What Is Actually in a GEO Retainer

“What am I actually paying for each month” is the fair question, and most agency pages answer it with a list of nouns. Here is the version with the work in it.

The four things a retainer should contain

1. Continuous measurement. A fixed set of buyer questions, sampled repeatedly across ChatGPT, Claude, Perplexity and Google AI Overviews, tracked against named competitors. Fixed matters: change the questions and you have destroyed the baseline. Repeatedly matters more, because these engines give different answers to the same question on different runs, so anything sampled once is noise.

2. On-site work. Content and schema on your own pages, written so a model can lift a claim without guessing. This is the visible half and the smaller half. It should also include the path after the click: the page a buyer lands on has to make it easy to ask for a quote or a demo.

3. Off-site work. Getting you accurately represented on the third-party pages the engines actually read in your category: review platforms, the listicles that keep getting cited, community threads, directories. This is the slow, unglamorous half, and on the evidence it is where most of the movement comes from.

4. A report you can audit. Not a score. The questions asked, how many times, what came back, who else was named, and which sources were cited. Then, separately, what happened on your site and in your pipeline: relevant visits, enquiries, and which of those enquiries your sales team accepted.

If a proposal is missing number three, you are buying content marketing with a new label.

What the months actually look like

Month one is baseline and access. Building the prompt corpus from your real buyer questions, running the first full sample, and mapping which URLs the engines cite in your category. It also includes the unglamorous checks: whether robots.txt blocks GPTBot, ClaudeBot or PerplexityBot, and whether what the engines currently say about you is even accurate. Expect corrections rather than wins.

Months two and three are mostly off-site. Claiming and completing review profiles, contacting the publishers of the listicles that keep appearing in the citation map, fixing stale facts about you wherever they already exist, and earning genuine presence in the communities your buyers use. In parallel, the on-site work starts: the comparison page, the pages that answer your category’s real questions, schema.

Do not expect the number to move much in this window. The measurement exists so you can see leading indicators rather than guess.

Month four onward is compounding and correcting. Re-running the same prompt set, watching which specific questions moved, and following the citation map wherever it has shifted. The useful signal here is not the headline percentage. It is “we are now named on four questions we were absent from in March”.

What should be in the contract

• The prompt set, in writing, and a clause that it does not change without your agreement.
• Sampling frequency per prompt. Ask for this specifically. “500 prompts monthly” can mean 500 questions asked once, which is a large amount of noise.
• Access to the raw answers, not only the dashboard. If you cannot read what the engine actually said, you cannot check the parsing, and extracting sentiment and position from a paragraph is harder than it looks.
• Which engines, named.
• What happens to the work if you leave. Content and profiles you paid for should be yours.
• What counts as a good enquiry, and how you will tell the agency which leads were useful. Without that, the only thing anyone can report is visibility.
• How the money splits between the agency’s fee and the budget for the work itself, and what happens to budget that is not spent.

The clauses worth refusing

A guaranteed share-of-voice figure. Nobody can promise this. The engines change, your competitors are working too, and the mechanism is not deterministic.

A results date. The “60-120 days” figure that circulates in this industry is a vendor convention, not a research finding. No primary evidence establishes time-to-first-citation in AI answers. We grade it as unproven in our evidence register even though a confident number would be easier to sell.

llms.txt as a deliverable. It is cheap to ship and no major engine honours it. Google has said explicitly that it ignores it. Shipping it is fine; charging for it as a visibility lever is not.

Anything citing “40% improvement” without qualification. That figure comes from a 2024 paper measured on GPT-3.5 using a redistributive share metric, and a 2025 benchmark contests it. Both are in our research library.

Do you need a retainer at all

Often not, and this is the part agencies leave out.

If you have not measured yet, you do not know whether you have a problem worth a monthly commitment. Run the measurement yourself first; it is free and takes an afternoon.

If the result is that the engines name nobody in your category, you may need one excellent page rather than an ongoing programme. If the result is that you are absent while three competitors are always named, the work is mostly off-site relationship and profile work, and the honest question is whether you want to do that in-house or hire it out.

A retainer earns its cost when the category is contested, the prompt set is large enough that manual sampling stops being practical, and somebody has to own the off-site work every week.

How we structure ours

We no longer sell a visibility retainer. A citation is not a lead, and a monthly fee measured only in mentions gives nobody a reason to ask whether the work produced anything commercially useful.

What replaced it is the Demand Programme. You give us a commercial objective, the buyer you want, a budget and permission to execute. We run the four kinds of work above against that objective, and measure the enquiries your sales team accepts as well as the visibility underneath them. It is typically £4,000–£8,000 a month ex VAT, split into a fixed fee agreed in advance and an execution allowance spent only on work you approve. There is no fixed minimum term; the agreement sets a review point instead, because the off-site work does not fit in a month, and you should be able to stop if the review shows it is not working. Prices and the split are on our pricing page.

We also publish our own numbers, including when they fall. Our most recent run put us at 11.1% share of answers, against 12.8% sixteen days earlier, during a period when we were actively publishing. One engine could not be read on the later date, so the two figures are not a like-for-like comparison, and neither one measures pipeline. That is the kind of reporting you should expect from whoever you hire.

Start with the free demand snapshot and decide afterwards.