What should a GEO contract say about results, exit and who keeps the data?

The scope clause should list the deliverables by layer, the languages covered, the number of pages and prompts the fee buys, and the end date. Generative engine optimization (GEO) work falls into three layers with different horizons: access (robots.txt, redirects, text served without JavaScript), content and entity (a page per buyer prompt, and one consistent name and description of the company everywhere it appears), and brand mass (how often independent sources name you). A scope that names the layer tells you what is being bought and when a result can reasonably appear; our results timeline guide sets out those horizons.

For a fixed-scope pilot, the clause should read like a bill of materials. Our 90-day pilot names: a baseline of 30–50 prompts per language; access fixes; up to ten prompt-specific pages; entity alignment across the site, profiles and structured data; corroboration outreach (getting independent sites to describe you the way you describe yourself); remeasurement every two to four weeks; a day-90 report. The pricing guide lists what each tier includes, so the contract can point at a published scope.

Two items belong in the scope clause and are often left out. First, a page-to-prompt map: which buyer question each page answers, drawn from the prompts where you are invisible at baseline. A page count without a prompt behind each page is a content quote, whatever the proposal calls it. Second, the exclusions: what the agency will not do (paid placements, link buying, reviews), so that nobody discovers it in month two.

The measurement clause should define the prompt set, the assistants, the session conditions and the two metrics in enough detail that a third party could repeat the measurement. Concretely: 30–50 buyer questions per language, fixed on day one, with a copy delivered to you before any work starts; ChatGPT, Perplexity and Gemini at minimum, per language; each prompt run several times in clean sessions (logged out, or a temporary chat with memory off); and a change log if any prompt is added or removed later. Our measurement guide is the version we sign.

The two metrics need one-sentence definitions in the contract itself. Citation rate is the share of prompts in the set where an assistant cites one of your pages. Share of voice is your share of brand mentions across the answers to the same set, measured against the competitor list recorded at baseline, which is also fixed on day one. A proprietary score with no published definition cannot be written into a contract, because nobody can say later whether it was met.

The reporting half should fix a cadence and a content list. In the pilot we remeasure every two to four weeks; on the retainer, monthly. Each report contains the repeat runs of the original prompts, the comparison with the baseline per assistant and per language, the changes made that period, the plan for the next one, and the raw material behind the numbers: saved answers and screenshots. Referral traffic from assistant domains is read from your own analytics, which the agency does not alter. The red-flags guide describes a report with the prompts missing from it. Few agencies offer this unprompted: in our September 2026 check of the 60 GEO agencies that ChatGPT, Perplexity and Google AI Mode cited in our runs of 14–15 September, only 13 (22%) published their measurement method in full and 1 of 60 published raw run data, so the contract has to demand both.

The results clause should define a review procedure: when it happens, which data are compared and what each party does if there is no progress. The agency does not control generated answers, so promises of a position or date require careful scrutiny. A screenshot of one favourable answer does not replace the agreed measurement. Any contractual remedy for a missed target should be stated separately.

Our version, which the contract references as the 90-day policy: at day 90 the same prompt set is re-run under the same conditions and compared with the day-one baseline. If citation rate and share of voice have stayed inside the baseline’s own run-to-run spread, we say so first, with our analysis of where progress stalled. You then choose one of three options: a free diagnostic month, a change of scope, or a stop. The 90-day guide explains what each option contains and how a lack of progress is investigated layer by layer.

Compare results per assistant and per language so an overall average does not hide differences. Include interim checkpoints in the same clause. If the second measurement shows no progress, check crawler access and the recorded answers then, rather than waiting for day 90. An absence of crawler visits for a few days does not, by itself, prove an access fault.

The exit clause should treat the pilot and the retainer differently, and attach no fee to leaving with the data. The pilot is a fixed-scope project with a fixed price and a defined end date; it should not roll into a retainer unless you sign for one. The retainer runs month to month; the contract should state the notice period as a number of days and say that notice is given in writing. In our terms it is 30 days, and no reason is required.

The clause should also cover early termination of the pilot. Our position, stated in the 90-day guide, is that you can stop before day 90 and receive everything produced up to that point; the only thing you forgo is the day-90 measurement itself. Whatever a vendor’s answer, it should be in the contract rather than in a call.

The exit clause should refer to the data and access clauses. Named agency accounts are revoked on the end date; materials are handed over within the agreed deadline, regardless of who ended the engagement or why. The client should be able to continue independently or with another provider using the work and data already produced.

The data clause should list, item by item, what belongs to you in every exit scenario, and set a handover deadline in days. The list we sign: the prompt set; the day-one baseline with its raw answers and screenshots; every interim measurement; the server-log extracts used in reports; the pages and structured data written for your site; the change log of what was edited where and when; and every report. Our terms hand these over within 10 business days of the end date, in formats you can open without our tools: spreadsheets for the runs, Markdown or HTML for the pages. Ask any vendor for their number.

Three things the clause should say explicitly, because they are where disputes start. The prompt set is yours from day one, not from the end: you hold a copy before work begins, so every later report can be checked against a list you have seen. The pages are yours on publication, whoever drafted them, and remain on your domain. And logs travel with every report by default, because a report is an interpretation and the logs are the evidence.

The clause should also settle whether the agency may keep an anonymised copy. We ask for permission to publish a before-and-after dataset with the brand removed, since a pilot is designed to produce one, and you can decline. Our questions guide explains why that dataset is the evidence a buyer should ask for, and why we do not yet have a client one to show.

The access clause should keep every account in your name and give the agency the narrowest access that lets it do the listed work. Accounts that matter for GEO: the CMS, DNS and hosting, robots.txt and redirects, analytics, Google Search Console and the equivalents for Bing, and the profiles that carry the company’s entity (LinkedIn, directories, review platforms). The agency gets named user accounts on those, not the owner login; the clause says who at the agency holds each one.

Log access-layer changes with their date and author: robots.txt, redirects and the canonical host, the site's single main address with or without www. These affect page access. OpenAI documents about 24 hours for robots.txt changes to be picked up; that is not a deadline for a crawler visit or for gaining or losing citations. The change log helps trace a regression to the changes actually made.

The clause should also say what the agency does not own: no domains registered in its name, no hosting on its account, no content published on properties it controls as a substitute for your site. Analytics stays yours, which is what makes referral traffic from assistant domains a check on the agency’s numbers rather than one of them. On the end date you revoke the named accounts; the agency confirms in writing that it holds no other credentials.

The pricing clause should state a fixed price for a fixed scope, say what the price buys in prompts, pages and languages, and rule out automatic renewal. Our pilot is priced at $2,500–6,000 for the scope listed above, and the retainer at $1,500–3,000 a month, month to month; the pricing guide carries the current figures and what each tier includes. A price with no scope attached cannot be compared with another quote, and a scope with no price attached cannot be budgeted.

Three items to check before signing. Whether the pilot fee is due in full up front, in instalments, or against milestones. Whether the retainer price can change, and with how many days’ notice. And whether anything is billed outside the fee: tools, paid placements, translations. If a vendor recommends paid listicle placements, that decision has a separate cost and should be a separate line, as our guide on listicle placements explains.

A renewal clause that requires you to opt in, rather than opt out, is the simplest protection. The pilot ends on its end date with a report and the three options; if you choose to continue, a new document with a new scope starts.

An exclusivity clause should either be absent or be narrow, mutual and time-limited. Broad exclusivity in your favour (the agency will not work with anyone in your industry) is rarely worth what it costs, because it raises the price and cannot be verified. Broad exclusivity in the agency’s favour (you will not hire another GEO or SEO vendor during the term) removes your ability to get a second opinion on the same prompt set, which is one of the cheaper checks you have.

What is worth writing down is conflict disclosure: the agency tells you if it takes on a direct competitor in the same language and category during the term, and you may end the retainer on the standard notice. Our own terms contain no exclusivity in either direction and include that disclosure line. If a vendor asks for exclusivity, ask what it protects that the measurement clause does not.

The table below condenses the nine clauses into a one-page check you can run on any GEO proposal, including ours. The second column is the minimum wording; the third is what a missing clause typically leaves unverifiable. It is a checklist for a buyer, not legal advice; have a lawyer read the contract in your jurisdiction.

If you would rather test a vendor before reading its contract, ask for ten prompts in one language where you are invisible today, the current answers, the specific changes it would make and a check of crawler access. That is the content of our free AI visibility audit, and it shows, before any clause is signed, whether the vendor measures the way its contract will claim.