How to brief and evaluate a GEO partner
Nine questions that separate a supplier who has done this from one who has read about it, plus what a good proposal contains and the clauses worth putting in the contract.
The short answer
Brief on the outcome and the constraint, not the tactics: what you want to be found for, which competitors you lose to, and what you will not do. Then evaluate on nine technical questions with known answers.
What a good brief contains
The questions you want to be the answer to
Ten to twenty, in buyer language. This is more useful than a keyword list because it is what will actually be tested, and writing it forces internal agreement. How to monitor AI search visibility covers building the set.
The competitors you actually lose to
Not your aspirational set. The ones that come up in lost deals. This becomes the share-of-voice denominator.
What you will not do
No astroturfing, no paid reviews, no undisclosed sponsored coverage. Say it in the brief so nobody proposes it and everybody is on record.
Your current state, honestly
Whether your organic search works, whether your docs are public, whether anyone has looked at robots.txt this year. Suppliers who scope against a flattering picture deliver against a flattering picture.
Who will act on the output
The single best predictor of whether an engagement produces anything. A brief with no named owner produces a document.
How you will judge it
Agree the metrics and their definitions before work starts, not in the first QBR when the numbers are already in front of you.
Nine questions to evaluate on
Most of these have a correct answer. They are not trick questions, but they do separate a supplier who has run this work from one who has read about it.
| Ask | A good answer |
|---|---|
| Which crawler controls whether we appear in ChatGPT citations? | OAI-SearchBot, per [OpenAI's crawler documentation](https://platform.openai.com/docs/bots). If they say GPTBot, they have confused training with retrieval, which is the most consequential error in the category. |
| Does your citation rate include unlinked brand mentions? | No, and they report the two separately. Blending roughly doubles the headline number. |
| How many times do you run each prompt? | Several, across a window. A supplier running each prompt once is reporting noise. |
| Do you average platforms into one score? | No. A blended score hides which platform is failing. |
| Which domains currently carry citations in our category? | They should be able to find out quickly, or admit they have not looked. "Your site" is the wrong answer. |
| What does Google-Extended control? | Grounding for [Gemini](/services/discover/ai-search/gemini) and Google's AI surfaces. If they call it a training opt-out, they will misconfigure it. |
| Do you sell llms.txt? | Ideally not, with the evidence to hand. Selling it is a reasonable proxy for how the rest of the offer is built. |
| What would you do if we do not need you yet? | They should have a version of this answer. A supplier with only one product has only one recommendation. |
| Can we see the raw responses behind a number you report? | Yes, always. A metric you cannot inspect is a metric you cannot check. |
What a good proposal contains
- A baseline before a plan. Any proposal that specifies twelve months of deliverables before measuring anything is selling a template.
- Metric definitions written out, including what is excluded. If they use our published standard, or their own equivalent, that is a good sign.
- Per-platform reporting, with the platforms named.
- An explicit split between on-site work and off-site citation sources, with the off-site part sized honestly as slow.
- A statement of what they will not do, unprompted.
- An exit that leaves you the prompt set and the data, not just a slide deck.
Clauses worth putting in the contract
You own the prompt set and the response data
It is your category's measurement baseline. Losing it at the end of an engagement means starting again, which is a lock-in mechanism whether or not it was designed as one.
Method changes start a new series
Restating history to fit a new method is how a reporting line becomes fiction. Put it in writing.
Disclosure of any tool they own
Suppliers who also sell software have an interest in which tool you buy. Ours is WriteWorks, and we say so wherever we mention it. Ask the same of anyone else.
No undisclosed paid placement
Sponsored coverage presented as earned misleads your buyers and is increasingly labelled by publishers anyway. The exposure is yours, not theirs.
Frequently asked questions
Should we run a paid pilot?
A paid baseline is usually better value than a paid pilot. The baseline is the thing you need regardless, it is portable to another supplier, and it tells you whether the engagement is worth scoping at all.
How long should the first engagement be?
Long enough for the off-site work to show anything, which is a quarter at minimum, but scoped so you can stop. Twelve-month commitments before a baseline are a supplier managing their revenue, not your risk.
What if their answers to the nine questions are all correct?
Then you are talking to someone competent, and the remaining questions are commercial: capacity, fit, and whether your team can act on what they produce. That last one causes more failed engagements than supplier quality does.
Sources
- Google Search Central, hiring an SEO. Google's own questions to ask a prospective search supplier.
About Surge45 Team
AI Search & GEO Specialists
Surge45 is the digital discovery and growth strategic advisory for SaaS. We help software companies become the answer across Google, AI search and communities, then turn discovery into pipeline. We also build WriteWorks, our content engineering platform for AI search.
Related Articles
Get More SaaS Marketing Insights
Join 10,000+ SaaS marketing leaders who receive our weekly newsletter with actionable strategies.
Subscribe to Newsletter