AI Visibility Client Onboarding Checklist for Agencies

A practical handoff guide for agency delivery managers: the access, brand data, prompts, competitors, and baselines to lock down before AI visibility work begins.

Updated on: 2026-10-01

The onboarding calls that go sideways almost always have the same cause. Someone sold an AI visibility retainer, the contract got signed, and then the delivery team inherits a client name, a domain, and a vague promise to "get them showing up in ChatGPT." No agreed prompts. No competitor list the client actually recognizes. No one who can approve the first report. Three weeks later you are still waiting on Search Console access nobody decided you needed.

This checklist is the part after the sale. It assumes the engagement is approved and scoped. The job now is to gather, approve, and access everything required to produce a defensible baseline and start delivery without stalling at review. If you are still in the pitch stage, the piece on going from a prospect audit to a scoped GEO proposal is the one you want first.

AI visibility here means something specific: whether and how a client's brand is named, recommended, positioned within a generated answer, or cited as a source in AI answer engines like ChatGPT, Claude, Perplexity, and Gemini. That is different from organic rankings, different from AI-driven referral traffic, and different from Google's own generative-search reporting. Keep those separate from the first day, because they measure different things and clients collapse them constantly.

A domain and a brand name are a starting point. Inside SEOforGPT, the workflow is to set up a separate client workspace with the website and brand details, then review suggested discovery prompts that you can edit before anything runs. That workspace structure matters operationally, because it keeps each client's prompts, competitors, and reports isolated and makes the approval step before report sharing explicit.

Here is the field set I would not skip.

Client identity. Canonical brand name, common misspellings and abbreviations, parent and sub-brand relationships, and every official domain or country or language site. This reduces missed mentions and false matches. Decide early who on the client side can approve entity aliases and which products must be treated as separate brands.

Offer and audience. Products and services in scope, priority audiences, real use cases, geography and language, and buying stage. Just as important: the claims the client does and does not make. This grounds prompt selection in the client's actual market instead of generic category questions.

Brand guardrails. Preferred descriptions, prohibited or outdated claims, regulated topics, sensitive wording, and the name of a legal or compliance reviewer. Model-generated descriptions are not approved client messaging. If you let an AI-written summary slip into a report as fact, you own that mistake.

Prompts. Existing customer questions pulled from sales, support, or approved research, plus priority discovery, comparison, alternative, and problem-framed questions. Record the exact wording and a reason for each priority prompt. When the prompt set changes later, keep the change visible so comparisons stay honest.

Competitors, in two sets. The client's confirmed strategic competitors and exclusions in one field. Competitors newly observed in AI answers in a separate field. The second set often surprises the client, and it should be reviewed, not silently swapped in. A brand that an engine keeps naming is worth discussing, but it is not automatically a strategic peer.

The fastest way to lose a week is to request broad admin access the scope never needed. The second fastest is to collect a client's personal password instead of named access.

Decide per system whether you need it at all, then request the lowest role that does the job.

One nuance worth flagging to the client: analytics access is not required to monitor AI answers. Prompt and model monitoring runs on the defined prompt set. Search Console and GA4 add organic search and post-click context, which is useful for a broader outcomes baseline, but it is a distinct permission request that the scope should justify. If you do not need it, write down the decision not to request it rather than asking by reflex.

An initial reading is a snapshot. It is not a trend, and it is not proof of anything yet. Say that to the client out loud during onboarding, because the first report will otherwise be read as a verdict.

The reason to retest is not caution for its own sake. A 2026 arXiv study ran prompts across four AI search engines and four Swiss-German verticals and found day-to-day source overlap of roughly 34 to 42 percent, with brand overlap higher at about 45 to 59 percent but still well short of stable. The authors recommend repeated measurement, and they are explicit that their run counts and time windows come from their specific engines, markets, and method. Use that to justify retesting your client's prompts, not to treat their sample size as a universal rule.

For each prompt in the baseline, capture:

That last point is the one teams get wrong. SEOforGPT's tracking distinguishes a completed negative result from an unavailable provider result, and that distinction matters. An unavailable run is not evidence the brand was absent. If you log a provider timeout as "not mentioned," your baseline is already wrong and every later comparison inherits the error.

Preserve the answer text, not just a score. When a client asks why their visibility number moved, you want to show the actual answer, the engine, and the date, not defend a figure you can no longer explain.

This is where a lot of AI visibility reporting quietly loses credibility. A single "AI performance" number that blends mentions, citations, positions, and traffic is easy to build and impossible to defend. Agree on distinct measures before launch.

My read, after enough of these reports: the agencies that hold these apart from day one have far easier quarterly reviews. The ones that merge them into a hero metric spend those reviews explaining why the hero metric is fine, which is a worse conversation than talking about five honest signals that each moved for a reason.

To make this concrete, here is a fictional example. It is illustrative, not a documented client result, and it carries no invented scores or uplift.

A regional B2B software firm signs on. During onboarding the client confirms three brand aliases including an old product name still in the market, two priority markets with two languages, and a strategic competitor list of four companies. The agency drafts fifteen priority prompts from the client's sales and support questions, records the reasoning for each, and reviews the engine-suggested discovery prompts, trimming the ones that target questions nobody asks.

The first run across the available engines surfaces two competitors the client never mentioned. Those go in the observed-competitors field for client review, not the strategic set. One engine returns an unavailable result for three prompts; those are logged as unavailable, not as absences. The agency freezes the baseline, captures every answer, and gets client approval on the prompt set and competitor definitions before sharing anything. The client is told in writing that this first read is a baseline.

Nothing in that example requires a visibility score to be useful. The value is a clean, auditable starting point everyone agreed to.

Work stalls at review more than anywhere else, so name the people. Client decision-maker, day-to-day contact, agency owner, and the content, brand, privacy, and legal approvers. Agree on approval turnaround and an escalation path. If one reviewer can hold the whole pipeline, you want to know that before the first article is drafted.

On reporting, SEOforGPT produces an agency-reviewed, frozen, read-only snapshot shared through a white-label report with a tokenized URL. Its own page notes that anyone with that URL can view the report. Treat the link as shareable access: set an expiry where it fits, and keep sensitive information out of a report unless the client has approved the distribution method.

Resolve data handling during onboarding rather than after an upload. Establish whether prompts or uploaded material could include personal, confidential, or regulated information, which providers receive data for a given feature, what retention and deletion terms apply, and who removes access at offboarding. One practical detail on the input side: SEOforGPT's Additional Knowledge workflow accepts DOCX, TXT, Markdown, and pasted text. PDF files are not supported, and PowerPoint files are not supported either, so plan to convert slide decks and PDFs to text before handing them over. The uploaded file is not retained after its text is extracted.

Before you launch, confirm the agreements are real and not assumed: a client-approved prompt set, competitor definitions, markets and languages, named approval owners, reporting recipients, and a shared definition of what counts as a completed observation. Then verify the team can reach the properties the scope needs, the first run is reviewable, and the client understands the baseline is a starting line.

No. Prompt and model monitoring runs on your defined prompt set across the answer engines, and that does not depend on analytics access. Search Console and GA4 add organic search and post-click outcomes, which help a broader baseline, but they are a separate permission request. Only ask for them when the scope calls for traffic and conversion context, and request the lowest role that covers the task.

There is no universal number, and anyone quoting one as a rule is guessing. Base the count on the client's real buyer questions, the markets and languages in scope, and your plan budget. The research that recommends repeated measurement is explicit that its own figures came from one study's engines and methods. Start with a prompt set the client can approve and defend, then retest it on a schedule rather than chasing a magic quantity.

Not necessarily. An engine naming a brand repeatedly tells you it is part of the answer pattern for that question, which is worth understanding. It does not confirm the brand is a strategic peer. Record observed competitors in a separate field from the client's confirmed set and review them together. That conversation often sharpens the client's own view of their market.

Because AI answers vary across runs, models, and days, a single reading is a snapshot of that moment. Day-to-day overlap in both sources and brand mentions is far from perfect in the available research, so an early result read as proof of success or failure will mislead the client. Frame it as the starting point you will measure change against, and the quarterly reviews get much easier.

Waiting on access and approvals you never defined. Decide which systems you actually need, request named access at the lowest useful role, and name every approver with a turnaround expectation before work starts. The delivery problems I see repeatedly are not technical. They are missing decisions that onboarding was supposed to make.