Why do I ask what will be delivered first, instead of how many recommendations I can get?

When I choose a GEO vendor, I first check whether they can leave the work behind. Which page was changed specifically, which material was added, which buyer questions were observed — all of it should be traceable back to the original files. If the answer stays at “improve authority” and “increase exposure,” I still cannot tell what they actually plan to do.

What an export business owner wants is for buyers to find the company during the purchasing process and to have a reason to get in touch. A brand name in an AI answer is only one observation along that path. It may appear in the reference materials, or it may appear in a supplier list; those two roles mean different things for a purchasing decision.

In my fixed-question monitoring I ran into exactly this distinction: earlier, when Tianwen was mentioned, it was mainly as a source for selection guides and checklists; only later did I observe answers that listed it among service providers. So I will not rewrite “cited our article” directly into “recommended our service.”

If you have not yet sorted out the difference between diagnosis, page remediation, and ongoing monitoring, you can start with GEO Service Contents and Deliverables. A longer quote is not necessarily more reliable; what matters is whether every piece of work has an object, an output, and a way to be accepted.

What does an acceptable deliverables list look like?

I suggest turning the quote into a worksheet that can be signed off line by line. The acceptance criteria below are my suggestions, not a unified market service standard; the specific scope still has to be agreed according to your product, your language, and your existing website.

A diagnosis report should also be able to determine the next step. If brand information is inconsistent, for example, it should point out which public pages the conflict appears on; if a page cannot support an answer, it should point out which fact the buyer needs that is missing. With only an overall score, I have no actionable object to change.

Delivery boundaries need to be clear before the quote. Remediating your own website, checking third-party materials, running monitoring, and analyzing inquiries involve different materials and permissions. The vendor must not hide evidence it failed to obtain inside “overall good performance.” You can use Materials a GEO Diagnosis Should Deliver to check the report, and then use Scope to Confirm Before a Quote to align both sides’ responsibilities.

I also confirm how these materials can be used after the engagement ends. Whether the original question files, the annotation rules, and the page change records can be exported, and whether whoever takes over can keep verifying them, should be stated in the deliverables list. If all you have is one overall score in the vendor’s dashboard, you lose the basis for comparison once the person changes. This check is about whether the materials remain usable; it does not require believing any promise about results first.

Once I have the raw answers, how should I verify them?

I pick one record from the summary table and trace it backwards to the original question text, the session conditions, and the full answer. If the report says “entered the supplier list,” the original text should genuinely place the company in that list. If there is only a source link and no list position, that label needs to be changed.

Whether the question itself inserts the brand also has to be checked. A need the buyer raises naturally is not the same measurement as “please describe the advantages of company X.” Prompts can be used to test whether brand materials are correct, but a result obtained with a brand in the prompt cannot be treated as the answer a buyer would get without one.

The environment description determines how that result should be read. We once verified that a skills directory in the runtime environment carried local information into the context. For observations that simulate an unfamiliar buyer, the vendor should state whether the session had historical memory, whether it had internal files, and whether it was explicitly isolated; without that description, the question of environment contamination remains open.

A successful run is no substitute for a contamination check either. Our T1 records have a complete successful exit log, but the original script had no contamination-detection code. So the absence of a contamination marker file does not prove there was no contamination. This made me value the detection action itself more than the pretty status label in the report.

Statistical rules also have to withstand counterexamples. The T1 audit once misjudged a similar string in another company’s website address as a hit for our own brand, and that record was removed after correction. A team willing to hand over its false positives and the correction process at least gives you a way to verify; if it hands over only the final number, there is no corresponding original to check.

Why can’t our monitoring be written as a guarantee of results?

I list the observation dates together with the denominators. T0 is 2026-09-17 and T1 is 2026-10-12; each round used a fixed set of 24 questions, with 3 independent sessions each, for 72 answers in total. Under the strict criteria, Tianwen mentions went from 3/72 to 6/72, and being named in a service-provider list went from 0 to 2. The source is the fixed-question panel T0/T1 records.

These are only two rounds of observation. Keeping the questions consistent does not mean model outputs have no variance, and it certainly does not mean that some remediation in between caused all of the change. Over the same period the total number of cited domains went from 322 to 245, with only 90 overlapping between the two rounds. That shift reminds me that the citation pool itself is also changing.

Strict brand-name matches and the broader criterion that includes our own domain also need to be kept separate. The broad criterion went from 8/72 to 7/72, and the increase in the strict criterion must not be used to cover up this observation. A report should keep each definition and not switch statistical methods when it suits the vendor.

These records come from DeepSeek headless probes. They cannot represent all AI platforms, nor can they establish changes in buyer clicks and inquiries after seeing an answer. Our research materials state clearly: the link from citations to inquiries still has no corresponding metric.

So when I read a monthly stage report, I read Misleading Metrics in a GEO Monthly Report first. Some changes are worth further research; where a change lacks evidence of its cause, I can only write it down as an observation, not as attribution.

How should the contract be written so we are less likely to talk past each other later?

Acceptance clauses need to be written as executable actions. Do not just write “optimize website content”; specify the target pages, the scope of changes, the fact-checking requirements, the delivery format, and the inspection method. For ongoing monitoring, agree on the question version, the sampling environment, how failed samples are handled, and how raw materials will be retained.

Completed work and observation results need to be recorded separately. If the page has been changed as agreed, that deliverable can be accepted; if the model did not cite it this round, the observation result should still be reported truthfully. Conversely, one answer that happens to name the brand cannot substitute for a page or material that has not yet been delivered.

When materials are missing, ask the vendor to state the gap concretely. Is it insufficient product documentation, a conflicting company name, or a buyer question with no supporting basis? Which page it is used on, which sentence, and who needs to supply what material should all be spelled out. That is how the owner can decide whether to supply more material or to narrow the scope of this stage.

If a contract involves a “recommendation guarantee,” I first check Evidence and Risks of Promising AI Recommendations. Then, following the GEO Contract Acceptance Checklist, I turn unverifiable adjectives into actual outputs.

What should I do when a quote comes with incomplete materials?

I first ask whether the missing piece affects the judgment. If there is no business attribution data yet, a basic diagnosis can go ahead, but it must not be sold as a plan already validated by inquiries; without fixed questions and raw answers, it is not appropriate to sign a monitoring project whose main basis is “growth in mentions.”

When comparing quotes, align the scope first. If one plan delivers only monitoring screenshots while another includes fact-checking, page changes, and raw records, the amounts were never in the same delivery scope to begin with. My approach is to put both sides’ materials into a single acceptance table and mark the blanks explicitly.

The final decision comes down to whether you can explain what this money bought. Whether the vendor is willing to explain failures, false positives, and unmeasured items can be checked against Common Red Flags of GEO Vendors. Work that can be signed off item by item is what suits item-by-item payment and review.

What other common questions come up before signing?

A vendor promises to get AI to recommend me. Can that be a basis for signing?

I would ask for the promise to be turned into verifiable deliverables and observation criteria. Fixed question sets, raw answers, annotation grounds, and actual page changes can be accepted; whether a given model answer recommends the brand cannot be used to guarantee future results.

If mentions increase, does that prove the GEO service worked?

No. The questions, the environment, and the annotation method have to stay consistent, and sampling variance has to be taken into account. Even if an increase is observed, the change cannot be attributed directly to a particular service, let alone used to estimate inquiries.

With a limited budget, what should I ask the vendor to deliver first?

Ask for baseline materials first: the target buyer questions, a description of the sampling environment, the complete raw answers, and item-by-item annotations. Without a baseline, there is no way to tell whether later reports are repeating the same measurement or switching questions to manufacture good-looking results.

Can I just look at the success screenshots a vendor shows?

Screenshots can help locate evidence, but they should come with the original question text, the session conditions, the complete answers, and the records of misses. Showing only favorable screenshots cannot explain the results across the whole monitoring scope.