How can a marketing agency run a free AI visibility audit to win new clients?

We win a fair share of new conversations with a free AI visibility audit, mostly because it shows the prospect something concrete they cannot get from their current agency. It takes us a short working session to produce, and it works best when we stay disciplined about what it claims.

We build these in Promptwatch, since that is where we run client programs, so the examples below are from that setup. The method works with any tool that can run a prompt panel and show sources.

What the audit is for

The audit has one job: to start a conversation about a paid first phase. It should not try to solve the prospect's problems for free. If it does, you have given away the engagement. If it is too thin, it looks like a lead-gen form.

We aim for a one-page summary with an appendix. The summary has four findings and one recommendation per finding. The appendix has the prompts, the engines, the dates, and the sources.

Choosing prompts

This is the step that decides whether the audit is credible. We write 20 to 40 prompts per prospect in three groups:

• Category discovery. "Best [category] for [use case]" in the language a buyer uses.
• Competitor comparison. "[Prospect] vs [main rival]" and "alternatives to [rival]".
• Brand. What the engines say when someone asks about the prospect directly.

We add the prospect's two or three real competitors, ask the prospect who they worry about, and include location where the business is local. We keep the whole list in the appendix, so nothing looks like a cherry-picked demo.

Setting up the run

In Promptwatch we create a pitch project for the prospect. Agency plans include unlimited pitch projects, so audits do not eat into client capacity, and they come with unlimited prompts and all tracked models. Pitch projects are the single biggest reason agency plans suit new business work. Agency plans start at $199 a month (Kick-off) and go to $399 (Growth) and $799 (Scale).

Before we load the prompts, we set the Brand Book and aliases, so the tool recognizes the prospect's brand under its common variants. Wrong aliases make the share-of-voice numbers wrong, and a prospect who spots an error stops trusting the rest.

The panel runs daily per model. For a pitch, we want at least a few days of data, because a single sample is weak evidence and answers change from run to run. If we need something the same day, we say so in the audit and label it a first look.

What we capture

1. Appearance and position. Visibility Score and Position in Answer for the prospect and its competitors.
2. Share of Voice. How the prospect compares with the named rivals on the same prompts.
3. Sources. Citation analytics shows which pages and domains the engines cite, including Reddit and YouTube. We highlight any domain that cites a competitor and not the prospect.
4. Access. Where possible we check robots.txt and llms.txt and run Promptwatch's free Agentic Readiness Check. We cannot see crawler logs for a prospect until they connect a log source, so we say that plainly and offer it as phase one.

Promptwatch also has a no-login AI Visibility Report, which is useful for a quick first scan before a call, and the GEO Inspector Chrome extension, which shows ChatGPT fan-outs and the sources fetched versus cited. We sometimes use the extension live in a meeting, because watching the fan-outs appear tends to land better than a slide.

How we present it

We keep the language careful.

• We show what the data says today, with the date and the prompt list.
• We never promise the prospect will be recommended after our work. We cannot control that.
• We show the data as a sample, not a verdict, and we mention that answers vary.
• We choose one finding the prospect can check in thirty seconds. A competitor named in ChatGPT where they are absent works well.
• We recommend three things at most: one content gap, one access or technical issue, one off-site opportunity.

We use Promptwatch's PDF reports and white-label branding on exports where we can. We do not assume a white-label plan, since the plan details are not specified, so we confirm during setup.

Turning the audit into a paid phase

The follow-up proposal is narrow: a fuller prompt panel, a crawler check with their developer, a content gap analysis and a first monthly report. We do not name outcomes. We describe the evidence we will collect, and what decisions it will inform.

If they sign, the prompt list and the competitor set we built for the pitch are the starting point for the client program, so the audit work is not wasted.

Mistakes we avoid

• Running only ChatGPT. Prospects often differ by engine, and the difference is the story.
• Using too few prompts or only brand prompts. They flatter or alarm without informing.
• Showing a single screenshot as proof.
• Overselling the audit as a guarantee.

Where to start

Write 20 prompts for a friendly prospect or one of your own contacts, run them across the main engines, and build the one-page summary. If you do not yet have a tool, Promptwatch's free tools let you try the scan and readiness check without a subscription.