# How to Choose a GEO Agency Source: https://enginely.ai/guides/how-to-choose-a-geo-agency/ > A buyer's checklist for choosing a generative engine optimization agency: what to ask, what a good answer sounds like, the claims that should end a conversation, and when to hire nobody. Generative engine optimization is roughly two years old as a commercial category, which means nobody in it has a ten-year track record and everybody's deck says the same six things. The usual buying signals do not work here. This is the checklist we would use if we were buying instead of selling. It is written to be usable in a meeting, and it includes the questions that would rule us out. ## How do you choose a GEO agency? Judge a generative engine optimization agency on its measurement method rather than its case studies. Ask which exact prompts it will track, how often it re-tests them, how it separates a real change from normal answer variance, and what it will show you in a month where nothing improved. An agency that cannot answer those four questions specifically is selling reporting, not measurement — and in a field this young, the method is the only thing that can be checked. - **Ask for the method**, not the results — results here are unverifiable without one - **Guarantees are disqualifying**, because no platform sells inclusion - **Variance handling** separates a measurement practice from a screenshot - **Ownership of accounts** should be yours during and after - **A flat month** should appear in reporting as a flat month - **Sometimes the answer is nobody** — the foundation comes first _Written by Enginely, which sells this service. The disqualifiers section names two things that should rule us out._ ## Nine Questions, and What the Answers Should Sound Like Ask these of every agency on your list, including us. The distinguishing feature of a good answer is that it is specific and slightly boring. ### Which exact prompts will you track, and who approves them? Good: a written set you sign off, built from how your customers actually describe their problem. Bad: “we monitor your brand across AI” with no list, or a list you first see in month three. ### How often do you re-test, and does location vary? Good: a stated cadence, with location and phrasing varied deliberately because both change the answer. Bad: testing that happens when a report is due, from one location, in one phrasing. ### How do you separate a real change from normal variance? Good: repeated runs, and a stated rule for what counts as movement. Bad: a screenshot of one good answer. Generated answers differ between runs on identical input — an agency that does not raise this has not looked closely. ### When we appear, can you show which page or source was cited? Good: the citation, traced to a specific URL or third-party source, so you know what to strengthen. Bad: a mention count with nothing behind it. ### Which parts of this are technical SEO, content, digital PR and local search? Good: a plain breakdown, because most GEO work is those four things aimed at a different scoreboard. Bad: a claim that it is an entirely new discipline requiring an entirely new budget. ### Who owns the accounts, profiles and content afterwards? Good: you do, throughout and after, with no transfer negotiation. Bad: anything held in the agency's name. An agency holding your Google Business Profile is managing a hostage, not a relationship. ### What will you tell us in a month where nothing improved? Good: a direct answer describing what a flat month looks like in their reporting. Bad: visible discomfort, or a pivot to a metric that always goes up. Every programme has flat months; only some agencies show them. ### What would you refuse to do for us? Good: a real list — bought links, generated reviews, location pages for markets you do not serve, claims that cannot be substantiated. Bad: “whatever it takes.” In regulated fields that answer is a liability, not enthusiasm. ### Under what circumstances would you tell us not to hire you? Good: a specific scenario, readily given. Bad: none. An agency that has never turned work away has no threshold, and you cannot tell whether you are above it. ## Claims That Should End the Conversation These are not soft signals. Each one describes something that either does not exist or cannot be delivered, and an agency stating any of them is either misunderstanding these systems or misrepresenting them. - Guaranteed inclusion in ChatGPT, Gemini, Perplexity or AI Overviews. There is no submission process and no paid placement. - A proprietary file, tag or schema that gets you recommended. No markup compels inclusion, and llms.txt is not an opt-in to being cited. - Guaranteed position zero or a guaranteed featured snippet. The source and the exact passage are re-chosen on re-crawl. - A named contact inside a platform who can help. There is no such relationship to sell. - Case study figures with no method, date or query set attached — unverifiable by construction. - Reporting that only ever moves upward across every month shown. - Bought links, generated reviews or manufactured mentions described as authority building. - A recommendation to buy the full programme made before anyone has audited the site. _The last one is the most common and the least dramatic. A diagnosis that arrives after the proposal is not a diagnosis._ ## What a Proposal Should Contain Five things. If any is missing, ask for it in writing before signing rather than after — all five are cheap to state and expensive to discover the absence of. - Baseline: the approved prompt set, the platforms tested, the locations used and the dates captured. - Diagnosis: which stage is failing — access, understanding, retrieval, corroboration or representation — and the evidence for that conclusion. - Scope: the specific changes proposed, with named owners and dates, and an explicit note of what is excluded. - Measurement: the metrics reported, the re-test cadence, and the stated rule for separating movement from variance. - Exit: who owns the accounts and content, notice period, and what you keep on the last day. ## When Enginely Is the Wrong Choice A guide about choosing an agency, published by an agency, is only useful if it can rule its author out. These are real, and the first two are the ones that come up most. - You need lead volume this quarter. This work compounds over months; paid media solves the immediate problem and we will say so. - Nobody internally can approve a specific factual claim in writing. In dentistry, law and healthcare the bottleneck is sign-off, not writing, and no agency can remove it. - Your website cannot be changed within a reasonable timeframe. If the pages cannot be edited, prompt research has nothing to act on. - You want a guaranteed outcome in a contract. We will not sign one, and you should be wary of anyone who will. - You need a single-platform specialist. We measure Google AI Overviews directly and report other surfaces separately rather than claiming one instrument covers all of them. _There is also a good case for hiring nobody yet. If your site cannot be crawled or your core information is wrong in public, that is foundational work — worth fixing before anyone runs a prompt set against you, whether or not we are the ones fixing it._ ## Frequently Asked Questions ### How much should GEO cost? There is no defensible benchmark yet, which is itself worth knowing — the category is too young and the scopes too different for a market rate to mean anything. What you can compare is what is inside a price: the number of prompts tracked, the platforms and locations covered, the re-test cadence, and how much implementation is included rather than recommended. Two proposals at the same price frequently contain entirely different amounts of work. Enginely publishes a fixed price for its Accelerator programme so at least one reference point is checkable. ### Should we hire a GEO specialist or ask our current SEO agency? Ask your current agency first, and ask them the measurement questions above rather than whether they do GEO. Most of this work is technical accessibility, entity clarity, content and third-party evidence, which a competent SEO agency already does. What is genuinely additional is prompt-level measurement — a tracked question set, run repeatedly, with variance handled. If they cannot describe that, you are looking for a specialist to add measurement, not to replace them. ### How long before we can judge whether it is working? Long enough for several re-tests, which in practice means three to six months. Technical and factual corrections can show sooner, but a single favourable answer is not evidence and neither is a single bad one. What you can judge much earlier is whether the agency is doing what it said: is there a baseline, an approved prompt set, a stated cadence, and a report that mentions something that did not work? ### Are case studies worth anything in this category? Only with a method attached. A citation count with no query set, no dates and no location is unverifiable, and generated answers vary between runs, so a screenshot proves less here than it would in traditional search. A case study that names the queries, the dates and the locations can be re-run by you. If it cannot be re-run, treat it as a claim rather than evidence. ### Should we ask for a guarantee? Ask, and treat a yes as disqualifying. No platform sells inclusion in a generated answer, so a guarantee is either a misunderstanding or a commitment the agency knows it cannot be held to. What is reasonable to require in writing is method: the prompt set, the cadence, the reporting, ownership of your accounts, and the notice period. ### Is it a problem that this guide was written by a GEO agency? It is a reason to read it sceptically, which is the correct way to read anything an agency publishes about its own category. The test we would apply: does it contain criteria the author fails? This one names five situations where Enginely is the wrong choice and recommends hiring nobody in some cases. Judge it on whether the checklist is usable against us as easily as against anyone else.