How Do You Vet a GEO Agency? Google Published a Three-Question Framework in June 2026

Most guidance on choosing an agency is written by agencies, which limits how useful it can be. This framework is different: it comes from the platform itself, it is published at a URL you can send to a vendor, and it is deliberately neutral about who provides the service.

The context makes it more significant. GEO has grown fast enough to attract both serious practitioners and opportunists, and until mid-2026 a buyer had no authoritative reference to distinguish them. Google's June documentation changed that by embedding GEO vetting into the hiring guide — the document businesses consult when deciding whether to engage an SEO at all — rather than leaving it as a technical footnote.

The question Google poses is worth quoting directly:

"If an SEO has advice on optimizing for AI experiences (also known as 'AEO' 'GEO' services), is their advice aligned with Google Search's official guidance on optimizing for generative AI features?"

That single sentence gives you permission to ask a question that previously sounded confrontational.

What to ask: for each significant recommendation in the proposal, which official Google documentation supports it?

What good looks like: a credible vendor grounds claims in developers.google.com/search rather than in proprietary "AI ranking factors" that cannot be verified. If a pitch cannot point to a documented basis for a recommendation, that recommendation is opinion — which is not automatically wrong, but should be labelled as such.

The red flag: claims about how AI systems rank or select content that are presented as fact but sourced only to the vendor's own analysis. Observation-based inference is legitimate and often valuable; presenting it as verified mechanics is not.

A fair caveat: not everything worth doing is in Google's documentation. Multi-platform work covering ChatGPT, Perplexity and Claude sits outside Google's guidance entirely, and honest vendors will say so rather than dressing it up as Google-endorsed. The test is not "is everything documented" but "does the vendor distinguish between what is documented and what is inference."

What to ask: cross-check the specific tactics on offer against Google's May 2026 AI optimisation guide.

Google names six tactics in that guide as unnecessary or problematic for its AI features:

• Tactic: llms.txt / AI-specific files | Google's position: Not needed — "you don't need to create new machine readable files, AI text files, markup, or Markdown"
• Tactic: Content chunking | Google's position: Not needed — systems understand multiple topics on one page
• Tactic: AI-specific rewriting | Google's position: Not needed — "you don't need to write in a specific way just for generative AI search"
• Tactic: Special AI schema | Google's position: Not required for AI features (structured data still earns rich results in regular Search)
• Tactic: Manufactured brand mentions | Google's position: Warned against
• Tactic: "Approved by Google" claims | Google's position: Warned against — Google does not evaluate third-party services

What good looks like: a proposal whose core deliverables are not built on this list.

The red flag: a roadmap centred on llms.txt generation, content chunking or special AI schema. These are things Google has publicly said do not help with its AI features, which means you would be paying for work the platform itself calls unnecessary.

The important scoping caveat: this list applies to Google's AI features only. ChatGPT, Claude and Perplexity use different retrieval models and may respond differently. A vendor recommending llms.txt specifically for AI agents or documentation tooling is on defensible ground; a vendor recommending it for Google AI Overviews is not. The distinguishing question is: which engine is this recommendation for? Vendors who conflate all AI platforms into one undifferentiated "AI optimisation" are over-generalising.

What to ask: what data is your "AI visibility score" built on, and is it presented as a measurement or an estimate?

Google's stated position:

"Third-party tools don't have access to our internal ranking data. They can't guarantee performance. Any predictions are their own and like predictions generally, may not happen."

This is general language covering every optimisation tool — established SEO platforms and newer AI-visibility products alike. No external tool sees Google's ranking signals, AI grounding logic, or internal scoring.

What good looks like: a vendor who says plainly that their tracking is observational — we run these queries, we log what appears — and presents scores as estimates for directional benchmarking.

The red flag: any promise of a specific number. "We'll get you cited in X% more AI Overviews" or "guaranteed AI Mode placement" is a claim Google says is structurally impossible for an outside tool to substantiate. Treat a guarantee as a disqualifier rather than a reassurance.

There is a fourth question you can now ask, thanks to a change that arrived alongside the documentation.

On 3 June 2026, Google began rolling out generative AI performance reports in Search Console, initially to a subset of site owners and expanding outward. These surface impressions, pages, countries, devices and dates for content appearing in AI features. One significant limitation at launch: click data is excluded — you can see that you appeared, not what it earned.

Why this matters for vendor selection: until now, AI visibility was measurable only through third-party estimates. Search Console reporting is the first first-party measurement layer, which makes AI visibility reportable against Google's own data rather than a vendor-defined metric.

The question to ask: will you use Search Console's AI performance reports as a source of truth once available to us, and how will you reconcile them with your own tracking? A vendor who intends to integrate honestly with Google's reporting is in a different position from one competing with it.

Three practical notes for Singapore businesses running this audit.

Ask which platforms are covered. Google's guidance covers Google's surfaces. If your buyers use ChatGPT and Perplexity — and for Singapore B2B, many do — you need a vendor who tracks those separately and says honestly that Google's guidance does not apply there.

Ask about local specifics. Google's guide names Google Business Profile and Merchant Center as levers for AI visibility in local and shopping contexts. For a Singapore local service business, a vendor who does not mention Business Profile is missing something Google explicitly flagged.

Watch for the grant angle. If a vendor is positioning services as PSG-eligible, verify their pre-approved status directly on the GoBusiness portal rather than taking it on trust — and remember EnterpriseSG requires companies to apply directly and charges no application fee. A vendor charging a fee to "secure" a grant is a separate red flag from anything in Google's framework.

A closing note on how to use this. No credible vendor should object to being asked these three questions — the standard is public, neutral, and authored by the platform. A vendor who becomes defensive when asked to align their recommendations with Google's published documentation is telling you something useful, at no cost to you.

Q1: What are Google's three questions for evaluating a GEO or SEO vendor?

First, does the advice cite official Google documentation rather than unverifiable proprietary "AI ranking factors"? Second, is their AI-specific advice aligned with Google's May 2026 guide on optimising for generative AI features — or does it rely on tactics Google explicitly lists as unnecessary? Third, do their tools acknowledge that no third party has access to Google's internal ranking data, and present scores as estimates rather than measurements? These come from Google's 5 June 2026 update to its "Do you need an SEO?" hiring guide and its new page on third-party SEO tools, services and advice — making them a free, public standard any business can apply.

Q2: Which GEO tactics should make me question a proposal?

Six that Google names directly: llms.txt and other AI-specific files, content chunking, rewriting content specifically for AI, and special AI schema are all stated as unnecessary for Google's AI features; manufactured brand mentions and any claim of being "approved by Google" are actively warned against. If a proposal's core deliverables are built on this list, you would be paying for work Google has publicly said does not help with its AI features. The one important nuance: this applies to Google's surfaces only, so a vendor recommending llms.txt specifically for AI agents or documentation tooling — rather than for Google AI Overviews — is on defensible ground.

Q3: Can any agency guarantee AI citations or AI Overview placement?

No. Google's own documentation states that third-party tools have no access to its internal ranking data, cannot guarantee performance, and that any predictions are the vendor's own. No external tool sees Google's ranking signals, AI grounding logic or internal scoring, which means an "AI visibility score" is modelling rather than measuring. Any pitch promising a specific number — a percentage lift in citations, guaranteed AI Mode placement — is making a claim Google says is structurally impossible for an outside party to substantiate. The honest version is probabilistic: strong fundamentals improve the odds.

Q4: How can I measure AI visibility independently of what my agency reports?

Two ways. Google began rolling out generative AI performance reports in Search Console from 3 June 2026, showing impressions, pages, countries, devices and dates for content appearing in AI features — the first first-party measurement layer, though click data is excluded at launch. Independently of any tool, you can also run your ten most commercially important queries monthly across ChatGPT, Gemini, Perplexity and Google AI Mode, logging whether your business is named and whether the description is accurate. That takes about twenty minutes, costs nothing, and gives you a check against whatever your vendor reports.

Q5: Does Google's vendor framework cover ChatGPT and Perplexity as well?

No, and asking a vendor about this scoping gap is itself a useful test. Google's guidance applies to its own AI features — AI Overviews and AI Mode — which are rooted in Google's core Search ranking systems. ChatGPT, Claude and Perplexity use different indexing and retrieval models and may respond to factors Google dismisses. An honest vendor will distinguish clearly between recommendations grounded in Google's documentation and recommendations for other platforms based on their own observation. A vendor presenting all AI optimisation as one undifferentiated discipline, or implying Google's guidance covers all platforms, is over-generalising.

Mayson AI works with Singapore businesses on SEO and AI search visibility, and we are happy to be measured against Google's published standard — including on the parts where our recommendations are observation rather than documentation. If you want a second opinion on a GEO proposal you have received, book a consultation.

Quotations are from Google Search Central: "Do you need an SEO?" and "Google Search's guidance on using third-party SEO tools, services, and advice" (both updated 5 June 2026), and the AI optimisation guide (15 May 2026). Verify current wording at developers.google.com/search.