How to Choose an AI Search Agency That Actually Delivers

Your sales team is already being mentioned in AI answers, but nobody can tell you whether the mentions are accurate, how often competitors appear instead, or whether any of it influences pipeline. Meanwhile, several agencies are offering “GEO,” “AEO,” and AI visibility packages that look different in presentation but often rely on the same unverified screenshots.

Choosing an AI search agency is therefore a measurement decision before it's a creative one. The right partner will define the prompts, preserve the raw evidence, separate genuine visibility from random model output, and connect answer-engine exposure to commercial outcomes. The wrong one will sell access to a dashboard, promise placement in ChatGPT, and leave your team debating vanity metrics.

Clarify Your AI Search Goals Before You Talk to Anyone

A marketing lead buried in AI-search pitches needs a one-page decision brief before taking another sales call. Start by naming the surfaces that matter to your customers, such as ChatGPT, Perplexity, Google AI Overviews, Microsoft Copilot, and Gemini in Workspace. Don't track every engine because a vendor supports it. A buyer researching software may use ChatGPT for category education, Perplexity for comparison research, and Google AI Overviews for a product-specific question. Those are different moments with different commercial implications.

Assign each surface a job

Map each engine to a buyer journey stage:

Category discovery: Track broad, non-branded questions where your company competes for recommendations.
Evaluation: Track comparisons, alternatives, use cases, and competitor-adjacent prompts.
Conversion support: Track branded questions, product details, pricing context, implementation concerns, and objections.
Post-sale confidence: Track support and trust questions that can reinforce or damage the customer relationship.

Then decide what outcome you're buying. Some teams want share of model, meaning the frequency and prominence of their brand across a defined prompt set. Others care about AI-referred sessions, branded search lift, or pipeline associated with AI-assisted discovery. An agency that sells content production may be a poor fit if your actual need is reproducible measurement across engines.

Practical rule: Choose one primary KPI and one guardrail metric before a vendor shows you its dashboard.

For example, make citation share against named competitors the primary KPI, with qualified AI-referred sessions as the guardrail. Or choose AI-assisted pipeline as the primary KPI and citation accuracy as the guardrail. Don't use a dozen “north-star” metrics. That gives the agency room to declare victory somewhere, regardless of business impact.

Write the scope before the pitch

Your brief should identify the audience, priority product lines, competitor set, engines, approval owners, content constraints, and technical dependencies. Assign one person to approve strategy, another to review claims and brand language, and a technical owner to provide site and analytics access.

Set a baseline citation share for your brand against named competitors before work begins. If you can't establish that baseline internally, ask the agency to make its prompt methodology and raw observations part of the proposal. The sales call should test whether the vendor can measure your problem, not whether it can describe the future of search.

Get Your Technical and Data Foundations in Order

An agency can improve weak foundations, but you shouldn't let it use basic technical debt to create an unlimited scope. Prepare the data and access checklist first, then ask the vendor to separate prerequisites from optimization work.

The pre-engagement checklist

Structured data: Confirm crawlable schema for Product, Organization, FAQ, and Author entities. Your SEO or engineering owner should validate that the markup matches visible page content and remains consistent across templates.

Crawl controls: Review robots.txt, sitemap coverage, canonicalization, rendering, and sensitive paths. A technical SEO owner should document what crawlers can access and what they shouldn't. Ask whether the agency expects an llms.txt policy, and use the LLMrefs llms.txt generator as one practical way to prepare a draft for review.

Logs and search data: Provide server logs, Google Search Console, and equivalent webmaster-tool access where appropriate. Your analytics or engineering team should clarify retention, permissions, and export procedures before the agency starts.

Referral measurement: Ensure GA4 or an equivalent analytics platform can identify referrals from chat.openai.com, perplexity.ai, gemini.google.com, and copilot.microsoft.com. Don't assume every AI visit will be cleanly attributed. Document known gaps and use dated answer captures alongside traffic data.

Content inventory: Tag pages by topic cluster, audience, commercial role, author ownership, evidence quality, and E-E-A-T signals. A content lead should identify which pages can be revised, consolidated, or retired.

Visibility baseline: Record branded and non-branded citation frequency in at least three answer engines. The measurement owner should preserve the prompt list, engine, interface, location, login state, device class, and collection date.

For data-heavy teams, the same discipline applies to external datasets. If your business depends on property or marketplace information, a guide to extracting Australian real-estate listings can help your technical team think through collection structure, consistency, and downstream data quality.

Flag missing prerequisites in the statement of work. Require a fixed remediation scope, named owner, acceptance criteria, and separate pricing for optional work. If an agency refuses to distinguish “required for measurement” from “recommended for growth,” it's protecting scope rather than protecting your budget.

Vetting Agencies With the Right Evaluation Questions

The AI search advertising market is moving from experimentation toward a significant commercial channel. Mordor Intelligence projects growth from USD 1.39 billion in 2025 to USD 3.25 billion in 2026, followed by USD 32.89 billion by 2031, with a projected 58.87% CAGR from 2026 to 2031 (Mordor Intelligence's AI search advertising market forecast). That expansion attracts serious operators, but it also attracts agencies repackaging conventional SEO with a new label.

eMarketer forecasts US AI search ad spending at USD 2.08 billion in 2026, representing 1.3% of total US search ad spend, and projects USD 25.93 billion by 2029, or 13.6% of overall search ad spend (eMarketer's AI search advertising analysis). The same forecast puts US AI ad spending at USD 32.03 billion in 2026, with more than 80% appearing beside AI content rather than inside chatbot conversations. Your agency should understand that near-term opportunity is concentrated around AI-enhanced search surfaces, not just chatbot prompts.

Ask for evidence, not theater

Ask how the agency measures share of model across engines. A credible answer includes the prompt taxonomy, collection conditions, refresh cadence, competitor definitions, raw exports, and rules for handling empty, conflicting, or hallucinated answers.

Ask what telemetry the team receives. Does it ingest analytics referrals, Search Console data, CRM outcomes, answer captures, and citation URLs, or does it present a proprietary score with no audit trail? Ask who performs the work, whether specialists are subcontracted, and who owns prompt design, technical implementation, content review, and reporting.

The agency should also explain its response to hallucinated brand information. Correction may require clearer first-party content, entity consistency, authoritative third-party references, and repeated validation. No vendor can guarantee a permanent placement in ChatGPT or any other generative interface.

Question to Ask the Agency	Red Flag to Walk Away
How do you calculate share of model across engines?	A single blended score with no engine-level results
Can we export raw prompts, responses, citations, and dates?	A proprietary dashboard that blocks raw-data exports
How often do you refresh the prompt set?	A static prompt list used indefinitely
What happens when an answer contains a hallucination?	A promise that content changes will guarantee correction
Who completes the work, and what is subcontracted?	Evasive answers about delivery personnel
Can we review a sample methodology document?	Refusal to share evaluation rules
How do you connect visibility to pipeline?	Reporting that stops at screenshots and mentions

If you're also evaluating broader workflow automation by AI agents, keep that work separate from answer-engine measurement unless the agency can show exactly how automation affects your tracked outcomes. For a wider vendor-screening framework, compare the questions in this search marketing agency evaluation guide, then insist each shortlisted provider answers in writing.

Designing a Pilot You Can Actually Trust

A pilot should behave like a controlled experiment, not a polished showroom. Limit the engagement to two or three priority query themes, one defined audience segment, and a named competitor set. A B2B software company might test “best tools for distributed finance teams,” “alternatives to Competitor A,” and “implementation requirements for enterprise finance software.” A property business might test buyer, seller, and suburb-specific themes.

Define the metrics before the agency edits a page:

Citation frequency: How often target answers cite your owned or earned sources.
Share of voice: How your brand compares with the named competitor set.
LLM referral traffic: Visits arriving from tracked AI surfaces.
Pipeline outcome: A CRM-linked stage such as qualified opportunity creation or influenced revenue.

The benchmarking method needs enough prompts to reduce noise. One independent guide recommends 75 to 150 prompts for a focused benchmark and 250 or more for a mature program, with balanced branded, non-branded, and competitor-adjacent intent (the repeatable AEO benchmarking methodology). For the pilot design described here, require at least 200 prompts per theme across ChatGPT, Perplexity, Google AI Overviews, and Gemini. That larger sample gives you a stronger read than a handful of handpicked screenshots.

Put guardrails around the result

Pre-register the baseline, evaluation prompts, collection conditions, minimum effect threshold, and decision rule. Decide in advance what qualifies as go, extend, or exit. If the agency changes the prompt set after seeing weak results, you no longer have a comparable test.

The Princeton GEO evaluation reported visibility improvements of up to 40% in generative engine responses, while emphasizing the need for controls and repeated measurements because query and variant output can vary (the Princeton GEO evaluation). Treat that finding as a reason to improve methodology, not as a promise for your brand.

Require raw prompt logs, dated answer captures, citation URLs, change history, and access to the tracking stack. Screenshot-only reporting is unacceptable. You should be able to reproduce the agency's score, inspect disagreements, and see whether a reported gain survives a held-out prompt set.

For a practical product and measurement walkthrough, request a demo of LLMrefs. The point isn't to outsource judgment. It's to make the agency's evidence inspectable.

Comparing Pricing Models and Engagement Structures

Pricing shapes behavior. A vendor paid only for hours has little incentive to simplify the program, while a vendor paid on a loosely defined “visibility lift” can choose the metric that makes its work look successful. Model the engagement around the risk you want the agency to own.

A monthly retainer works when you need continuous testing, content operations, technical coordination, and reporting. It rarely creates accountability by itself. Use a modest base fee for access and planning, then tie a meaningful portion to agreed outputs such as citation improvement, qualified AI-referred sessions, or validated prompt coverage. Don't pay for a metric the agency can inflate by changing the prompt set.

Project pricing fits a focused audit, schema rollout, entity cleanup, or content restructuring sprint. The statement of work should list templates, pages, acceptance criteria, dependencies, and revision limits. “AI optimization” is not a scope. “Audit these product templates, implement approved Product schema, and deliver validation notes” is.

Model	Best For	Risk Sits With	Watch Out For
Monthly retainer	Ongoing testing and coordinated execution	Mostly the buyer unless KPIs are explicit	Open-ended hours and vague deliverables
Project or sprint	A defined technical or content intervention	Shared, based on acceptance criteria	Scope creep and expensive change requests
Performance fee	Outcomes with reliable attribution	The agency only if measurement is independent	Branded-search definitions the vendor controls
Revenue share	Directly attributable commercial programs	Shared, with strict CRM rules	Disputes over source mix and attribution windows
Equity or barter	Rare strategic partnerships	Primarily the buyer	Lock-in without predictable delivery

Performance and revenue-share deals need exact attribution windows, source definitions, CRM rules, exclusions, and audit rights. If “AI-influenced lead” includes every lead who saw a branded answer at any time, the metric isn't useful.

Compare effective hourly cost, maximum cost per acquired AI-influenced lead, internal review time, and switching cost after six months. Equity, barter, and long exclusivity clauses usually create more lock-in than value for mid-market brands. The cheapest quote can become the most expensive once your team pays for missing data, repeated explanations, and reporting cleanup.

Contracts, SLAs, and the Red Flags That Should Kill the Deal

The contract should turn the agency's confident pitch into testable obligations. Start with the SLA. It must define response times, reporting cadence, named tools, escalation contacts, issue resolution, and the format of deliverables. “Increased visibility” belongs in a strategy document as an aspiration, not in an SLA as a service commitment.

Attach a measurement methodology exhibit. Lock the prompt set, engine list, evaluation criteria, collection conditions, data sources, refresh cadence, and calculation rules for the contract term. Add a right to audit the methodology and substitute your own tracking system at any time.

Protect ownership and leverage

Your company should own every prompt library, schema template, content asset, citation report, dashboard export, measurement script, and account-specific workflow produced for the engagement. Generic agency know-how can remain theirs, but work created from your data and paid for by your company shouldn't disappear when the contract ends.

Review these clauses with procurement and legal:

Renewal: Avoid auto-renewing terms longer than ninety days unless the renewal notice and exit process are clear.
Exclusivity: Reject language that blocks your in-house team or another specialist from working on adjacent SEO, content, analytics, or technical projects.
Termination: Don't accept early-termination penalties without reciprocal performance credits or a fair transition obligation.
Confidentiality: Cover your strategy, data, prompt sets, competitor research, and internal documentation, not only your brand name.
Exit package: Require raw exports, documented workflows, credentials transfer where appropriate, and a thirty-day transition window.

Contract test: If the vendor can't explain how you leave, it hasn't explained how you stay in control.

Watch for guaranteed ChatGPT placements, unexplained “AI authority” scores, and reports that show only favorable answers. Also reject a contract that permits the agency to alter the benchmark after launch. Answer engines change, but that's precisely why the agreement needs a documented change-order process, not unlimited discretion.

The agency should name the people responsible for strategy, technical work, content, analytics, and account management. If the sales strategist is the only visible expert and delivery is subcontracted without disclosure, you're buying a presentation rather than a team.

Running and Renewing the Engagement for Long-Term Wins

A successful pilot should produce a repeatable operating system, not a reason to approve an indefinite retainer. Set a monthly business review with a fixed dashboard covering AI citation share, prompt coverage, entity consistency, and attributed pipeline. Keep the definitions stable so a month-to-month change reflects performance rather than a reporting redesign.

Use a 90-day optimization cycle with distinct decision points:

Weeks 1 to 2: Audit regressions in content, schema, citations, entity references, and tracked answers.
Weeks 3 to 6: Ship approved content, schema, internal-linking, and authority improvements.
Weeks 7 to 10: Measure performance against a held-out prompt set, not only the prompts used to guide the work.
Weeks 11 to 12: Reallocate budget toward themes and surfaces that show validated commercial value.

Keep the source of truth in-house

The agency can run execution, but your brand should own the source-of-truth spreadsheet or database. Store prompt IDs, intent classifications, engine conditions, dated answers, citations, brand mentions, competitor mentions, page changes, and CRM joins. The agency submits deltas and recommendations. Your team approves the record.

Renewal should depend on more than a visibility graph moving upward. Ask whether the agency found prompt categories you weren't tracking, improved answer accuracy, defended visibility through a model or interface change, and trained internal staff to inspect the data. An agency that creates permanent dependency has delivered a liability, even if its dashboard looks impressive.

When a new answer surface enters the tracked set, define its inclusion criteria before adding it to the KPI trend. Agree on the change-order process, expected implementation effort, and whether historical data will be backfilled. Never let a vendor expand the scope and then present the larger dataset as proof of progress.

End cleanly when the work ends

Before renewal, request documentation of entity mappings, prompt libraries, measurement scripts, schema decisions, content change logs, source evaluations, and unresolved risks. Run a handoff session with your SEO, content, analytics, engineering, and revenue-operations owners.

LLMrefs is one practical option for teams that want to keep this measurement layer visible internally. It tracks brand mentions, citations, share of voice, and rankings across AI answer engines, supports prompt generation from keywords, and provides exports and API access for ongoing analysis. Use it as an auditable source of evidence, not as a substitute for business judgment.

The best AI search agency relationship leaves your team more capable than it was at the start. You should understand what changed, why it changed, how the result was measured, and how to continue without a knowledge silo.

LLMrefs helps brands and agencies monitor visibility across AI answer engines, inspect citations, compare competitors, and turn prompt data into measurable share-of-voice reporting. Visit LLMrefs to benchmark your current AI visibility and build a vendor evaluation process grounded in evidence rather than screenshots.