Ask every candidate to price the same written scope and show the evidence you will receive. A reviewable GEO agency proposal names the target country and category, the frozen buyer-question panel, platforms and sampling dates, original answers and URLs, product-fact approvals, specific pages or outreach work, owners, exclusions and a follow-up check. An AI visibility score or promised ChatGPT placement alone is not an acceptance criterion.
Why a generic GEO quote cannot be compared
“We will optimize 100 GEO keywords” could mean a content brief, repeated model questions, technical fixes, a monitoring subscription or an agency’s own score. Those are different purchases. Give all bidders the same product facts and buyer context before asking for fees. Otherwise a low quote may omit English editing, publishing, external source work or repeat measurement that another quote includes.
Start with a one-page brief: one product category, a target country, primary language, site and product URLs, key competitors, what the brand can approve, and the commercial question the work should help answer. Add the intended AI platforms, but allow candidates to explain why one may be unavailable or irrelevant. Ask for a fixed first phase and a separate optional ongoing scope.
For a hypothetical mini PC company, the brief might specify the US market, two current models, photo-editing and home-office buyers, ChatGPT and Perplexity, and an internal editor who can publish approved pages. This is an example of scope design, not a real client outcome or a default package.
The baseline must be reproducible
Require the proposed question panel before collection begins. Each question needs an intent label: category discovery, task fit, comparison, compatibility, pricing or objection. Capture the same question with platform, market, language, date, full answer and exact cited URLs. Report tool errors or unavailable answers separately. Keep a change log if a question must be rewritten after the panel is frozen.
Ask the agency to define mention, recommendation and citation separately. A brand named in a warning is a mention but not necessarily a recommendation. A cited page can be owned or independent and may not support the specific claim. A product card is a different surface from a prose answer. Show the numerator and denominator for every rate: for example, six brand mentions out of twenty valid answer captures, with failures disclosed rather than silently excluded.
A percentage calculated on only brand-name questions should never be presented as generic category visibility. A platform-specific rate in one country should not be presented as global share of voice. Demand the original observation rows so your team can inspect those boundaries.
- Prompt ID and exact wording
- Platform, market, language, date and run status
- Complete answer and all source URLs
- Brand, product and competitor mentions with classification
- The calculation and denominator behind each summary number
Specify the editorial and source work as actual outputs
The diagnosis should lead to a page-level backlog. For each proposed change, ask for the page URL, the buyer question it answers, the approved product claim it uses, the source for that claim, an editor, a review round and a publication owner. “Create GEO content” is not a deliverable until the number, format, approval route and rights are clear.
If the agency offers third-party source work, distinguish source research, a prepared pitch, submission, accepted coverage and a live URL. The agency can commit to research and outreach within an agreed scope. An independent editor or directory controls whether a contribution is accepted. If paid placements are proposed, the site, cost, disclosure and expected reader value belong in a separate approval line.
Require a handoff file even if your own CMS team publishes. It should contain final approved copy, page titles and descriptions, internal-link targets, any structured data recommendations with supporting visible content, and a record of what went live. An agency should not claim that a special schema type or an AI text file guarantees Google AI or ChatGPT citations; official platform documentation does not support that promise.
Price the same scope and name exclusions
Ask for separate lines for discovery and baseline, content and revisions, technical implementation, tool subscriptions or API usage, external outreach, translation or local review, repeat measurement and reporting. State who pays any third-party charge and whether it needs prior written approval. Do not compare the monthly retainer before comparing the work and access it buys.
Specify deliverable ownership, license to use editable files, account ownership, data export, confidentiality, handover on termination, approval deadlines and the process for correcting an inaccurate claim. A platform’s public feature list does not establish that every capability is in your price tier; a managed agency’s service page does not establish that every listed activity is in your retainer.
If one supplier offers only a diagnostic and another includes publication, ask each for an optional line item that makes their scopes comparable. The right comparison is the cost of reaching an approved and remeasured change, including the client’s staff time.
A practical acceptance test for the first cycle
A buyer can accept the baseline when the agreed panel and raw answer log are complete, exceptions are marked, the cited URLs open to the recorded claims, and the product fact sheet has an approval owner. Accept the content phase when the specified files or pages are delivered, the claims are approved, and the publication or handoff record identifies what went live. Accept follow-up when the same panel is rerun on agreed dates with both observation sets retained.
Consider one illustrative row: question “Which quiet mini PC suits family photo editing in the US?”; platform ChatGPT; capture date; answer text; three products named; two exact source URLs; one incorrect memory specification flagged; product team approves a corrected fact; editor updates the product page; analyst repeats the question later. That row supports a specific action and a check. It still does not prove that the update caused a later mention.
Reject acceptance clauses based only on “Top 3 in ChatGPT,” a proprietary score with no raw rows, or a promised number of independent editorial placements. Those results depend on systems and organizations the provider does not control. A useful contract rewards the work you can inspect and the care with which results are interpreted.
Copy this short RFP into your brief
“We are seeking a GEO partner for [product/category] in [country/language]. Please propose a fixed first phase using [agreed AI platforms]. Include a buyer-question panel and baseline methodology; a sample original answer and cited-URL record; the number and type of pages or editorial assets you will prepare or implement; our product-fact approval points; any external source work; the date and method for repeat checks; the complete file handoff; and a line-item price with exclusions. State what you cannot guarantee and provide one redacted sample of a previous delivery format.”
Ask each bidder to answer one further question: “What would you do if our product is accurately described on our website, but independent sources in the sampled answers only discuss competitors?” The answer should connect the observed source gap to a reader-useful plan and make no promise that a third party will publish.
See the agency-versus-software decision if you are still deciding what to buy. For a China-based brand, use the overseas shortlist method. BrandInAnswers offers managed research, English content and measurement work; the studio approach explains how to prepare a project brief.
Official references
Platform features and requirements can change. Check the current documentation when planning implementation.