What Does GEO Cost? Prices, Effort and Measurement for Visibility in ChatGPT, Perplexity and AI Overviews

The direct answer first. Making a mid-market B2B company visible in AI answers costs 11,000 to 14,000 euros in setup and 1,100 to 1,400 euros per month for a full implementation with ongoing measurement, based on our own documented projects in the DACH region. The technical groundwork alone, retrofitted, is considerably cheaper, and a good part of it you can do yourself. What nobody can sell you is a guaranteed placement, because there is no ranking to occupy. This article covers what the money actually buys, which parts you can handle in-house, how to tell a serious offer from an expensive signpost, and why access to your own content became a separate decision in 2026.

Why the question turned concrete this year

For a while, being visible in AI answers was a topic for early adopters. Two developments moved it into the budget conversation. The first is buyer behaviour: in a Gartner survey published on 9 March 2026 (n = 646 buyers who had completed a B2B purchase in the previous twelve months), 45 percent said they used AI tools during a recent purchase, and 67 percent said they prefer a buying process without a sales rep. If two thirds evaluate without talking to you and nearly half do it partly through an assistant, the assistant is a channel, not a curiosity.

The second development is infrastructure. Until recently, whether AI systems could read your site was a question nobody actively answered. That has changed, and we get to the details further down. Together the two mean the same thing: what used to happen implicitly now has a price tag and a switch.

What GEO actually consists of

The word covers a lot of very different work, which is exactly why offers are hard to compare. Broken into its parts, it becomes assessable, and you will notice that several parts do not need an agency at all.

Read that column on the right honestly and the cost question answers itself in part. Four of the seven items are things you do better yourself, because they require knowledge of your own business rather than a tool. What genuinely needs outside help is the delivery layer, the structured data, and the discipline to keep measuring after the initial enthusiasm fades.

What it costs

The two figures in the bottom row are not market averages; they are what two of our own documented projects cost, and both are described with their measured results in our article on AI visibility. Payback in those cases ran at three to five months, calculated on additional qualified enquiries. Treat them as an order of magnitude, not a quote: the single biggest variable is how much substance your site already carries, because GEO cannot make a thin site quotable.

The new half of the question: access became a decision

Until 2026, the question was only how to be findable. Now there is a second one: whether you want to be reachable, and by which kind of system. Cloudflare, which sits in front of a large share of the web, now separates AI traffic into three categories: Search, which indexes pages for later reference; Agent, which fetches live for a specific user request; and Training, which collects material to build model weights.

From 15 September 2026, Agent and Training are blocked by default on ad-supported pages, for domains newly onboarded to Cloudflare, for sites added by existing customers, and for free-tier accounts that have not changed their settings. Search remains allowed. Cloudflare has also renamed its pay-per-crawl offering to pay-per-use and is working with partners to compensate publishers when content appears in search results or agent queries, a shift from charging per fetch to paying per appearance.

Now the honest scoping, because the headlines make this sound universal. A typical B2B mid-market website carries no advertising and usually sits on a paid plan, so the default block does not apply to it. What does apply to everyone is that the switch now exists and someone should throw it deliberately. If your business case is being cited in AI answers, blocking Agent traffic is the one setting that would quietly undo everything else on this page. Check it once, write down the decision, and move on.

How to recognise a bad offer

• Guaranteed placements. There is no ranking, no ad slot and no parameter to set. A guarantee means either coincidence sold as a service or a misunderstanding of how these systems work.
• llms.txt as the deliverable. Useful, cheap, roughly an hour of work. As the centrepiece of a retainer it is a signpost sold as a road.
• Visibility scores without a method. If a provider reports a number, ask which queries, which systems, how often, and what the variance was. Without that the number is decoration.
• Content volume as the strategy. More articles do not create citability. Attributable, sourced, genuinely new statements do, and those are slower to produce.
• No mention of the delivery layer. Anyone who talks about AI visibility without asking whether your site renders without JavaScript has skipped the precondition.

Measuring it without fooling yourself

The method we use is deliberately unglamorous. Define roughly 30 questions your customers would genuinely ask, in their words, not your product names. Run them across several systems before anything changes, and record three things per answer: are you mentioned at all, is what is said about you correct, and is there a link. That is your baseline. Repeat at fixed intervals.

The part most providers leave out is the volatility. These systems change their sources constantly, so two measurements a fortnight apart can differ substantially without anything on your side having changed. That is why a single reading is close to worthless and a series over several months is not. It is also why we prefer to report the correctness of statements alongside the presence: being mentioned wrongly is worse than not being mentioned, and it is the one failure mode you can usually fix quickly.

When it pays off, and when it does not

GEO pays off when your customers research before first contact and your offering needs explaining, because that is exactly the situation in which someone asks an assistant rather than typing a product name. It pays off less, or not at all, when your business runs through tenders, long-standing framework agreements or local walk-in trade. No amount of citability changes a procurement process that never starts with a question.

The pre-check costs nothing and takes an afternoon: write down the 30 questions and ask them yourself. If your competitors appear and you do not, you have both the problem and the baseline. If nobody in your industry appears, the systems have no good sources yet, which is the cheapest moment there is to become one.

Conclusion

GEO is neither magic nor a new discipline requiring a new budget line. It is the old question of whether your company is findable and correctly represented, asked about a channel where you cannot buy placement and where roughly half of B2B buyers now spend part of their evaluation. The costs are assessable once you break the work into parts: several of them belong in-house, the delivery layer and the structured data belong to a developer, and the discipline of measuring belongs to whoever will still be doing it in month six. Anyone selling you a guarantee is selling coincidence. Anyone selling you an llms.txt is selling an hour. If you want to know where you actually stand before spending anything, ask the thirty questions, or let us run that baseline with you.

Frequently asked questions about GEO cost