GEO agency recommendations: Why AI answers change and how to choose
When you ask an AI, "Recommend an agency to manage my GEO," you get roughly six to eight names. Ask the same question tomorrow, and much of that list has been replaced with names you've never seen before.
From June 21 to July 21, 2026, PION posed this question to ChatGPT and Gemini 44 times and tallied every source attached to the answers. The answers cited 601 URLs across 152 domains, and not a single domain accounted for more than 10% of the total. Because the list is assembled anew with each call, collecting names only expands your list of candidates instead of narrowing it.
1. Which sources does AI look at when recommending an agency?
It cites content written by others before companies' own official websites. In the 44 measurements, the most frequently cited source was overseas community posts, which accounted for 48 of the 601 citations in total, a share of 8.0%. Next came the own site of the single most-cited Korean agency, followed by academic paper and technical document repositories. The own domains of the remaining Korean firms were scattered in small shares.
| Type of cited source | Share of 601 | Content |
|---|---|---|
| Overseas community posts | 8.0% | Reviews and firm-comparison threads |
| Domestic Korean agencies' own sites | 5.5% (top single site) | Service introduction pages |
| Academic paper / technical document repositories | 3.3% | GEO concept and research citations |
| Overseas agency directories | Around 2.0% | Firm-listing pages |
If a firm appears frequently in AI answers, that means there is a lot of third-party content covering that firm. The mentions may have increased because the firm does good work or because it worked hard on visibility; the answer list alone gives you no way to tell which.
2. Why do different names appear each time?
Ask the same questions 300 times and 648 firm names come up. That is what PION tallied after submitting 7 recommendation-type questions to ChatGPT and Gemini 300 times. On average, 7.8 firms were mentioned per answer, and each answer cited an average of 12.9 URLs. Even the most-mentioned firm appeared in only 35 of the 300 runs, or 11.7%.
AI generates each answer anew with every call, so there is no ranking that holds for days the way a first-page search result can. During measurement, AI sometimes opened its answer with a warning. The gist was that no firm can guarantee visibility on AI services, so you should be wary of any that advertises such a guarantee.
3. Five items to check before signing
What to look at in a proposal is the measurement conditions, not the list of case studies. Every company's portfolio looks similar, but measurement conditions are either written down in numbers or missing entirely.
| Item to check | What to ask for | If there's no answer |
|---|---|---|
| Measurement engine | Which of ChatGPT, Gemini, Perplexity, or Claude, and how many times, are measured | You can't verify performance reports |
| Prompt list | Whether the questions used for measurement are agreed in advance | They'll cherry-pick favorable questions to report |
| Baseline figures | Whether the pre-contract citation rate is provided as a number | There's no baseline to calculate improvement against |
| Scope of publishing | Whether it covers content production and publishing, or stops at strategy documents | Execution falls to the advertiser |
| Deliverable ownership | Whether published content and measurement data are the advertiser's assets | The materials vanish when the contract ends |
The measurement engines and prompts come first. If you don't fix these two when signing the contract, the agency will decide which questions go into the report three months later. That is why PION first maps brand-specific prompts during the diagnostic stage and starts by measuring responses across four engines (ChatGPT, Gemini, Google AI Overviews and AI Mode). The diagnostic procedure is laid out step by step in What is an outsourced AI visibility audit, and how does it work?.
4. Agencies that diagnose and agencies that also publish
GEO agency services come in a diagnostic/consulting type and a full-service type that includes publishing, and which one fits depends on whether the advertiser has someone in-house to write content.
| Category | Diagnostic/consulting type | Full-service type including publishing |
|---|---|---|
| Deliverables | Diagnostic report, improvement roadmap | Report + content production/publishing + monitoring |
| Who does the work | The advertiser's in-house team | The agency |
| Suitable when | An organization that already has content staff | An organization with no staff to handle publishing |
| Duration | Around 4 weeks | 3 months or more |
No matter how good a roadmap you receive, if no new pages go up for AI to read, the citation rate stays the same. What you can hand over when choosing the full-service type is covered item by item in Generative engine optimization: What can you outsource?. How the numbers change when you also hand over publishing is covered in How AI citation rates change when a GEO agency handles content publishing.
5. When, and by what, do you assess performance?
At least three months. The first one to two months go into establishing the baseline citation rate and setting up a publishing cycle, so judging performance by the numbers from that period leads to the wrong conclusion.
Use the following metrics to assess performance.
- Measured citation rate across four engines: the rate at which the brand or domain appears in the body of the answer
- Change in share versus the baseline: the size of the change rather than the absolute value
- AI referrer traffic: actual visits verified by isolating AI-service domains in GA4
- Mentions on external channels: how often the brand is named in communities and media outlets
- Indexing and technical metrics: whether AI crawlers can read the pages
If the citation rate rises but traffic does not follow, the brand is only appearing by name in lists. Adding pages only to your own site while external mentions stay at zero runs directly against the tendency described above for AI to pick up third-party sources first.
The structure of PION's managed GEO service is laid out in the managed GEO service overview. We adjust the contract format among fixed-fee, performance-based, and hybrid arrangements according to industry and goals; the details are on the pricing page.
If you want to receive only diagnosis and design and handle publishing in-house, you'll compare firms by their consulting deliverables. A comparison table of deliverables, including the prompt list, baseline figures and raw data, is in Choosing a GEO agency: What to compare before hiring a consultant.
Frequently asked questions
Can you recommend an agency to manage my GEO?
When PION put this question to ChatGPT and Gemini 44 times in June and July 2026, collecting names did not narrow the shortlist. The sources attached to the answers were scattered across 601 citations from 152 domains, and not a single domain exceeded 10% of the total. It is faster to narrow the candidates with verification items. Ask four things: whether the measurement engines and prompts are written into the contract, whether the baseline citation rate is given as a number, whether they handle content publishing, and whether ownership of published material belongs to the advertiser. PION starts by measuring responses across four engines (ChatGPT, Gemini, Google AI Overviews and AI Mode) and runs content production and publishing as well.
Can I just trust the list of GEO agencies that the AI recommended?
Over the same period, 648 firm names appeared across 300 runs of recommendation-type questions, so the list is useful only for gathering candidates. Even the most frequent one showed up in only 35 of the 300 runs, or 11.7%. The source AI referenced most was overseas community posts. Firms' official sites came below that. Once you've confirmed the names, verify them against business registration information and contract terms.
How long should a GEO agency contract run?
At least three months. In the first one to two months, the baseline citation rate is established and a publishing cycle is set up, so you cannot judge performance from the numbers for that period alone. Only after three months can you compare before and after publishing on the same basis.
How does a GEO agency differ from an SEO agency?
An SEO agency looks at rankings on the search results page, while a GEO agency tallies how often the brand is named in the body of AI answers. It's common for a brand to be on page one of search yet absent from AI answers, so a ranking report alone can't confirm AI visibility.
Are performance-based contracts possible?
Yes. PION adjusts the contract format among fixed-fee, performance-based and hybrid arrangements to fit your industry and goals. If you choose performance-based pay, first define in writing which metrics count as performance. AI answers vary even for the same question, so to avoid disputes later, base it on the average citation rate over a set period rather than a ranking measured once.
- Category
- Guide