A GEO Agency Cannot Promise ChatGPT Rankings. Here Is the Proof It Should Show You.

A credible GEO agency cannot promise that ChatGPT will cite you, that Google AI Mode will surface your page, or that a certain number of LLM visits will become leads. Those systems change their retrieval, ranking, and response behavior outside the agency's control.

What an agency can promise is much less sexy and much more useful: a clean baseline, a set of pages worth improving, a record of what it changed, and a way to tell visibility from actual business results. If a proposal cannot show those four things before work starts, it is not a GEO plan. It is a nice-looking guess.

"GEO agency" now covers wildly different work: technical SEO cleanup, a dashboard that runs 500 prompts, or llms.txt plus generic FAQs. Those services are not interchangeable.

Google's own guidance for generative AI search is unusually plain about the limits. Its AI features use the core Search ranking and quality systems. A page still needs to be indexed and eligible to appear with a snippet, and even meeting the requirements does not guarantee Google will crawl, index, or serve it. That is a useful correction to the idea that an agency can buy its way into an answer box.

Start with the work, not the vocabulary

GEO is a label, not a magic channel. The work should be easy to describe without saying "GEO" at all.

For a B2B software company, that can mean rebuilding a few product pages around implementation and comparison questions with current product detail and proof. For a local service business, it may mean fixing service areas and booking paths before publishing anything new.

The agency should be able to name the customer question, the page that will answer it, and the evidence the page needs. "We will optimize your brand for AI" is not a scope. It is a fog machine.

Google also says that AEO and GEO are industry terms and that its generative features remain rooted in ordinary search systems. That does not make the work pointless. It means an agency should fix fundamentals before selling exotic tactics. We broke down that distinction in our analysis of GEO versus SEO: AI search changes the surface where people discover you, not the need for pages that deserve to be found.

The evidence stack we would require

A real engagement needs a short, inspectable chain from question to outcome. We use five pieces of evidence.

1. A baseline that reflects how buyers actually search

The starting set should be small enough to review by hand. Twenty to thirty prompts is plenty for most teams. Split them by job, not by whichever keyword tool returned the biggest number:

• Discovery: "What is the best way to solve this problem?"
• Evaluation: "What is the difference between these options?"
• Decision: "Which provider is right for a company like ours?"

For each prompt, record the answer, cited sources, date, country, model or search surface, and whether the answer describes the company accurately. Screenshots without the prompt, date, and model are theatre. They cannot be compared later.

That baseline should sit beside Search Console and analytics, not replace them. Google recommends its first-party Search Console data and warns that third-party tools do not have access to its internal ranking data or the ability to guarantee performance in its guidance on outside SEO services. Prompt trackers can be useful. They are not an oracle.

2. A page-level opportunity list

The agency should show the exact URLs it wants to change and why each one made the cut. There are usually only three defensible reasons:

• The page already appears for relevant search demand but gives a thin or buried answer.
• The page has useful expertise or original data but is difficult to crawl, understand, or navigate.
• The site lacks a page for a high-intent question that customers repeatedly ask.

Everything else is secondary. A 200-page content plan before anyone has audited the current pages is a content quota disguised as strategy.

The deliverable should include the intended change, not just a score. "Rewrite the opening to answer the implementation question in one paragraph, add the named product constraint, and link to the pricing comparison" is a real recommendation. "Improve AI readability" is not.

3. A technical eligibility check before the content sprint

If the pages are not crawlable, indexed, fast enough to use, or clear about their canonical version, citation tracking will not rescue them. Google says eligibility for its generative features starts with the same indexing and snippet requirements as Search.

Audit rendering, robots directives, canonicals, sitemaps, duplicates, internal links, author and business details, and schema that matches the page. Adding schema merely because an AI tool requested it is not strategy; Google says there is no special generative-AI markup.

For a deeper measurement setup, our GEO ROI framework separates citation accuracy, assisted traffic, query intent, and platform coverage. That division matters. A page can be visible in an answer and still generate no worthwhile visit.

4. A change log you can audit

GEO work should not disappear inside a dashboard. The client needs a dated record of each change: URL, hypothesis, editor, before state, after state, and source used for any factual claim.

A typical entry might read: "August 19: replaced a generic introduction on /integrations/ with a direct answer to the buyer's setup question; added a screenshot of the actual workflow; linked to the implementation page; checked the page in Search Console." That is plain enough for a marketing lead to evaluate and specific enough for the next person to maintain.

A record keeps a citation spike from becoming proof that a random change worked. Search systems move, models change, and competitors publish. Without a log, every explanation afterward is storytelling.

5. A review that admits what the data cannot prove

The evidence should be read in layers:

1. Eligibility: Can search engines and users access the page?
2. Visibility: Does it appear in the relevant Google or assistant results?
3. Referral: Did identifiable visitors arrive from AI surfaces?
4. Business outcome: Did those visitors subscribe, request a quote, or buy?

The layers are connected, but they are not the same metric. An agency that calls citation volume "revenue" is skipping three steps.

Google's Generative AI performance report covers its own AI surfaces, not ChatGPT, Claude, or Perplexity. Analytics can show known referrals, but not which line drove the visit. We cover a practical read in our Search Console analysis. Report that uncertainty plainly.

A small Mintec snapshot, and the lesson inside it

In the seven days ending August 18, mintec.co recorded 16 sessions from known LLM referral sources: 8 from ChatGPT, 5 from Claude, and 3 from Perplexity. Those sessions produced 25 pageviews across 11 users.

That is useful evidence. It tells us that people are arriving from more than one assistant, and it gives us a starting point for the pages they visited. It is not proof that 16 people "found Mintec through GEO," and it is nowhere near enough to make a revenue claim. The list includes the blog index, product-oriented blog posts, policy pages, and location pages. A serious agency would inspect that mix before making a deck about growth.

This is why we would rather report a small number with clear limits than inflate it into a case study. The job is to learn which questions, pages, and formats deserve another cycle of work. Sometimes the answer is "keep going." Sometimes it is "the page is visible but irrelevant." Both are decisions a client can use.

How to read a GEO proposal in ten minutes

Open the proposal and look for these answers.

What will change? You should see URLs, content or technical actions, owners, and a sensible publishing order. A package that promises "AI authority" without named assets is impossible to manage.

What is being measured? The proposal should distinguish prompt visibility, Search Console performance, referrals, and conversion events. Ask which sources are first-party and which are estimates.

What would make the agency change course? Good work has a stop rule. If the target pages are not indexed, fix that first. If a prompt group produces no relevant visibility after the agreed review window, revise the content hypothesis instead of adding more of the same.

What cannot be promised? The answer should be direct: rankings, citations, crawl timing, and assistant wording are not guaranteed. If the proposal avoids that sentence, assume the guarantee is hiding in the sales call.

When not to hire a GEO agency yet

Do not buy an AI-search retainer to paper over a broken site, unclear offer, or missing conversion path. Fix those first. An agency can help research buyer questions; it cannot invent genuine expertise for you.

The first 60 to 90 days should produce page changes, source notes, a prompt baseline, and a clear account of what happened. Less glamorous than "rank in ChatGPT," but much harder to fake.