GEO agency: what it actually does, what it costs, and how to pick one

A new kind of supplier showed up over the past year: the GEO agency. Some have spent months working on the problem seriously; others changed the title of the proposal, added a paragraph about ChatGPT to the same old report, and raised the price. From the outside they look much alike, and the difference only becomes visible three months in, when someone has to show results.

This guide is for the buyer's side. What work GEO actually involves, which questions separate the people doing it from the people naming it, what it costs in 2026, and — the part no agency puts on its website — when you're better off hiring nobody yet.

Key takeaways

• +GEO doesn't replace SEO, it sits on top of it. If your site isn't crawlable or fast, there's nothing to optimise for AI yet.
• +The actual work is three fronts: extractable content, a consistent brand entity, and monthly measurement inside the models. A proposal that doesn't name all three is SEO with a new label.
• +The question that filters hardest: “how will you measure whether this worked?” With no defined prompt set and no baseline taken before starting, nobody can tell — you or them.
• +Nobody can guarantee showing up in ChatGPT. Anyone who guarantees it is selling something they don't control; the models have no ranking dashboard and no position API.
• +A reasonable monthly retainer starts around USD 500, and a one-off audit runs USD 800–2,000. Far below that is usually an automated report.
• +If you sell something nobody researches before buying, or your site isn't even in Google yet, GEO isn't your next expense.

What a GEO agency actually does

GEO is Generative Engine Optimization: the work of getting ChatGPT, Perplexity, Gemini and Claude to name and cite your brand when someone asks them about your field. On top of ordinary SEO, which remains mandatory, it adds three fronts classic SEO doesn't cover.

1. Content that can be extracted and quoted

A model doesn't “rank” a page: it lifts a fragment and cites it. That changes how you write. The answer goes at the top and complete in one paragraph, figures carry their unit and their date, comparisons go in a table or list, and every claim has to stand on its own outside the article's context. Writing that's excellent to read straight through can be terrible to cite, because its best idea is spread across five paragraphs.

2. The brand as a recognisable entity

Models work with entities, not strings. They need to be able to assert who you are, what you do and where you are, with no contradictions between sources. That means coherent structured data, the same name, address and phone everywhere, real authors with verifiable profiles, and presence in the sources models consult. It's the dullest part of the job and the one that weighs most, because a single contradiction — a phone number from one country and an address in another — breaks the resolution of the whole entity.

3. Monthly measurement inside the models

Without this, the other two fronts are an act of faith. You define a set of real questions from your field, run them every month in each model, and record whether you appear, where in the answer, and citing which source. It's manual, repetitive work — which is exactly why it's the first thing missing from cheap proposals.

Six questions worth asking before you sign

• +“How will you measure whether this worked?” If the answer doesn't include a concrete prompt set and a baseline taken before starting, there's nothing to compare against in three months.
• +“How much of this is technical SEO?” The honest answer is “quite a lot, especially early on”. If they say none, they don't understand how models reach content.
• +“Show me a case where you measured brand citations, not rankings.” It's the hardest test and the one fewest people can pass today, because the discipline is new. An “I don't have one yet, but here's how I'd measure it” is a better sign than an invented case study.
• +“What happens when the model changes?” It happens often and nobody controls it. Someone with a plan will say so without drama; someone without one will act like it doesn't happen.
• +“What do you need from my side?” A GEO project with no access to someone who knows the business produces content that's correct and generic. If they ask for nothing, they're planning to write without you.
• +“What do I keep if we stop working together?” The schema, the content and the documentation should live on your site under your name. If the work lives inside an agency tool, you rented instead of building.

Nobody can guarantee ChatGPT will name you. There's no ranking dashboard and no position API: there's a probability that moves with every model update. Anyone guaranteeing a result there either doesn't know that, or is counting on you not knowing it.

Signs you're being sold SEO under a new name

None of these proves anything on its own, but three together is reason enough to get a second proposal.

• +The monthly deliverable is a Google rankings report with an AI section that measures nothing verifiable.
• +The plan never mentions structured data, brand entity or external sources — only “optimised content”.
• +They talk about “keywords for ChatGPT”, which don't exist: a model doesn't index by keyword, it answers natural-language questions.
• +They promise results in 30 days. Technical fixes do land fast; a model starting to name you does not.
• +They charge for “optimising your llms.txt” as if that were the service. It's a text file: an afternoon's work, and no substitute for the real job.
• +They never asked what you sell or who to. Without that you can't define the prompt set, which is the heart of the measurement.

What it costs in 2026

The market is young and prices are scattered, but three bands are reasonably stable. A one-off audit — diagnosis, prioritised plan, nothing else — runs USD 800 to 2,000 depending on site size. A monthly retainer that includes execution starts around USD 500 and rises with content volume. Larger projects across several markets or languages are quoted separately and rarely come in under USD 2,000 a month.

Below that, what you're usually buying is a tool-generated report with little human work on top. It isn't a scam — the report exists — but the value in GEO lies precisely in the manual part: reading what the models actually answer, working out why they cite someone else, and fixing the content accordingly.

When not to hire anyone yet

There are three situations where GEO is the wrong spend, and it's worth hearing that from someone who sells it. First: your site is new or not indexed. If Google isn't crawling you properly there's no content for a model to find — fix the basics first. Second: you sell something nobody researches before buying, like a takeaway or a corner shop; there, the map and local reputation pay off more. Third: you're never going to publish content. GEO needs material to cite; with nothing new to say, the work stops at the technical layer and runs out in two months.

What a good first quarter looks like

As a yardstick: in the first two weeks there should be an audit with a prioritised plan and the baseline prompt measurement done — you need to know where you started. Weeks three to six go on the technical base: speed, crawling, structured data, a consistent entity and access for AI crawlers. From month two, content written to be cited plus the external mentions work. And every month, without exception, the report with the prompt set run again.

If by the end of the third month nothing has moved on any of the three measured things — rankings, model citations and external mentions — it isn't necessarily badly executed, but it does need a concrete explanation of why, and what changes from here.

At loco22 we do exactly this work, and we're happy for you to put the six questions above to us — we'd rather that than a sales call. If you want to see where you stand before deciding anything, the initial audit includes your baseline in ChatGPT, Perplexity and Gemini plus a prioritised 90-day plan. Tell us what you sell and who to, which is where this starts.