GEO for Agencies: How to Offer AI Search Optimization as a Service

This guide is an operating model, not a market-rate survey. Public evidence does not support universal claims about what percentage of agencies offer GEO, what a GEO retainer should cost, or how quickly a client will improve. The useful question is narrower: can your agency define a repeatable measurement method, identify work tied to that evidence, and price the delivery without promising an outcome you cannot control?

1. Start With the Measured Change in Search

The strongest reason to add AI search measurement is observable behavior, not a forecast. Pew Research Center analyzed 68,879 Google searches from 900 U.S. adults in March 2025. AI summaries appeared on 18% of those searches. People clicked a traditional result on 8% of visits with an AI summary, compared with 15% of visits without one. Only 1% of visits with a summary produced a click on a cited source. The study covers Google and a defined observation window, so it should not be generalized to every engine or audience. It does show why rankings and clicks alone no longer describe the whole search experience. See the Pew methodology and findings.

For an agency, that creates a concrete reporting gap. A client can rank in search while remaining absent from the answer layer above it. The service opportunity is to measure that layer across the prompts and engines that matter to the client, then connect changes to source citations, referral sessions, and downstream conversions without claiming that visibility caused revenue.

Qualify the service before you sell it

Write down six decisions before quoting a project:

• Prompt scope: Which buyer questions, brands, markets, and languages will be measured?
• Engine scope: Which answer engines belong in the panel?
• Sampling: How often will each prompt run, and how will repeated answers be summarized?
• Deliverables: Is the work limited to measurement, or does it include technical fixes, content, and source-layer distribution?
• Attribution: Which referral, conversion, and assisted-conversion data can the client share?
• Decision rule: What evidence triggers an action, a hold, or a change in direction?

This scope is commercially useful because it makes the offer checkable. It also prevents an agency from selling a vague promise to "rank in ChatGPT" when answers vary by engine, prompt phrasing, run, and time.

2. Use a Five-Phase Delivery Model

Phase 1: Establish the baseline

Build a prompt panel from real buyer questions, sales-call language, support tickets, and search data. Run the same panel across the contracted engines. Record brand mentions, citations, named competitors, position when observable, sentiment, and the exact answer timestamp. Keep raw responses so the client can inspect the evidence behind a summary metric.

A single answer is not a baseline. The 2026 preprint Don't Measure Once explains why probabilistic answer systems require repeated measurements across runs, prompts, and time. Treat visibility as a distribution. Report the sample size and collection window next to every rate.

Phase 2: Diagnose the gap

Separate each losing prompt into two response paths:

• Better-page gap: A competitor's guide, product page, or documentation answers the question with clearer facts, structure, or first-party evidence. Improve the client's closest existing page.
• Off-page authority gap: The answer cites review sites, community discussions, journalists, or industry roundups. The work is an earned-source plan, not another owned blog post.

Benchmark the actual cited pages. Compare their opening answer, named facts, update signal, source quality, table structure, visible FAQs, and structured data with the client's page. That diff becomes the action list.

Phase 3: Implement bounded changes

Prioritize pages tied to commercial prompts and change one coherent evidence unit at a time: a direct definition, a current product fact, a sourced table, a technical correction, or a clearer entity description. Record the page version and deployment date so later measurements have a usable before-and-after boundary.

Keep schema claims precise. Google's current guidance says that AI Overviews and AI Mode use the same SEO foundations as Search. It requires no special schema.org markup or AI text file. Structured data should match visible content, while crawlability, indexability, internal links, textual availability, and people-first content remain the foundation. Read Google Search Central's AI features guidance. Do not promise a citation because a page gained FAQ or Article markup.

Phase 4: Monitor the panel

Continue the contracted prompt and engine schedule. Track mention rate, citation rate, top-three rate, share of voice, and competitor pressure using disclosed formulas. Annotate launches, page changes, major campaigns, and measurement outages. Alerts should identify an observed change and its sample, not declare a cause.

Phase 5: Report decisions

A client report should answer four questions: what changed, how confident are we, what evidence explains the change, and what will we do next? Lead with verified wins and material losses. Include the prompt, engine, date range, denominator, cited URLs, and affected page for each recommendation.

Keep visibility, traffic, and revenue in separate rows. A mention is an answer-layer outcome. A referred session is a traffic outcome. A qualified lead or sale is a business outcome. Showing all three builds a useful attribution chain without turning correlation into a guarantee.

3. Price the Scope, Not the Category Hype

There is no defensible public benchmark for a "typical" GEO audit or retainer. Published agency price pages mix different prompt counts, engines, countries, content volumes, meeting schedules, and distribution work. Use them as competitor packaging research, not as proof of a market rate.

Here is an illustrative calculation, not a market benchmark. If a defined monthly scope requires eight delivery hours at a $150 loaded rate plus $200 in platform and reporting cost, delivery cost is $1,400. At a chosen 50% gross-margin target, the price floor is $2,800. A different prompt count, engine panel, content volume, or loaded rate produces a different answer.

Three workable commercial models

4. Choose Tools That Support the Evidence Chain

Agency tooling should connect diagnosis, monitoring, improvement, and delivery. A disconnected rank chart is not enough. Evaluate the stack against the client journey below.

• Technical Audit: crawlability, search readiness, performance, security, accessibility, and shareable evidence.
• Multi-engine monitoring: prompts, mentions, citations, competitors, sentiment, and trends with raw-answer access.
• Improvement workflow: prompt discovery, recommendations, content briefs, actions, and page history.
• Observed discovery: AI crawler and AI referral tracking to complement answer monitoring.
• Delivery: exports and developer access that fit the agency's reporting stack.
• Tenant boundaries: clear workspace, brand, member, and client-access rules.

Foglift links those layers in one product. Every plan includes unlimited five-dimension Technical Audits with shareable results and PDF export. Launch starts at $49 per month and adds five-engine monitoring, three brands, three team members, API access, CLI, MCP, and webhooks. Growth adds ten brands, ten team members, Buyer-intent Win Rate, and sitemap scanning. Foglift also includes Prompt Discovery, Query Fanouts, Watched Pages, Knowledge Bases, Ask Foglift, recommendations, content briefs, AI Crawler Analytics, and referral tracking according to plan entitlements.

Foglift has a Client Portal connection for client-facing reporting. Its per-plan allowance and access terms are not yet published, so confirm the current contract before including it in a client package. Verify any white-label reporting or Client Portal requirement separately. The product currently promises shareable Technical Audit results and PDF export, not a published white-label entitlement.

For a platform-by-platform review, see the AI visibility tools for agencies comparison.

5. Pitch the Evidence You Already Collected

Lead with the client's measured answer, not a category forecast. A useful pitch has four parts:

1. Show the exact buyer prompt and current answer.
2. Name the cited competitors and sources.
3. Explain whether the gap is owned-page quality or off-page authority.
4. Offer a bounded scope that will test a specific response.

A pilot should validate a method, not guarantee a lift. Define the panel, sample size, work delivered, review date, and decision rule in the proposal. If the answer layer does not move, the report should still explain what was tested, what the evidence ruled out, and what happens next.

6. Build Case Studies That Survive Scrutiny

A credible case study lets a reader reproduce the comparison. Capture the baseline before implementation and preserve the raw answers. Use the same engines, prompt set, inclusion rules, and rate formulas at the follow-up point.

Avoid invented before-and-after examples. A score jump or citation gain belongs in public copy only when it comes from a real, permissioned case with a disclosed method. If the client cannot be named, describe the anonymization rule and retain the underlying audit for internal verification.

7. Answer Client Objections Without Overclaiming

"We already do SEO. Why add this?"

SEO remains foundational. Google explicitly says the same search requirements and best practices apply to its AI features. The added job is measurement across answer engines, citation-source analysis, and a response workflow when the client's brand is missing. Show the client where existing SEO reporting stops and the answer-layer panel begins.

"AI search is too new to budget for."

Use the client's own baseline and the bounded Pew findings. Do not rely on a market-size forecast. If decision-stage prompts already produce AI answers that name competitors, the client can make a scoped measurement decision now. If they do not, monitoring may remain a lower priority.

"How will we measure ROI?"

Agree on an attribution ladder before work starts: answer visibility, cited-source visits, qualified actions, pipeline, and revenue. Use tagged landing pages, referral classification, analytics events, and CRM fields where available. Visibility supports the first step. It does not prove the later steps by itself.

"Can our team do this in-house?"

Yes, if the team can own the measurement design, monitoring stack, source analysis, implementation queue, and reporting cadence. Compare those capabilities with the agency scope. The decision should turn on capacity, expertise, and opportunity cost, not an unsupported claim about universal ramp time or tooling cost.

"What if the engines change?"

Expect provider and answer changes. Keep durable SEO fundamentals, visible source quality, and accurate structured data in the base layer. Keep engine-specific measurement rules versioned so the agency can explain discontinuities rather than hiding them.

"Our market is too niche."

Test that premise with real category prompts. A niche can have strong buyer questions, sparse answers, no answer layer, or a source set dominated by associations and documentation. Each result calls for a different decision. Market size alone does not predict how often a brand will appear.

Track the Practice Without Fake Benchmarks

Track revenue, gross margin, delivery hours, renewal, expansion, proposal acceptance, time to first measured result, and client evidence coverage. Set internal targets from your own baseline and capacity plan. Do not borrow universal thresholds from an unsourced article. A low renewal rate can reflect weak outcomes, poor fit, pricing, or reporting; the metric starts an investigation rather than supplying its conclusion.

A Four-Stage Internal Rollout

The sequence below can fit a month, but the calendar is illustrative. It does not guarantee a signed pilot, a visibility gain, a case study, or revenue by a particular date.

Build competencies before promises

An agency does not need a new job title before it can test the service. It does need accountable owners for measurement design, technical review, content or source-layer work, analytics, and client reporting. Run an internal pilot, have a second person audit the calculations, and require evidence review before sales copy or a client report goes out.

Create a dedicated service page only after those boundaries are clear. Explain the prompt and engine scope, the work delivered, what the reports measure, and what the agency does not guarantee. A precise offer is easier to trust and easier for an AI answer to quote than a page built around unsupported market claims.

Ready to test the workflow? Run an AI Visibility Check on your agency site, then review Foglift's current plan limits for the monitoring and developer-delivery scope you need.

Frequently Asked Questions

Sources & Further Reading

• Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results", July 22, 2025. Analysis of 68,879 Google searches from 900 U.S. adults.
• Google Search Central, "AI features and your website", updated December 10, 2025. Official eligibility, SEO, structured-data, and measurement guidance.
• Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024. Foundational controlled research on content visibility in generative-engine responses; its experimental gains are not organic ranking guarantees.
• Schulte, Bleeker, and Kaufmann, "Don't Measure Once: Measuring Visibility in AI Search (GEO)", arXiv preprint, April 2026. Measurement guidance for variation across runs, prompts, and time.