How many prompts should an LLM visibility audit cover?

For a B2B tech client we typically build out forty to sixty prompts across four categories. Brand prompts (5 to 10) about your company directly. Definitional prompts (10 to 15) for generic terms in your space. Comparison and selection prompts (10 to 15) for the shortlist-shaping queries. Transactional and pricing prompts (10 to 15) for bottom-of-funnel. Refine the set against the client's actual sales pipeline. The prompts that matter are the ones their real buyers ask, not the ones the marketing team thinks sound clever.

Why does running prompts logged in mess up the audit?

ChatGPT, Copilot and Gemini personalise based on your history. A logged-in audit gives you results biased to your own previous queries, which makes the data unrepresentative of what a fresh prospect sees. Use logged-out browsers in clean sessions. Set location explicitly for Google AI Overviews because it is geographically sensitive. Run priority prompts at least twice for Perplexity, where two consecutive runs can return different sources. Consistency matters more than perfection, but logged-out is the cleaner baseline.

How often should we run a full visibility audit?

For most B2B tech clients, a full audit twice a year is enough, with monthly automated tracking on the priority prompt set in between. Heavier cadence rarely surfaces new insight and usually creates noise. Tools like Profound or Athena handle the in-between tracking. Quarterly manual audits supplement the tooling because tools miss nuance and manual work surfaces edge cases. Citation patterns drift, so revisit the audit even where nothing material has changed in your own programme.

Auditing your visibility in Copilot and ChatGPT

If you do not know what ChatGPT and Microsoft Copilot are saying about your brand, you do not know what your buyers are reading. We’ve run this audit dozens of times for B2B tech clients and it never fails to surface something the marketing team did not realise. Sometimes it’s good news. Often it’s a competitor occupying a citation spot that should be yours, or a flat-out wrong description of what you do.

This piece walks through the audit process we use, including the prompt sets, the tooling, what to record and the common gotchas. It’s grounded in audits we’ve run for MSPs, SaaS firms and ERP consultancies through 2025 and into 2026.

What you’re actually auditing

A complete audit looks at three things in parallel:

Brand visibility. What do the LLMs say about your company when asked directly?
Topical visibility. Are you cited for the buyer queries that matter to your business?
Competitive visibility. Who else is being cited for the same queries, and what shape of content are they offering?

These are different lenses. Each one needs a slightly different prompt set. We’d argue you need all three to make sensible decisions.

Building the prompt set

The prompt set is the heart of the audit. Get this wrong and the rest does not matter.

For a B2B tech client, we typically build out forty to sixty prompts across four categories:

Brand prompts (5 to 10). Direct questions about your company.

“What does Acme Cloud Services do?”
“Who are the senior people at Acme Cloud Services?”
“Is Acme Cloud Services any good?”
“What’s Acme Cloud Services known for?”

Definitional prompts (10 to 15). Generic terms in your space.

“What is co-managed IT?”
“How does MDR pricing work for mid-market businesses?”
“What’s the difference between EDR and XDR?”

Comparison and selection prompts (10 to 15). The shortlist-shaping queries.

“Best MSPs for law firms in the UK”
“Top SAP S/4HANA consultancies for mid-market”
“How to choose a managed security provider”

Transactional and pricing prompts (10 to 15). Bottom-of-funnel.

“How much does outsourced IT support cost in the UK?”
“Average MSP pricing per user 2026”
“What’s included in a 24/7 IT support contract?”

We refine these against the client’s actual sales pipeline. The prompts that matter are the ones their actual buyers ask, not the ones the marketing team thinks sound clever.

Running the prompts

Run each prompt across the four major surfaces:

ChatGPT. Logged-out browser session. Default search behaviour on. Note the cited sources, the body of the answer and any visible reasoning.
Microsoft Copilot. Both the public surface (copilot.microsoft.com) and, where you can, the Microsoft 365 Chat surface inside the tenant of a customer-shaped persona.
Perplexity. Free version. Note that Perplexity returns different answers across sessions, so run priority prompts twice with a gap.
Google AI Overviews. Logged-out, location-set browser. Note that not all prompts surface AI Overviews, and the engine sometimes shows different sources for the same prompt minutes apart.

We typically also include Gemini and Claude where the client cares about completeness, but ChatGPT, Copilot, Perplexity and AI Overviews cover most of the practical visibility surface for B2B tech buyers.

For each prompt and surface, capture the answer text, the cited sources (URL plus title) and a screenshot. We use a shared spreadsheet with one row per prompt-surface pair. It’s tedious. It’s also where the insights come from.

What to look for

Once you’ve captured the data, the analysis falls into a few patterns:

Where you appear

Are you cited at all? In which surfaces? For which prompt categories? Most B2B tech clients we’ve audited appear well for brand prompts and patchily for everything else. The gap is usually in comparison and pricing prompts.

Where competitors appear

Which two or three names recur across the comparison and selection prompts? Those are the firms occupying the citation slots you want. Look at the pages those citations point to. They are usually well-structured, contain pricing or scoping detail and have an authored byline. Patterns repeat. We’ve broken down what to do when a competitor gets cited instead of you.

What’s wrong about your brand

Read the brand-prompt answers carefully. Is the description of your company accurate? Does it name the right product categories? Does it mention the right industries? We’ve seen LLMs describe an MSP as “a small IT consultancy” when the client had four hundred staff. We’ve seen Copilot confuse two clients with similar names. These errors propagate. They need addressing.

The shape of cited content

Look at the URLs being cited from competitor sites. What kind of page are they? Pricing pages? Comparison pages? Case studies? Glossary pages? Blog posts? The citation pattern usually reveals which content shape is working in your category. We’ve covered the writing-craft side of this in writing content that AI search engines actually cite.

Tooling

Manual audits are valuable but slow. For ongoing work, tools speed things up:

Profound. Define a prompt set, run it on a schedule, track citation share over time across multiple LLMs.
Athena. Similar category, with stronger reporting on competitive share of voice.
Semrush AI tracking. Convenient if you already use Semrush.
Ahrefs brand mentions. For tracking the wider third-party signal that feeds LLM authority.

We typically combine one ongoing tool with quarterly manual audits. The tool catches drift. The manual audit catches nuance the tool misses. We’ve put the two side by side in Profound vs manual audits to help teams decide where to spend.

We covered the broader tracking question in tracking AI search traffic.

Common gotchas

Things we’ve watched go wrong:

Running prompts logged in. ChatGPT, Copilot and Gemini personalise. Logged-in audits give you results biased to your own history. Use logged-out browsers, or accept the bias and run the audit consistently the same way.
Forgetting location settings. Google AI Overviews is geographically sensitive. Set location explicitly to a representative buyer location.
Treating one run as the truth. Especially with Perplexity, two consecutive runs of the same prompt can return different sources. Run priority prompts at least twice.
Auditing the wrong prompts. If your prompt set is what you want buyers to ask rather than what they actually ask, your audit is measuring fiction.
Ignoring the body of the answer. Citation appears, but is the body of the answer flattering, accurate or damaging? A citation that surfaces a negative review is not a win.

What to do with the audit

The audit is a means, not an end. The output should drive a clear list of actions:

Fix factual errors about your brand. Often by tightening your About page, adding clarifying schema, getting an accurate description into Wikipedia or trade publications. This often moves quickly because LLMs do recrawl.
Build the missing pages. If competitors are being cited for “best MSP for law firms in Birmingham” and you serve law firms in Birmingham but have no page targeting that combination, build the page. We’ve covered the structure in Google AI Overviews and MSP citation.
Rewrite the underweight pages. Some pages exist but do not get cited. Apply the structural rewrite framework from writing content that AI search engines actually cite.
Address the third-party signal. Some gaps cannot be fixed on your site. They need mentions in trade press, analyst recognition or podcast coverage. We covered this in brand mentions vs backlinks in AI search.

We typically deliver an audit alongside a sequenced action plan with the highest-impact items first. A reasonable team can work through the most valuable fixes inside a quarter.

Cadence

For most B2B tech clients, a full audit twice a year is enough, with monthly automated tracking on the priority prompt set in between. Heavier cadence rarely surfaces new insight and usually creates noise.

If you’d like a second opinion on your AI search strategy, drop us a line. Citation auditing is one of the most useful starting points for any AI SEO engagement, and it’s almost always more revealing than the marketing team expects.