itself even if the LLM theory turns out wrong.
- Being quotable is a writing problem: specific claims get lifted, hero copy never does.
Frequently asked questions
As a separate retainer, we would say no for almost every mid-market shop. The mechanisms are unpublished, the measurement is not reproducible, and the concrete recommendations turn out to be things good SEO and good writing already require. Pay for clearer pages, correct structured data and genuine subject expertise. If those also help you get cited, excellent — but you did not gamble to find out.
Look at your referrer data and give the assistant domains their own segment or channel group in whatever analytics you run. It is a small job and it is the only honest number in this whole discussion: real sessions, from real people, that you can compare to conversion. Everything else on offer measures what a model told a script. Start with the sessions you can actually count.
Only if your content is the product you sell. For a shop, being absent from the systems people increasingly use to decide what to buy is a strange way to protect a catalogue you publish for free anyway. The nuance worth having is separating crawlers used for training from those fetching a page to answer a live question — blocking the second removes you from the answer, which is the opposite of what you want.
There is no established, widely honoured standard here, and proposals in this space appear and fade quickly. Maintaining a parallel copy of your content for machines is also how the two versions drift apart and one of them starts lying about your prices. One accurate, well-structured page for everybody is easier to keep true, and truth is the part that matters when something quotes you.
We do this for a living — Shopware, Node.js, React, ERP integration and automation for B2B.
Talk to an engineer