Before You Hire a GEO Agency: 4 Green Flags and 5 Red Flags

How Do You Tell a Real GEO Agency From a Rebranded One?

You tell them apart by what they can show you before you pay. A credible GEO agency can produce a documented baseline of how AI systems currently describe your brand, the exact prompts it measured, and the raw responses behind every number. A rebranded one leads with a score, a guarantee, or a content quota.

Ask a search engine for the best accounting software for construction companies, and you get a ranked set of pages to explore. Ask an AI platform this instead:

We are a 70-person construction company operating across three locations. We need job-costing, approval workflows, and integration with our payroll system. Which accounting platforms should we consider?

Now the platform is not simply retrieving pages. It is interpreting requirements, comparing options, and recommending specific brands. That is the problem Generative Engine Optimization is meant to address. Your company must still be discoverable, but it must also be understood accurately and presented as a credible answer.

The demand has created a wave of agencies selling GEO, Answer Engine Optimization (AEO), and AI SEO. The terms are used inconsistently. In this guide, GEO means the work of improving whether a brand is accurately discovered, understood, cited, and recommended in AI-generated answers. Some agencies have real expertise. Others have added a new acronym to a familiar service.

If you are evaluating an AI search partner, four green flags separate the operators from the repackagers.

1. Do They Treat GEO and SEO as One Connected System?

A credible GEO agency will not tell you that SEO is obsolete. Google's own guidance states that the best practices for SEO remain relevant for AI features in Google Search, and it sets a concrete eligibility bar: to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet (Google Search Central, AI features and your website).

The retrieval behavior underneath makes the point sharper. Google documents that both AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response. For the construction-company prompt, those related searches might cover:

• Accounting platforms with construction job-costing
• Software for multi-location finance teams
• Construction payroll integrations
• Alternatives to the buyer's current accounting system
• Customer reviews and implementation experiences

ChatGPT search follows a comparable pattern, turning a request into one or more targeted search queries before composing an answer (OpenAI Help Center, ChatGPT search, verified August 2, 2026).

You are not being retrieved once. You are being retrieved several times, against questions you never wrote.

That means a capable partner has to examine the entire discovery system, not one layer of it:

• Can search and AI crawlers access the important content?
• Can search engines index and understand it?
• Does the site clearly explain what the company does and whom it serves?
• Are relevant pages connected through a sensible internal-link structure?
• Do credible third-party sources confirm the company's claims?
• Is the brand associated with the categories and problems it wants to own?

GEO and SEO require different measurement. They should not operate as isolated strategies. If you want the full breakdown of where the disciplines diverge, we mapped it in SEO vs AEO vs GEO.

Ask them: "Show me how technical SEO, content, authority-building, and AI visibility fit into one roadmap."

The green flag: They can connect every GEO recommendation to a broader search and brand-discovery strategy.

2. Do They Establish a Verifiable Baseline Before Recommending Work?

Before an agency proposes new content, technical changes, or digital PR, it should show you where your brand stands today. Without a documented starting point, neither side can prove what improved.

A useful baseline answers:

• Which buyer prompts are being measured?
• Which AI platforms are included?
• How often is the brand mentioned or cited?
• Which competitors appear more frequently?
• Which sources influence the answers?
• Is the brand being described accurately?
• Which pages are surfaced, and which are ignored?

Start with first-party data where it exists. Google announced Search Generative AI performance reports in Search Console on June 3, 2026 (Google Search Central Blog). The Search report groups impressions from generative AI features by page, country, device, and date, and Discover has its own separate report (Search Console Help).

That is a genuine first-party baseline, and it is deliberately narrow. Verified August 2, 2026: the reports are still rolling out to a subset of sites, and they currently report impressions rather than clicks, click-through rate, position, or query data. They also cover Google surfaces only. Nothing in Search Console tells you what ChatGPT, Claude, Gemini, or Perplexity say when a buyer asks which vendor to choose. That is why a complete program needs more than Search Console.

An initial program should track a documented set of buyer-relevant prompts, brand mention and citation frequency, the pages and domains being cited, competitor share of voice, and copies of the underlying AI responses.

The prompt set matters as much as the resulting score. If an agency measures vague or irrelevant questions, the dashboard looks impressive while telling you almost nothing about real buying decisions. We wrote about why the decision-stage prompt is the one that pays in Decision-Stage AI Visibility.

As the program matures, measurement expands to results segmented by platform, market, audience, and buying stage; brand sentiment and factual accuracy; changes in citations, source influence, and competitor share of voice; and AI referrals, leads, and conversions where attribution is available.

At Pixelmojo we built Radar to connect those layers instead of collapsing them into a single visibility score. Radar audits technical readiness and content signals, then measures citations, source influence, hallucinations, brand disambiguation, and prompt-level share of voice across ChatGPT, Claude, Gemini, and Perplexity. One disclosure worth demanding from any vendor, including us: Radar queries the provider APIs, not the consumer apps, which can layer different models, memory, personalization, and live web search on top. A number is only defensible if you know exactly what was measured. That is the argument behind A Score You Can Defend.

Does a Good Technical Readiness Score Mean You Are Visible in AI Answers?

No. Technical readiness makes a site reachable and interpretable. It does not make a brand recommended. Our own audit data separates the two cleanly.

One definition first, because the number below is meaningless without it. Prompt share of voice is the percentage of brand mentions you capture when a fixed set of category prompts is run against multiple AI engines. If a buyer asks four engines five category questions and your brand is never named in any answer, your share of voice is zero. If a competitor is named in half of them, theirs is not.

We audited two independently operated websites inside the same partner network. One scored 33 overall. The stronger site scored 53, with 83 for crawlability, 82 for AI readiness, and 80 for brand resolution. Its structured data was present but incomplete at 57.

Despite that stronger foundation, both sites recorded zero prompt share of voice. Their AEO scores sat at 25 and 26, and their citation scores at 49 and 45.

That pattern is not a quirk of two sites. Across all nine domains Radar audited between July 1 and August 2, 2026, taking the latest audit for each, every single one recorded zero prompt share of voice. Eight of the nine scored 70 or higher on crawlability. Average crawlability was 80 and average AI readiness was 79, while the average AEO score was 28 and the average citation score was 47.

Three caveats, because a number like that deserves them. First, every domain was measured on five category prompts across four engines, so this is 180 separate chances to be named and zero mentions, not one unlucky query. Second, zero is not a floor in the metric: Pixelmojo's own audits in the same window returned a non-zero share, so the scale does register a mention when one exists. We excluded ourselves from the aggregate so the figures describe the client set rather than us. Third, this sample is self-selected. People who run an AI visibility audit are disproportionately people who already suspect they are invisible, so read this as the shape of the problem among companies looking for help, not as a market-wide rate.

The takeaway is not that technical readiness failed. It made the stronger sites easier for systems to access, interpret, and distinguish, and that work is a precondition. But accessibility alone did not translate into visibility inside AI-generated recommendations.

This is also why a composite number needs a caveat, including ours. Radar reports a single AI Readiness Score that averages the dimensions which completed for your domain, and those dimensions span both readiness and visibility. A composite is only safe when the layers underneath it are reported separately and you can open each one. Radar splits them explicitly into an infrastructure group and a monitoring group for exactly that reason. The danger is not the average. It is an average with nothing inspectable beneath it, because that lets a rising headline number coexist with a recommendation share still sitting at zero. Ask any vendor, us included, to show you the layers before you accept the score. We unpacked that failure mode in AI Monitoring vs AI Technical Readiness.

The point is not to replace one black-box score with another. Every finding should lead back to evidence you can inspect: the prompt, the response, the cited source, the affected page, and the recommended action.

Ask them: "Show me the baseline, the prompt set, and the evidence behind your visibility score."

The green flag: They can demonstrate what is happening before asking you to pay them to change it.

3. Can They Create Content Your Competitors Could Not Publish?

Content remains central to search visibility, but publishing more content is not the same as becoming more authoritative. The test is whether a competitor could publish the same article without your experience or your data.

An article titled "10 Tips for Choosing Accounting Software" can be produced by almost any company, or generated in seconds by a model. It gives a retrieval system little reason to treat one brand as more useful than another.

A stronger article draws on implementation experience to explain which requirements construction companies commonly overlook, where multi-location approval workflows break down, how different accounting platforms handle job-costing, what integration problems appear during implementation, and which questions buyers should ask before requesting a demonstration.

That content carries experience, evidence, and judgment a competitor cannot easily reproduce.

Google's own self-assessment guidance points the same direction. It asks whether content provides original information, reporting, research, or analysis; whether it provides substantial value compared to other pages in search results; whether it is mass-produced or spread across a large network of sites so that individual pages do not get care; and whether it demonstrates first-hand expertise and depth of knowledge (Google Search Central, Creating helpful, reliable, people-first content).

In practice, citable content combines four qualities.

Evidence

Original research, documented tests, methodologies, benchmarks, customer observations, and clearly sourced facts. This is the layer most content programs skip, and it is the layer retrieval systems reward. We made the full argument in AI Visibility Is an Evidence Architecture Problem.

Expertise

First-hand lessons, subject-matter commentary, implementation detail, and informed opinion that goes beyond summarizing existing articles.

Clarity

Direct answers, useful definitions, descriptive headings, appropriate tables, and explanations that make complex information easier for people to understand.

Trust

Visible authorship, publication and update dates, supporting sources, transparent limitations, and consistent company information across the web.

This is not an argument for writing mechanically for the machine. It is an argument for making your information useful, understandable, and verifiable.

It is also where a specific sales pitch should worry you. Google's AI optimization guide states plainly: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them" (Google Search Central, AI optimization guide, page last updated 2026-07-10, verified August 2, 2026). Other platforms and crawlers handle such files differently, so maintaining one can still serve platform-specific purposes. We covered the nuance in Google Says You Don't Need llms.txt.

The red flag is not the file. It is anyone selling llms.txt, special AI markup, forced content chunking, or any other isolated tactic as a universal shortcut into AI answers.

Ask them: "Show me one piece of content you would create that our competitors could not publish without our experience or data."

The green flag: Their content plan is built around your evidence and expertise, not a quota of generic articles.

4. Is Their Engagement Model Something You Can Inspect?

A sound methodology still fails when the working relationship is vague. Before signing, you should be able to describe the first 90 days without the agency in the room.

AI search platforms change quickly. Prompt behavior shifts, retrieval sources change, and a brand that appears consistently one month may disappear or be described differently the next. Your agency needs an operating model for detecting and responding to that.

Establish these before you sign:

• What will be measured before any work begins
• Who creates and approves the prompt set
• Who owns the prompts, reports, response data, and accounts
• Which changes will be prioritized during the first 30, 60, and 90 days
• How frequently results will be reviewed
• How recommendations and implemented changes will be documented
• What happens when a platform changes its retrieval behavior
• Which outcomes define success

Reporting should include more than a monthly score. It should show what changed, what the team did, what evidence supports the recommendation, and what will be tested next.

You should also retain access to the underlying data. A client should not lose its measurement history, prompt library, or strategic learning because an agency relationship ended.

Ask them: "Walk me through the first 90 days, including the baseline, deliverables, reporting cadence, and what we will own."

The green flag: They can describe exactly how the engagement works before asking you to commit.

What Are the Red Flags When Hiring a GEO Agency?

Green flags tell you what good looks like. Red flags disqualify faster. Five are worth treating as hard stops.

The pattern connecting all five: each one moves the evidence out of your reach. A guarantee replaces measurement with a promise. A content quota replaces judgment with volume. An unexplained score replaces a trail with a number. Anything you cannot audit, you cannot manage.

Your GEO Agency Evaluation Checklist

Before selecting a partner, make sure you can answer four questions with evidence rather than assurances:

1. Integration. Is GEO connected to the company's broader SEO, content, and authority strategy, or sold as a separate product?
2. Baseline. Will you receive a documented baseline supported by a real prompt set and inspectable responses?
3. Originality. Is the proposed content based on knowledge or data genuinely specific to your organization?
4. Operating model. Are the first 90 days, reporting process, success measures, and ownership terms clearly defined?

A serious GEO partner satisfies all four. A partner who satisfies three and waves at the fourth is usually strongest at selling and weakest at proving.

One practical move before any of those conversations: run the technical baseline yourself. It costs nothing, it takes about a minute, and it changes the meeting. When you already know your crawler access, structured data, and AI readiness position, you can tell within minutes whether an agency is describing your site or describing a template.

The Bottom Line

No agency can guarantee what an AI platform will say about your company. That is the honest starting point, and any pitch that skips it is selling something else.

What a strong partner can do is make your brand easier to discover, easier to understand, more credible to reference, and less likely to be misrepresented. They can measure the result, show you the evidence, and respond when the environment changes.

That requires four connected capabilities: durable SEO foundations, transparent measurement, original content, and a disciplined operating model. Everything in this guide is a way of testing whether those four exist before money changes hands.

Start With Evidence, Not a Pitch

Before booking an agency call, run a free Radar check to see how AI crawlers access and interpret your site. Use the findings as a baseline for your internal team, or take them into your next agency conversation and see whether the agency's read matches the data.

The free check covers six technical readiness tools. It will not tell you what ChatGPT says about you, and no honest tool claims otherwise. It will tell you whether your foundations are the problem, which is the first question any credible partner should be answering anyway.

Ready to evaluate a partner with evidence instead of assurances?

• Run the free check - Six technical readiness tools, about 60 seconds, no card
• See how Radar scores - The full measurement methodology, including what we do not measure
• Talk to us - Bring your baseline and we will tell you what we would fix first