Branded vs Unbranded Prompts: How to Stop Rigging Your AI Visibility Test

Picture a dashboard that says your brand shows up in two of every five AI answers. The team celebrates. Then someone opens the prompt list and finds that nearly half the prompts contain the company name. The number was never about buyers finding you. It measured whether a model recognises a name you typed into the question yourself.

This article is for whoever owns the AI visibility prompt list: a marketing lead, an SEO, a founder with a spreadsheet. It covers what each prompt type can and cannot tell you, how to write the unbranded kind in buyer language, and the reporting rule that keeps the whole exercise honest.

Three prompt types, three different questions

AirOps splits prompts into two groups. Branded prompts name your company, so they test whether an assistant knows you and describes you correctly. Unbranded prompts state a need without naming a vendor, so they test whether the assistant picks you unprompted (AirOps, June 2026). A third type deserves its own line: prompts that name a rival, such as "alternatives to [Rival]" or "[You] versus [Rival]". They show how you are placed when a buyer is already comparing.

Mixing the three into one rate hides the only part that matters for growth. Branded results will run high and unbranded results low, and that gap is the normal shape of the market, not a bug in your tracking. Report three lines.

The AirOps article also offers one illustration of how low the unbranded line can go: a 0.34% average mention rate on a single target prompt (AirOps). It is one vendor dataset on one prompt, so use it to calibrate your expectations, not as an industry norm.

Where unbranded prompts should come from

Not from your product page. Writer's guidance on buyer conversations says to pull 20 to 30 questions from sales calls, support tickets and win-loss interviews, keeping the buyer's own wording, then cut the list to 10 to 15 grouped into a few themes (Writer). Writer sells enterprise AI, and its own pipeline pulls transcripts from a call-recording tool each morning, which most teams will not copy. The principle survives without the automation: start from what buyers actually said.

Compare two ways of asking the same thing.

Rigged:    "Which AI content platform is best for B2B?"
Realistic: "There are two of us on marketing and the CEO wants twice the
            output by spring. What should we look at before hiring?"

The first prompt carries your category vocabulary and a leading "best". The second describes a situation and lets the model decide which category applies. If you want a method for finding more of these, see how to do prompt research for AI search.

• Strip any word only your own team uses.
• Keep role, team size and constraint where the buyer gave them.
• Drop "top", "leading" and "best", which push the model toward a ranked list.

Freeze the list, then change it on purpose

Give the list a version number and a date. Do not edit it mid-quarter. When a prompt must change, add the replacement and run the old one in parallel for a month, so the trend line has an overlap to stand on. Run every prompt several times in at least two assistants, since answers vary run to run, as covered in why AI answers change every run.

Include a few prompts where you expect to be absent. They are the most useful rows in the sheet because they point at the content gaps.

A worked example with invented numbers

Take a fictional invoice-automation vendor for freight firms. Its list holds twenty prompts: eight unbranded, six rival-named, six branded. After three runs in two assistants the vendor is named in five of six branded prompts, two of six rival-named prompts and one of eight unbranded prompts. Blended, that reads as eight of twenty, two in five. Split, it reads: well known, rarely discovered.

The fix follows from the split. Publish pages that answer the situation-style questions, earn mentions on the third-party sources that the answers already cite, and re-run in a month. A blended two-in-five would have prompted none of that.

Why the third-party line matters more than your own site

Writer reports a Bain and ScrunchAI analysis in which 89% of unbranded prompts were answered with third-party sources, across about 500 million citations (Writer, second-hand). Treat that as second-hand and vendor-adjacent. If it is even roughly right, an unbranded win depends on reviews, comparisons and press more than on your own blog, and your prompt log should record which domains each answer cited.

Call it: the blended score deserves to go

Most guides say to report the split and, if you must, a blended number as well. Drop the blended number. A weighted average of three different measurements has no meaning that survives a question from the board, and its value depends on a choice you made yourself: how many branded prompts to include. Move that dial and the score moves without any change in reality. That is exactly the rigging this article opened with, performed politely.

The split worth defending is simple. Unbranded is the headline, because it is the only line a competitor cannot inflate by changing the question. Branded is a quality check on accuracy. Rival-named is a positioning check. A small category undermines this advice: with only three real vendors, nearly every unbranded prompt names all three, and the unbranded rate stops saying anything. There, lean on rival-named prompts and on how you are described.

Search Console now offers a parallel check on your own demand. Google's branded queries filter separates branded from non-branded queries in the Performance report, using an AI-assisted classifier rather than a regular expression, and it reached all eligible sites on 11 March 2026 (Google Search Central). If branded search in Search Console is rising while the unbranded AI line sits flat, awareness is growing without discovery, which is the same story told from the other side. Google's separate Search Generative AI performance reports, which began on 3 June 2026 and reached all sites by 31 August, add impressions in AI features by page, country, device and date, though not clicks (Google Search Central).

What a cheap audit of your current list looks like

Open the sheet and add one column: prompt type. Sort by it. If the branded rows outnumber the unbranded ones, the old score is inflated by construction, and you can say so without re-running anything. Then add a second column for the buyer source of each unbranded prompt: a call, a ticket, a survey answer. A prompt with no buyer source is a prompt somebody on your team imagined, and imagined prompts tend to borrow the team's own vocabulary. Delete or flag them. Ten prompts with a real origin beat thirty without one, because the first set can be defended in a meeting and the second cannot.

Signs the list is leaning your way

• Your brand appears in most of the prompts that were meant to be neutral. Count the branded ones.
• Prompts echo your slogans or category label.
• Every row is a "best of" list request.
• Rivals never appear in any answer, which usually means the prompts are too narrow.
• Answers come from one logged-in account. Personalisation can colour results, so use clean sessions where you can.

Rebuild the buyer-language questions every quarter from fresh calls, because wording drifts. For the metric definitions behind each line, see the complete measurement guide, and for ways to improve the unbranded line, the GEO guide for B2B teams.