# Before you hire a GEO agency

> A credible GEO provider sells measurable work, such as baselines, content and technical fixes, and repeatable re-tests. It does not sell guaranteed AI rankings. Start with a small paid pilot, write sampling rules, deliverables, data ownership and banned tactics into the contract, and judge results on shares measured over many runs.

- URL: https://geo.casa/best-practices/choose-agency/
- Guide · Observed in practice
- Reviewed: 2026-10-28

This guide is for the person who signs off on a GEO provider: a marketing lead, business owner or procurement manager. It explains which promises are credible, which are red flags, and what the contract should say. Reading it takes about twenty minutes. Running the pilot it recommends takes one to three months.

## The short version

1. Know what GEO can and can't deliver before you take any sales calls.
2. Reject anyone who guarantees rankings or sells mass-published articles.
3. Ask how they measure, and expect repeated sampling, not screenshots.
4. Start with a paid diagnosis and pilot, not a year-long contract.
5. Write deliverables, sampling rules, data ownership and banned tactics into the contract.
6. Check results yourself with the same question set before renewing.

## 1. Know what can be promised

No AI platform documents a way to buy placement in its organic answers, or exactly how it chooses sources. That limits what any provider can honestly promise.

> **Observed in practice:**
> NBD AI Think Tank's GEO Red Book lists what GEO should not promise. The list includes: that any AI platform will recommend a brand in any question; that GEO will bring a set amount of sales or leads in the short term; and screenshots from a single run as evidence of results. It also rules out claiming causation without an experiment design, control group, time window and sample size. (Translated from Chinese.)
>
> Sources: 每经AI智库: GEO红皮书（2026）：生成式引擎优化的边界、风险与治理 (unpublished report, not publicly available; readers can't check it)

What a provider can credibly deliver:

- a baseline of how assistants answer your customers' questions today;
- fixes to your own site and listings (access, facts, structure);
- accurate content on your site and in places assistants read, published under your name;
- corrections of wrong facts at their source;
- repeated measurement that shows the trend.

## 2. Red flags

The fastest way to filter providers is by what they offer.

> **Officially documented:**
> Google says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem", because its core ranking systems focus on quality content and other systems block spam.
>
> Sources: [Google Search Central: Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)

> **Officially documented:**
> Google's definition of spam includes "attempting to manipulate generative AI responses in Google Search."
>
> Sources: [Google Search Central: Spam policies for Google web search](https://developers.google.com/search/docs/essentials/spam-policies)

> **Officially documented:**
> China's 2026 Qinglang campaign treats using GEO for malicious marketing as AI data poisoning. It also covers promoting or teaching poisoning methods and selling related tutorials and tools.
>
> Sources: [新华网（中央网信办通知）: 中央网信办部署开展“清朗·整治AI应用乱象”专项行动](https://www.news.cn/politics/20260430/5a335918ff184d098c169c3dc0401508/c.html)

> **Observed in practice:**
> CCTV's 3·15 show reported that some providers mass-produce false information to steer AI answers. In its test, a GEO tool published more than ten promotional articles about a fitness band that did not exist. Two major AI models then recommended it near the top.
>
> Sources: [第一财经: 3·15揭露GEO“洗脑”：AI大模型被投毒问题日趋严重](https://www.yicai.com/news/103087422.html)

> **Observed in practice:**
> Wisers' event monitoring (15–19 March 2026) records that after the broadcast, the market regulator in Beijing's Huairou district said it had opened an investigation into the company behind the GEO tool named in the show. (Translated from Chinese.)
>
> Sources: 慧科讯业（Wisers）: 315曝光AI投毒，品牌如何做好GEO营销？ (unpublished report, not publicly available; readers can't check it)

Walk away if you hear any of these:

| Offer | Why it's a red flag |
|---|---|
| "Guaranteed first place in DeepSeek / ChatGPT" | No platform documents a way to buy it, and answers vary run to run (section 1) |
| "We publish hundreds of articles a month for you" | Mass-produced content is what the CAC and Google target |
| Fake reviews, rankings or "independent" comparison posts | Spam policies, consumer law and the Qinglang campaign |
| "We see your rankings inside Google's AI" | No tool sees inside Google ([Claim: GEO tools can see your rankings inside Google's AI](https://geo.casa/measurement/third-party-ai-tools/)) |
| Results shown only as screenshots | One run proves nothing |
| They won't name where they publish | You can't check for hidden paid placements |

> A guaranteed AI ranking is either a misunderstanding or a lie. Either way, it shouldn't be in your contract.

## 3. Ask how they measure

Measurement is where weak providers are easiest to spot. Ask them to show a past report and explain each number.

> **Observed in practice:**
> In September 2025, Miaozhen Marketing Science Institute ran 2,025 queries across five product categories and three models: one typical question per category, 135 runs per model. It concluded that results reach basic stability only at 30 or more runs per prompt. Low-visibility brands need 60 runs or more.
>
> Sources: [中国商务广告协会AI营销应用工作委员会 · 秒针营销科学院: 2025年中国GEO行业发展报告](https://g4-img-all.oss-cn-beijing.aliyuncs.com/video/50108560/01503a0d1eaeb4429ee3c0014738e9ea.pdf)

Good answers include:

- a fixed question set built from real customer questions, separated into brand and category questions;
- the platform, app or API, account state, location and date of every run;
- shares with sample sizes, per platform;
- raw answers you can inspect.

> **Observed in practice:**
> Analysys's 2026 report on choosing GEO suppliers, based on company interviews, user surveys, public information and its own monitoring, expects buyers to favour suppliers that give specific, measurable and verifiable results and accept checks by independent third-party monitoring. It warns that suppliers offering only vague "buzz" or many screenshots will quickly lose trust. (Translated from Chinese.)
>
> Sources: 易观分析: 中国GEO行业应用场景成熟度与供应商选择分析报告2026 (unpublished report, not publicly available; readers can't check it)

The standard method is described in [Measuring AI visibility by sampling prompts](https://geo.casa/measurement/ai-visibility-measurement/).

## 4. Start small

Buy a diagnosis and a short pilot before committing to a large retainer.

> **Observed in practice:**
> Aifenxi's 2026 market report says most Chinese GEO providers charge a combined fee under quarterly or annual framework contracts. A base fee covers tool subscriptions and monthly operations. A performance fee is tied to visibility or first-recommendation rate on a preset question set. Billing by token use is emerging for monitoring and automated content. (Translated from Chinese; the report does not publish its survey method in detail.)
>
> Sources: 爱分析: 2026爱分析·中国GEO市场研究报告 (unpublished report, not publicly available; readers can't check it)

If part of the fee depends on performance, the question set and sampling rules decide what you pay. Agree them before work starts, and make sure brand-name questions, which almost always return your brand, don't count towards the target.

A pilot should cover a limited set of questions and platforms over one to three months. At the end you should be able to see the baseline, what was changed, and the re-test.

## 5. Put it in the contract

> **Observed in practice:**
> The GEO Red Book's contract checklist covers: goals (platforms, scenarios, question set size, competitors, target metrics); scope; deliverables; acceptance standards (metric definitions, sample size, sampling frequency, re-test rules, data retention, dispute handling); data ownership; confidentiality; banned practices; change handling; fee structure; and exit and handover. (Translated from Chinese.)
>
> Sources: 每经AI智库: GEO红皮书（2026）：生成式引擎优化的边界、风险与治理 (unpublished report, not publicly available; readers can't check it)

> **Observed in practice:**
> Its sample clauses say back-end metrics such as site traffic, leads and sales are observed by the company and are not a direct attribution promise by the provider, unless a separate experiment is agreed. The provider must not fake cases, credentials, awards, reviews or endorsements. Content touching medicine, finance, education or minors needs approval from a named company approver before publication. (Translated from Chinese.)
>
> Sources: 每经AI智库: GEO红皮书（2026）：生成式引擎优化的边界、风险与治理 (unpublished report, not publicly available; readers can't check it)

Our minimum list:

- **Deliverables:** question set, baseline report, sampling logs, change log, monthly reports, handover pack.
- **Publishing:** every piece published in your name or a disclosed partner's, approved by you. A full list of URLs.
- **Banned tactics:** fake reviews, undisclosed paid placements, mass-produced articles, invented data.
- **Data ownership:** questions, raw answers and reports belong to you and survive the contract.
- **Exit:** handover of accounts, content and data within a set period.

In China, industry standards give buyers a reference point:

> **Observed in practice:**
> The China Advertising Association of Commerce (CAAC) published a five-part GEO group standard on 14 September 2026. Its parts cover general principles and terms, provider evaluation, pricing and measurement, trusted corpora, and process compliance and safety.
>
> Sources: [界面新闻（东方财富转载）: 我国互联网广告服务行业首份GEO团体标准正式公布](https://finance.eastmoney.com/a/202609233882093664.html)

The standard is voluntary. It is useful as a checklist; compliance with it is not a guarantee of quality. For regulated sectors, see also [GEO for healthcare, legal and financial businesses](https://geo.casa/best-practices/regulated-industries/).

## What not to do

- **Don't sign a year-long contract before a baseline.**
- **Don't pay for volume of articles.** Pay for verified fixes and measured change.
- **Don't accept results you can't reproduce** with the same questions.
- **Don't let the provider own your question set or data.**
- **Don't outsource accountability.** Content published for you carries your name. Approve what goes out ([GEO regulation in China](https://geo.casa/foundations/geo-regulation-china/)).

## How you'll know it's working

Re-run the pilot's question set yourself, or with a second tool, and compare shares across rounds.

> **Observed in practice:**
> A provider tracking one client studio saw its recommendation rate move between 10% and 68% within two months, ending at 40%. The provider itself concluded that claims of a "stable first place" did not match its observations. (Translated from Chinese.)
>
> Sources: 合肥图灵象人工智能科技有限公司: 2026普拉提行业首份GEO白皮书 (unpublished report, not publicly available; readers can't check it)

Expect movement. Judge the trend over several rounds, and ignore changes smaller than the sampling error. Report it the way [For marketing leads: measuring GEO and reporting it upward](https://geo.casa/best-practices/measure-report/) describes.

> Pay for work you can check. If you can't rerun the provider's numbers yourself, you haven't bought measurement.

## Evidence at a glance

| Recommendation | Evidence |
|---|---|
| Reject guaranteed rankings and seeded content | Official (Google, CAC); observed (3·15 case) |
| Demand repeated sampling | Observed (Miaozhen study) |
| Prefer third-party-verifiable results | Observed (Analysys supplier report) |
| Start with a pilot; agree question set before performance fees | Observed (market report); our recommendation |
| Contract checklist | Observed (GEO Red Book); our recommendation |
| Use the CAAC standard as a checklist | Observed (news report of publication) |
| Judge trends, not single results | Observed (vendor tracking data) |

## Sources

- [Google Search Central: Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- [Google Search Central: Spam policies for Google web search](https://developers.google.com/search/docs/essentials/spam-policies)
- [新华网（中央网信办通知）: 中央网信办部署开展“清朗·整治AI应用乱象”专项行动](https://www.news.cn/politics/20260430/5a335918ff184d098c169c3dc0401508/c.html)
- [第一财经: 3·15揭露GEO“洗脑”：AI大模型被投毒问题日趋严重](https://www.yicai.com/news/103087422.html)
- 慧科讯业（Wisers）: 315曝光AI投毒，品牌如何做好GEO营销？ (unpublished report, not publicly available; readers can't check it)
- 每经AI智库: GEO红皮书（2026）：生成式引擎优化的边界、风险与治理 (unpublished report, not publicly available; readers can't check it)
- 易观分析: 中国GEO行业应用场景成熟度与供应商选择分析报告2026 (unpublished report, not publicly available; readers can't check it)
- 爱分析: 2026爱分析·中国GEO市场研究报告 (unpublished report, not publicly available; readers can't check it)
- [中国商务广告协会AI营销应用工作委员会 · 秒针营销科学院: 2025年中国GEO行业发展报告](https://g4-img-all.oss-cn-beijing.aliyuncs.com/video/50108560/01503a0d1eaeb4429ee3c0014738e9ea.pdf)
- 合肥图灵象人工智能科技有限公司: 2026普拉提行业首份GEO白皮书 (unpublished report, not publicly available; readers can't check it)
- [界面新闻（东方财富转载）: 我国互联网广告服务行业首份GEO团体标准正式公布](https://finance.eastmoney.com/a/202609233882093664.html)