In short
- Ask finalists to demonstrate their work on a real buying question and a page you own.
- Require a reporting method that another reviewer can follow from the underlying answers.
- Reject guaranteed citations, fabricated reviews and undisclosed paid rankings presented as independent recommendations.
Define the assignment before comparing agencies
Start with a problem your team can describe without using the phrase "AI visibility." Buyers may misunderstand a product limitation. Your company may be absent when they ask for suppliers meeting a specific requirement.
Write that problem into the brief. Specify the market and who buys the product. Explain what your team can change, and what needs approval from another department.
Our recommendation is to compare agencies against this same assignment. A generic demonstration gives you little basis for judging whether the agency understands your constraints.
Separate advice from implementation in the scope. If you need someone to edit production pages, ask for that responsibility explicitly. If your team will implement recommendations, confirm its available capacity before buying a large backlog.
The guide to evaluating a GEO agency proposal develops this distinction through an illustrative product-page assignment. Use it to frame the work you want finalists to demonstrate.
Ask for a sample you can inspect
Give the finalist an existing page and a buyer question. Request a small worked example with source material. An anonymised example can also help, provided you can understand its constraints.
For a fictional scheduling product, the question might concern technicians working offline. Ask the agency to identify the missing explanation. It should flag unknown product facts for verification before drafting claims.
Read the proposed wording with someone who knows the product. Does it answer the buyer's question accurately? Can your team trace each important assertion to documentation or an accountable internal source?
Then examine the implementation instructions. The sample should identify the affected URL and the proposed change. Ask how the agency will check the published result, including whether the important text is accessible.
Google's AI guidance includes crawl access and textual content among its established recommendations. 1 An agency should explain how its technical work addresses a concrete issue on your site.
Ask what the sample leaves unresolved. A provider willing to identify a missing fact is easier to evaluate than one presenting every assumption as settled.
Examine the method behind a case study
A case study is useful when you can reconstruct what happened. Request the starting position and comparison period. Ask which pages changed and whether other campaigns ran during the same period.
Look for the underlying question set. Did it remain stable? Were questions naming the client reported separately? Which engine actually produced the answers?
Ask for the completed-answer count behind any visibility percentage. Request the counting definition and the treatment of failed runs. If the method changed, the comparison needs an explicit qualification.
Inspect a sample of answers selected from the reporting period. Include unsuccessful observations. Check whether the answer merely names the company or recommends it for the relevant task.
Treat citation charts with the same care. Bing explicitly says citation counts don't establish placement within an answer. 2 A chart of linked pages therefore needs further evidence to support a recommendation claim.
For traffic claims, ask for the acquisition source and relevant landing pages. For revenue claims, ask how the agency connected the opportunity to that source. Keep customer recollection distinct from analytics attribution.
An agency may have legitimate confidentiality limits. Request a redacted example or a permissioned reference instead. If neither is available, record the remaining uncertainty when making your decision.
Compare what the engagement will leave behind
Read the proposal as a description of work your team must be able to accept. Content volume alone is an incomplete basis for that decision.
| Area | Evidence to request | Decision it supports |
|---|---|---|
| Product understanding | A sample checked against product documentation | Whether the agency can represent the offer accurately |
| Implementation | A proposed change with a published-page check | Whether your team can release the work |
| Measurement | A fixed question panel with underlying answers | Whether progress can be reviewed consistently |
| External coverage | Proposed outlets with payment and selection details | Whether the coverage fits your audience |
| Handover | Sample exports and a description of retained files | Whether you can continue after the engagement |
Use the GEO proposal review sheet to record each finalist's response. Write the missing evidence beside the relevant decision. Avoid turning an unanswered question into an invented numerical score.
Confirm access requirements before signing. Specify who approves copy and who can publish it. Ask which work pauses if those people are unavailable.
Clarify whether measurement access belongs to your company or the agency. Request export formats before the engagement begins. Confirm how you will retrieve historical answers if you stop using its service.
Recognise promises the evidence cannot support
Treat a guaranteed citation or fixed AI ranking as a reason to stop the evaluation and request a precise explanation. Don't accept a screenshot as proof of control over future answers.
Google says meeting its requirements doesn't guarantee that content will be crawled, indexed or served. 1 A proposal promising guaranteed inclusion in Google's AI features contradicts that documented limit.
For other engines, require equally specific evidence rather than carrying a promise across platforms. Our buying recommendation is to reject guarantees of organic citations or recommendations.
An agency can commit to deliverables it controls. Examples include reviewing a defined set of pages or supplying a report with retrievable observations. Keep acceptance of that work separate from the outcome being investigated.
Scrutinise packages built around schema or special machine-readable files. Google says its AI features need no special schema or new AI text files. 1 That statement concerns Google, so preserve its scope.
Our standard is broader as a purchasing rule. We don't accept schema, FAQ formatting or llms.txt as promised citation levers. Nor do we accept added statistics or artificial content chunks on that basis.
Require a practical reason for each change. Structured data can serve a documented eligibility requirement. Google's product snippets, for example, require specific properties. 5 A short section can make a difficult explanation easier to read. Neither justifies an unsupported visibility guarantee.
Reject fabricated reputation and concealed placements
Ask how the agency intends to obtain reviews. Reject invented customer experiences and accounts presented as independent buyers. For Google Maps, paid reviews and fabricated engagement violate the platform's published policy. 3
Apply the same purchasing standard to community work. Require honest affiliation and compliance with the community's rules. Exclude vote manipulation and fabricated discussions from the assignment.
Paid lists deserve a direct question about who receives money. Ask whether payment affects inclusion or position. Identify any ownership connection between the agency and the publisher.
Our recommendation is to evaluate a disclosed sponsorship as advertising. Require its own audience rationale. Reject a paid ranking sold as an independent endorsement or as guaranteed access to AI answers.
Google's link-spam policy covers buying links to manipulate rankings. It distinguishes appropriately qualified advertising and sponsorship links. 4 Don't let a proposal blur those different practices.
Also reject hidden instructions aimed at manipulating answers and mass production of low-value pages. Google's spam policies address hidden manipulation and scaled content abuse. 4
Ask the provider to put these exclusions in writing. If it refuses to explain how coverage is obtained, the work cannot be evaluated properly.
Evaluate the agency again after work starts
At the first review, compare delivered work with the agreed sample. Confirm that the changes actually reached the public pages. Record anything awaiting your team's input.
Review the report against the AI visibility review sheet. Check the same questions and engines over comparable periods. Require visible denominators and retain observations that don't support the hoped-for result.
If the outcome is flat, ask for the next investigation and its rationale. A useful response might identify an inaccurate source or a measurement change. More pages should require a documented need.
Evaluate whether the agency corrects its own conclusions when evidence changes. Ask it to explain which observations remain uncertain. Keep implementation quality distinct from changes that coincide with the engagement.
Before renewal, revisit the original buying problem. Decide which unresolved issue justifies further work, and whether the current provider has shown it can investigate it.
Apply these questions to a Red-engage proposal too. Choose a team whose work you can inspect and whose conclusions you can challenge.
Sources
- Google Search Central, AI features and your websiteAccessed Back to the text, first mention of source 1
- Microsoft Bing, Introducing AI Performance in Bing Webmaster Tools Public PreviewAccessed Back to the text, first mention of source 2
- Google Maps, Prohibited and restricted contentAccessed Back to the text, first mention of source 3
- Google Search Central, Spam policies for Google web searchAccessed Back to the text, first mention of source 4
- Google Search Central, Product snippet structured dataAccessed Back to the text, first mention of source 5
