How to choose the best GEO agency: a practical scorecard

When you search for the best GEO agency, similar proposals can be hard to compare. A dashboard and confident promises do not tell you how the work will be done. The hard part is not finding agencies. It is comparing them on what they can prove, what they will actually do and what you will own at the end.

This article gives you a buyer scorecard for that comparison. It assumes you already know the basics of generative engine optimization; if not, read our practical GEO and AEO buyer guide first, which covers the questions to ask, the internal-team versus agency choice and what a quote should explain. Here we go one step further: how to score proposals, what proof to demand before signing and how to define acceptance at 30, 60 and 90 days.

A note on fairness. Auden AI is a GEO provider, so we have an interest in this topic. The scorecard is written to be applied to every provider, including us. “Best” here means best fit for your needs on evidence you can check, not a ranking of agencies.

Starting with an AI-generated shortlist? Our ChatGPT, Claude and Grok recommendation walkthrough shows real answer excerpts, selected source checks and the questions to resolve before using this scorecard.

Start with the decision, not the vendor

Before you score anyone, write down the decision the engagement must support. Three examples:

• “Find out whether AI answers describe our services accurately for our five most important buyer questions, and fix what is wrong on our site.”
• “Determine whether our product pages give AI services and buyers the facts they need to compare us, and close the gaps.”
• “Set up a repeatable observation process our team can run after the agency leaves.”

A written decision keeps proposals comparable. An agency that answers a different question, such as “increase AI visibility,” with no defined questions, services or markets, is harder to evaluate and easier to oversell.

The scorecard

Score each proposal against the eight criteria below. The weights are our suggestions for a typical first engagement. They have not been validated against outcomes, and you should change them to reflect your priorities. A team with strong internal developers might lower the implementation weight; a regulated business might raise data handling.

Use a simple scale for each criterion, then multiply by the weight:

• 0 – not addressed, or addressed only with assurances.
• 1 – described in general terms, without examples.
• 2 – described specifically, with a sample deliverable.
• 3 – demonstrated with a sample deliverable you could inspect and reproduce.

The maximum is 300. Treat the total as a structured conversation aid, not a verdict. Two reviewers should score independently and then discuss any criterion where their scores differ by two points or more.

Proof deliverables to request before you sign

The fastest way to separate method from marketing is to ask for artifacts, not adjectives. Request anonymized or synthetic samples of the following. A provider unable to share a sample should be able to explain exactly what you would receive.

• A raw observation record: one saved answer with the exact question, service, mode, date, time zone, context state and every cited URL.
• A claim-by-claim source check: claims from an answer, each labeled against its cited source as supported, partial, contradicted or unknown, with omitted caveats noted separately.
• A prioritized gap list: each item linking an observed problem to a proposed change, an owner and the evidence needed.
• A change log template: what changed, where, when, who approved it and why.
• A sample report that separates observations, completed work and business data, with limitations stated in the report rather than in a footnote.

See the full sample, not just the good answers

Any rate an agency reports needs a denominator you can see. A claim such as “mentioned in 40% of answers” means little until you know how many answers were sampled, from which services, on which dates, for which questions and under what settings.

Ask each provider to commit, in writing, to disclose:

• the full list of questions tested, including any that produced unfavorable results;
• the number of captures per question, per service and per period;
• how questions were chosen, and who chose them;
• whether captures were made with prior conversation context, signed-in accounts, or location or language settings;
• which captures were excluded and why;
• which tools or services were used to collect the data, and any usage fees passed on to you.

Selective reporting is an easy way for a small sample to become a persuasive slide. Insisting on the full denominator, including unfavorable results, is the simplest safeguard.

Three different things that are often blurred

GEO reporting often mixes numbers that answer different questions. Ask each provider to keep them apart.

What Google’s generative AI reporting shows

Google’s Generative AI performance report documentation, reviewed on 22 October 2026, describes impressions from AI Overviews and AI Mode, with page, country, date and device views. Google says the report rolled out worldwide on 31 August 2026. It may not appear when a site has too few impressions. It does not establish clicks, conversions, what every answer said or visibility in other AI services. Ask a provider to explain its date range and aggregation limits alongside any reported numbers.

Google also documents a separate Search generative AI inclusion control. Include is the default; child properties can inherit a parent setting. This is separate from AI training controls and should be reviewed through your organization’s normal approval process. We have not accessed any organization’s Search Console data for this article.

In a proposal review, ask the agency to label each metric with its source and its limits. Request a sample report that keeps platform impressions, captured answer observations and actual customer actions separate. A polished dashboard is useful only when someone can explain what its numbers count.

What else could explain a change?

The most tempting claim in GEO reporting is “we changed the page, and the answer improved.” Sometimes that is true. Often something else changed too. Ask each provider how it records and handles at least these other possible causes:

• model or product updates by the AI service;
• changes in the sources the service retrieves, including competitors’ pages and third-party reviews;
• seasonality, promotions and news events;
• differences in capture settings, such as account state, location, language and prior context;
• other marketing activity running at the same time;
• sampling variation: answers can differ between runs even when nothing has changed.

Stronger designs include repeated captures before and after a change, a set of comparison questions or pages that were not changed, and a pre-agreed statement of what result would count as no effect. Even then, a careful provider describes results as observed associations under stated conditions, not proof of cause.

Implementation ownership and change approval

Many GEO engagements stall at the point where findings meet the website. Agree in advance who does what.

Ask whether the agency recommends changes, implements them or both, and whether implementation is in the fee. Require a change log that records each edit, its approver and its date, so later observations can be read against what actually changed.

Cost and contract questions

The buyer guide lists what a quote should explain. When you reach contract stage, these additional questions are worth raising. This is not legal advice; involve your own legal adviser for contract terms.

• Who owns the observation records, claim-by-claim source checks and reports, and in what format will they be handed over at the end?
• Are third-party tool or usage fees passed through, and how are they capped?
• Are any subcontractors involved, and do they handle your data?
• What happens to your data and account access when the engagement ends?
• Can you terminate if agreed deliverables are not met, and what notice applies?
• Are any outcome claims in sales materials reflected in the contract? If not, treat them as illustrations, not commitments.

Be cautious of any proposal that guarantees rankings, recommendations or a specific number of AI citations. No provider controls how an AI service composes its answers.

30, 60 and 90-day acceptance criteria

Acceptance criteria should describe deliverables and process quality you can verify, not business results the agency cannot control. The examples below are a starting point to adapt.

Notice what is absent: a target for citations, rankings, traffic or revenue. You can and should track business measures in your own systems, but the agency’s acceptance should rest on work it can be held to.

Putting the scorecard to work

Send every shortlisted provider the same written brief and the same request for sample deliverables. Score independently, compare notes, then ask follow-up questions only where the evidence was thin. The best GEO agency for you is the one whose evidence is clearest, whose responsibilities are explicit and whose work your team can check and continue.

Apply the same standard to Auden. If you want to see a small example of separating what was fetched from what it means, our Auden AI Checker guide shows an open-source local tool that reports observations from one public page without claiming to predict rankings or AI recommendations.

How to read “best GEO agency” lists

Start with the publisher. Does it sell GEO services? Does it rank itself? A clear scoring method helps you review a list, but it does not remove the publisher’s interest in the result.

Check the review date, the evidence behind each score and what “best” means. Lowest price, hands-on implementation and enterprise reporting are different needs. Look for sources you can open. Keep a fictional sample separate from a completed client project. A useful list can help you build a shortlist; it cannot choose for you.

Asking ChatGPT, Claude or Grok to recommend an agency

Use an assistant to gather questions and candidates, then check its answer. For example: “Recommend a GEO agency for a US business. Explain the fit, name your sources and mark anything you cannot verify.” This is a suggested prompt, not a captured test result.

Save the full answer, exact question, date, displayed mode and cited links. Repeat in fresh conversations. If a name changes, record that too. Open the sources and check service scope, prices and evidence directly. Being named in one answer does not prove that an agency is best, or that other people will see the same answer. The prompt above is a suggestion, not a captured test. For real captures of a similar neutral question, see our 22 October walkthrough. Read its conditions first. One answer per product does not show how often any provider appears.

Your shortlist worksheet

Copy these fields into a document. Complete one record per provider using the same brief.

• Need: buyer question, market, priority pages and work your team needs help with.
• Provider and date: name, offer URL and date checked.
• Evidence: sample report, raw answer records and source checks you inspected.
• Responsibility: who researches, approves, publishes and checks the changes?
• Total commitment: service fees, tools, extras, minimum term and exit terms.
• Open questions: missing proof, unclear costs and claims to verify.
• Decision: shortlist, seek clarification or decline—and why.

Illustrative example: a proposal has a strong dashboard but no raw records. Mark the evidence field “not yet supplied” and request one complete sample. Do not award full points because the presentation looks polished.

To apply the same checks to our offer, read Auden’s GEO service scope, delivery process and pricing. Auden publishes this guide; this link is to our own services.

Sources and editorial notes

Google documentation reviewed 22 October 2026. The scorecard, weights and acceptance criteria are Auden’s suggestions; they are not validated against outcomes and do not guarantee any result.

1. Google Search Console Help: Generative AI performance report (Search)
2. Google Search Console Help: Search generative AI control
3. Auden AI: GEO and AEO practical buyer guide
4. Auden AI: Install Auden AI Checker from GitHub