Buy a testable service, not a guaranteed recommendation
An AI search visibility agency should explain what it will change, how it will measure the result and which outcomes remain outside its control. A proposal that promises to make your business “the answer” without defining the platform, question and testing conditions is not a usable scope.
Start with a business problem: prospective buyers cannot find your product documentation, answers misstate your services, or competitors appear in relevant comparisons. These problems require different work. Browsing AEO and GEO agencies is more useful once you know which problem you are buying help with.
What AEO and GEO actually mean
Answer engine optimization, usually called AEO, describes work intended to help content answer questions directly. Generative engine optimization, or GEO, describes work intended to improve how a business or its content appears in generated answers. Agencies use these labels differently; require the proposal to translate them into deliverables.
Neither label establishes a separate technical shortcut. Google says its established SEO practices remain relevant to AI Overviews and AI Mode, with no additional requirements or special optimizations needed for inclusion, according to its guidance on AI features and websites.
Distinguish the outcomes:
- **Mention:** the answer names your business.
- **Citation:** the answer links to a source associated with your business.
- **Recommendation:** the answer presents your business as a suitable choice.
- **Visit or conversion:** someone reaches your site or takes a measurable action.
Ask agencies to report these separately. A citation to a troubleshooting article is not the same result as a recommendation to buy your product.
What the evidence says about clicks
Visibility and website traffic are not interchangeable. In its observational study of US browsing behavior, Pew Research Center found different click patterns on Google pages with and without AI summaries:
- Traditional-result clicks occurred on **8%** of visits with an AI summary, according to Pew’s study.
- Traditional-result clicks occurred on **15%** of visits without an AI summary, according to Pew’s study.
- Browsing sessions ended after **26%** of visits with an AI summary, according to Pew’s study.
- Browsing sessions ended after **16%** of visits without an AI summary, according to Pew’s study.
These are historical observations, not a forecast for your website or proof that summaries alone caused the differences. They concern Google browsing, not every AI assistant. Use them to challenge any proposal that treats more AI exposure as an automatic increase in traffic.
What an agency can reasonably promise
Technical access and implementation
A useful audit should document crawler access, indexability, internal linking, rendering and whether important information is available as readable text. Require an issue register with evidence, ownership and acceptance criteria—not an unexplained readiness score.
Crawler policies need platform-specific treatment. OpenAI distinguishes OAI-SearchBot, used for search discovery, from GPTBot, associated with potential model-training use, in its crawler documentation. A proposal should explain which access settings it recommends and why, rather than treating search access and training permission as identical.
Require approval before changing crawler rules. Passing an access check should mean the technical change worked, not that a citation is guaranteed.
Clearer, better-supported content
A content scope should identify the buyer questions to address, the evidence needed and who approves factual claims. Useful work might include clearer service boundaries, product specifications, comparison criteria, original research methods and accessible support documentation.
Ask content marketing agencies to show the source material behind their proposed answers. Restructuring accurate information can be useful; mass-producing pages that repeat unsupported claims is not a substitute for expertise.
Consistent information beyond your website
Scope corrections to inaccurate business descriptions, outdated product details and conflicting public information. If external coverage is included, specify the research and outreach work separately from publication outcomes.
When comparing public relations agencies, ask whether the engagement buys outreach, sponsored placement or earned editorial coverage. Do not accept an external mention as proof that an AI system will use it.
How to make the results testable
Agree on the test before making changes
Build the question set from actual sales conversations, support requests and buyer research. Include discovery, comparison and factual-verification questions. Keep branded questions separate from questions where the buyer has not named your business.
For each observation, record:
- The exact question and any preceding conversation.
- The platform, mode and test date.
- Language, location and account state where observable.
- The complete answer and its linked sources.
- Whether your business was mentioned, cited or recommended.
- Whether the description was accurate and relevant.
Keep the agreed questions stable during the comparison period. Label new questions separately so that changing the test cannot quietly improve the reported result.
Separate observations from business outcomes
Require reporting in distinct layers: completed work, observed answer changes, and website or commercial outcomes. Preserve unsuccessful tests as well as favorable examples.
If an agency reports a visibility rate, demand the underlying observations, denominator and scoring rules. The result describes its monitored question set—not your share of every AI answer. Repeat observations under documented conditions and acknowledge variation rather than presenting a screenshot as a durable ranking.
Do not accept blended organic traffic as a clean measure of AI-search performance. Google says traffic from its AI features is included in the overall Web reporting in Search Console in its AI search documentation. Ask the agency to identify what is directly observable, inferred or unavailable.
Where feasible, compare changed content with similar unchanged content. Record other influences, including campaigns, seasonality and website releases. This improves interpretation without pretending the agency has run a perfectly controlled experiment.
Promises that should stop the sale
Reject guaranteed placement, permanent citations, universal visibility and claims of a special submission mechanism without platform documentation. Google explicitly says compliance does not guarantee crawling, indexing or serving in its AI features guidance.
A guaranteed inclusion timetable is **not published** in that guidance. Do not let an agency replace that missing platform commitment with an unsupported deadline.
Also challenge:
- Reports containing only successful screenshots.
- “Share of voice” without a disclosed question set.
- Proprietary scores without definitions or exportable evidence.
- Claims that structured data guarantees recommendation.
- Attribution that assigns every new inquiry to AI visibility.
Put the buying decision into the contract
Use the same due diligence you would apply when choosing an SEO agency: inspect relevant work, interview the delivery team and establish ownership of content, accounts and reporting data.
The agreement should state the baseline method, implementation responsibilities, review process, reporting exclusions and handover requirements. Tie acceptance to inspectable deliverables and agreed testing—not outcomes controlled by a search provider.
Review our agency evaluation methodology when researching candidates, then request evidence specific to your website and market. The strongest offer is not a promise to control AI answers. It is a clear plan to improve the information available, test what changes and explain what the evidence cannot establish.