What might someone searching this English phrase be looking for?

When I read generative engine optimization agency, I first understand it as a purchasing question: someone wants a team to improve brand and content visibility in AI answers. That reading still has to be checked against the buyer’s specific need. Some want their product material to be cited; others want their company to enter a supplier shortlist. The two involve different objects of work.

This batch of work orders recorded a GSC signal for that query: 1 impression, average position 2. Without a full time window and more samples, it can only suggest something worth watching; it cannot prove the size of demand, let alone serve as evidence that “this service has already won the market.”

I would rather the client state which kinds of questions they want to be seen in, in what language, and what they want the buyer to do after seeing them. Only with the goal written clearly can you judge whether the agency’s deliverables have a matching object. The related research sits at the research site’s field-test record index; it is not a performance guarantee.

What exactly were the two times we were named?

The baseline round T0 was 17 September 2026, and the monitoring round T1 was 12 October 2026. The fixed panel covered different topics, languages and ways of asking, and 72 answers were recorded per round; the provider topic T6 had 12 answers that round. Tianwen was not in the T0 provider list; in T1, two answers carried the provider identity.

One Chinese prompt described a mechanical parts exporter looking for a domestic website-building and Google SEO team, and the answer put Tianwen on the list. Another Chinese prompt asked directly for professional providers, and the answer listed Tianwen Network under the “website building and SEO integrated” category. This is the position in the raw answer, not a professional endorsement of the companies on the list by this study.

Strict mentions and the broad measure overlap, so they cannot be added together. Early T1 records also contained false positives; after the audit, the valid broad count is 7/72. Copying the “8 times” from an older report would carry already-removed errors back into promotion.

Both appearances of the provider identity were in Chinese prompts; none of the six answers to the English provider prompt in T1 mentioned us. Anyone searching English agency terms should pay particular attention to this boundary. A fuller breakdown of identities can be checked against how to monitor brand mentions, citations and shortlists.

What limitations must this record disclose?

This is a single round of sampled observation, with no causal conclusion. Pages, public information and the source set may change between the two rounds, and the answers themselves fluctuate. I cannot take a mention going from absent to present and conclude that one particular rewrite caused the result.

The total number of cited domains moved from 322 to 245, with only 90 overlapping between the two rounds. That gap is why I do not treat a single list as a stable position. Re-checking the same prompts in future may keep, drop or change the category, and that needs to be recorded.

The run was healthy, so can it be treated as free of contamination?

No. The 72 execution records in T1 succeeded, which only shows that the run finished. The script has no contamination-detection code, yet a claim of “zero contamination” was made; that zero has no corresponding test behind it and should not be used as proof of quality.

That round ran in a clean environment, but the canary only re-ran the local read-denied and leakage-word checks; it did not re-run the full dirty-versus-clean environment comparison. This limitation has to be kept. If an agency has only success statuses and cannot produce contamination checks and failure material, I will hold it to the same standard and ask it to supply the missing evidence.

Can the research environment and other AI products stand in for each other?

This study used dsh’s DeepSeek headless environment. It is not a simultaneous test of all AI products, and it certainly cannot replace every question a real buyer might ask. If the prompts, the model or the environment change, the result should be recorded as a new condition rather than promoted on top of the old one.

How do you judge whether an agency’s evidence is credible?

I look at the prompt first, then the answer. Whether the question carries a brand, and whether it specifically asks to “recommend this company,” changes what the evidence means. Then I check the source URL and its position in the text: is the company name in a citation list, or in a list of selectable providers? Is the answer merely quoting a selection guide the author wrote?

Attribution also has to be checked. We once misrecorded a cited page as belonging to this research site; after checking, the T0 source turned out to be a guide on tianwenwangluo.com. tianwenseo.com still has no citation record of its own across these two rounds. Sites under the same brand cannot borrow each other’s page results.

Then I ask about failure samples. If only the answers that mention the brand are shown, there is no way to know which prompts produced no mention. If material has been filtered, the filtering rules and the original scope should be stated. What you need is a record you can go back to and that can be overturned, not a screenshot you can only admire. You can use the checklist for auditing AI recommendation promises to press on each point.

What should a GEO service deliver?

I would ask it to start from the business question and deliver the fixed prompts, a page map, source checks and an explicit change list. At the page level it has to be able to explain which answer the buyer is missing, what existing material fills the gap, and which facts cannot be confirmed. Lengthening an article, or renaming a service, does not count as done.

Monitoring deliverables, in turn, should include dates, models, independent-session conditions, complete answers, source links and the annotation definition. The raw material has to map to the summary table, and environment gaps should be handed over together with failure samples. A contract can write these up as checkable work instead of making the uncontrollable mention result the sole acceptance item.

Existing technical SEO and content problems should also be described as in scope or out of scope. If a source page will not open or its core facts do not match, the agency should explain how it handles that, rather than talking only about citation packaging. You can prepare the scope of purchase against GEO service deliverables and the website answer and evidence checklist.

A monthly report has to show both changes and unresolved questions. Rounds with no new citations also need an explanation: were the prompts kept the same, did the source pages change, which hypotheses remain unruled out. The corresponding material is in what a GEO monthly monitoring report must record.

Different teams should also be compared on the same scope. Some provide monitoring only; others also handle page changes or checks of public information. I would have each of them list what is included and what is not, and state who actually does the work and what material the client must provide. Comparing package names and one mention screenshot will not tell you whether you are buying research, implementation or a periodic summary.

Whether the deliverables can be handed over is also worth asking. Whether the prompts and raw answers are kept, whether page changes have a source, and whether the annotation rules can be explained all affect whether a later team can continue the observation. With only summary scores and no raw material, it becomes hard at the end of an engagement to judge how the old results were produced; even if the scores look good, there is no reusable research material.

What is still missing between citation and business?

A citation can show that content entered a particular answer, but not whether any reader visited. A mention can make a company a candidate, but not whether the buyer made contact after checking. Visits, actual messages, qualified demand and later sales each need their own records.

Our internal records still state plainly that the path from “cited to inquiry” has no verified metric. So these two mentions cannot be used to estimate how much demand a client would receive, and data from projects in other categories cannot substitute for verification.

If there is real contact, I check the source description, the landing page, what was actually received and the business fit, and leave unknown attribution as unknown. Form tests and WhatsApp clicks are recorded separately and do not count as clients. When traffic appears without demand, you can trace the gap along the check path from visits to inquiries.

What questions can you ask directly in the next conversation?

• Which real buyer questions do you plan to monitor, and how are the language and the way of asking decided?
• Can a screenshot be traced back to the complete prompt, answer, source and sampling record?
• Where exactly does the name appear, and how do you distinguish a guide citation from a shortlist?
• What will be changed on the page, where do the facts come from, and how is the old content preserved?
• Can the samples where the brand did not appear be delivered alongside the rest?
• What will be watched for business results, and which results are currently unverified?

These questions are not meant to make a team guarantee a mention every round; they test whether it is willing to state clearly the conditions and limits under which results arise. If the other side can point out which judgments still lack evidence and say how it will check next time, I have material to keep comparing.

I also ask the team to explain the opposite result: the page has been changed but the next round brings no citation, so how will it record and check that? If the answer is only to keep changing prompts and adding screenshots, the original comparison loses its object. Being able to keep unfavorable results and state the hypotheses still under test supports a purchasing decision far better than packaging every round as a success.

After being named, we still have to answer the same questions. A recommendation list did not verify our cases for us, and it does not make the purchasing decision for the client. Choosing an agency should rest on needs fit, maintainable deliverables and complete evidence; a name on a list is only a lead that still needs checking.

What other common questions are there about agencies?

Does being named twice by AI prove that a GEO service works?

It only proves that the provider identity appeared in two answers that round. There was no causal experiment and no verified inquiry path, so it cannot be used to promise client results.

Did these two mentions come from English-language buyer questions?

No. Both occurred in Chinese-language provider prompts. None of the six answers to the English provider prompt in T1 mentioned us, so it cannot be promoted as a verified English-buyer recommendation.

Can this study represent all AI search products?

No. The study used dsh’s DeepSeek headless environment and a fixed prompt set, and answers from that environment cannot be generalized directly to other products, models or all buyers.

What should you ask for first when choosing a GEO agency?

Ask for the fixed prompts, the complete answers, the sampling dates and environment, source URLs, failure samples, page changes and definitions of the business metrics. A mention screenshot with no raw material should be kept as unverified.