Citelytics · Blog

For agencies

What a Good AI Visibility Baseline Looks Like Before You Change Anything

A practical, repeatable baseline helps agencies separate what they observed from what they changed—and scope the work profitably across clients.

Published: 2026-11-18

A baseline is not a screenshot of a brand appearing in an AI answer. It is a dated, repeatable record of what a defined set of questions returned, under defined conditions, before your agency made changes. That distinction matters when a client asks whether your work caused a shift, and when your team has 30 accounts to report on without turning every review into a bespoke investigation.

The baseline’s first job is operational: freeze the starting conditions. Its second is commercial: show what is included in measurement, what is not, and what evidence would justify further work. It should never imply that a measured result guarantees future recommendations or mentions.

Define the measurement unit before collecting data

Write down the unit you will compare later. A useful default is one client, one market, one language, a fixed prompt set, and a named AI surface. Keep different products or service lines in separate prompt groups when a single blended score would hide important differences.

For a first pass, use 12–20 prompts: roughly four to six each for category discovery, comparison or shortlist, and specific need or use case. Include branded and non-branded questions, but label them separately. These are working agency defaults, not industry standards. The right size is the smallest set that covers the client’s priority journeys and can be rerun consistently.

For each prompt, save the exact wording, date and time, platform and mode, market or location settings where relevant, and the complete response. Record whether web search was used, plus any cited sources and linked pages. ChatGPT’s own documentation notes that search responses may contain citations and that citations can be incomplete, outdated, or incorrect; preserve the answer and inspect important sources rather than treating a citation as proof.

Make the record useful, not just scoreable

A percentage alone is a poor handoff. Store the raw answer alongside a small, consistent annotation schema. For every response, mark whether the client is named, whether the description is accurate, whether a competitor is named, whether a source is cited, and whether the cited source belongs to the client. Add a short evidence note for ambiguous cases. Keep “not mentioned” distinct from “incorrectly described.”

Report the numerator and denominator. For example: “Client named in 5 of 16 fixed prompts; cited as a source in 2 of those 5.” Include the prompt group and platform beside the figures. Do not collapse different surfaces, locations, or prompt types into one agency-wide score unless the client understands exactly what that aggregation means.

Use a freeze-and-review workflow

A low-friction baseline can fit into a standard onboarding sprint. One person owns the prompt set; another checks the captures and labels. That second review is especially useful for descriptions, source attribution, and cases where the model names a brand indirectly.

If a response is missing or a platform changes materially during collection, log that fact rather than silently replacing the observation. Preserve the original record and note any rerun separately. This is not bureaucracy: without an audit trail, later teams cannot tell whether a changed result came from changed client conditions, changed prompts, or a different collection process.

Explain the boundary between visibility and traffic

A visibility baseline describes observed responses in a defined sample. It does not establish how often the client appeared across all AI answers, prove why a response occurred, or forecast leads. Keep those limits in the report’s first page, not in a footnote after a confident-looking chart.

Where available, pair response observations with first-party analytics, but don’t imply that the data sources measure the same thing. Google says AI Overviews and AI Mode are included in the overall “Web” search traffic in Search Console; that is not a separate count of every AI mention. Treat visibility observations and measured site visits as separate evidence streams, then discuss their relationship cautiously.

Package the baseline so the service can scale

Set a standard allowance before selling the service: for example, one fixed prompt set, two named surfaces, one market, and one baseline report. Quote separately for extra markets, prompt expansion, manual source verification, or custom analysis. Track actual setup and QA time for the first five clients; use the median to revise your onboarding fee and delivery estimate, rather than assuming every account will take the same effort.

For dozens of clients, consistency is a margin control. Use one capture template, one label dictionary, a versioned prompt library, and a central exception log. Let account teams add client-specific prompts only through a documented change request. That protects comparability and stops “just add a few questions” from becoming unpriced research.

The client-facing close is simple: this is the starting observation, under these conditions, on this date. We will rerun the same set and report what changed, alongside any documented changes in the measurement conditions. That is a credible service promise. A promise that AI systems will recommend the client is not.

Sources

Running this for clients?

The Agency plan is priced per fresh scan, not per location — the client list is unlimited, so a sleeping client costs nothing.

See plans

More on this

← All articles