Branded vs Unbranded Prompts: Why You Split Them and How

A branded prompt contains your company name. "What does Acme Payroll cost per employee?" or "Does Acme integrate with QuickBooks?" The engine will almost always mention Acme in the reply, because you told it to. So the useful question is not whether you appear. It's whether the price, the integration list and the founding year in the answer are correct.

An unbranded prompt describes a need. "Best payroll software for a 20-person company" or "How do small agencies run payroll across two states?" Nobody is named. Here the useful question is whether the engine names you, and next to whom.

So the two sets produce two different metrics:

• Branded set: error rate and error severity. Mentions are expected and carry no signal.
• Unbranded set: mention rate and the company you keep. Accuracy matters too, but presence comes first.

If you're still deciding whether to track both, the short answer to should I track branded prompts, unbranded prompts, or both is both, in separate columns.

Say you run 10 branded and 10 unbranded prompts. The branded set returns a mention nearly every time. If you average mention rate across all 20, half your score is guaranteed. Add five more branded prompts next quarter and your "visibility" rises without anything changing in the market.

The reverse happens with accuracy. Unbranded answers that never mention you contain no facts about you to check. Fold them into an accuracy score and the error rate gets diluted by answers that had nothing to get wrong.

There's a second layer of mixing nobody can see. As of October 2026, retrieval is discretionary on every major platform. OpenAI says ChatGPT search "rewrites your query into one or more targeted queries", and the model decides when to search at all. An answer given without search comes from training data with a fixed cutoff. OpenAI documents a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Any score is already an average of grounded and ungrounded answers. Don't stack a second average on top.

The fix is simple: two sets, two metrics, two lines on the report. Phrasing still matters inside each set, as covered in how phrasing changes whether AI mentions your brand.

Unbranded answers favor brands that are already widely discussed. A 2025 study of AI search citations (arXiv 2507.05301) found the top 20 news sources accounted for 67.3% of all citations for OpenAI models. Its data window was March to May 2025, so treat it as a pattern, not a current figure. The 99Visibility evidence review rates brand prominence as a strong association with no lever: you can't buy it this quarter, and no published study controls for it well.

That changes how you read your numbers. If you sell regional payroll software and your unbranded benchmark includes the three largest payroll companies in the world, you will lose every comparison. The gap measures size, not anything your site did.

Pick three to five competitors of similar size, funding stage or market. Report your mention rate against theirs. For larger players, track them separately as context, and see whether to track prompts that name competitors for how to handle those.

What you can influence is narrower and still real. When engines cite other sites instead of yours, being named on those sites is associated with appearing in the answer. That's an association, not a promised effect.

1. List your facts first. Write down the 10 to 20 facts that matter to a buyer: pricing tiers, core product names, integrations, locations, founding year, ownership. These become the answer key for the branded set.
2. Write the branded set. One prompt per fact or fact group. "What plans does Acme offer and what do they cost?" Avoid leading phrasing that contains the answer.
3. Write the unbranded set. Use the way buyers describe the problem, not your marketing terms. Cover the category, the main use cases and the key buyer constraints. The prompt set template gives categories to start from.
4. Tag every prompt. Mark each as branded or unbranded, plus a topic. Store the set with a version number.
5. Run under controlled conditions. Use a logged-out session or the API, with memory off. Run each prompt more than once, because answers vary between runs.
6. Record retrieval state. Note whether each answer showed citations. No citations usually means you can't tell whether it was grounded.
7. Score branded answers for accuracy. Count flagged answers divided by answers that actually returned. Record the worst error separately, so one wrong price isn't averaged away.
8. Score unbranded answers for presence. Mention rate, position if listed, and which peers appear alongside you.
9. Report per platform. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews are different systems. Google says even AI Overviews and AI Mode may use different models. Keep each in its own column.

If you want a quick first look before building this, the free checker on this site runs up to three buyer questions and gives a one-time snapshot.

• Branded and unbranded results sit in separate tables.
• Branded results show error rate, not mention rate.
• Unbranded results show mention rate, not accuracy alone.
• The error rate divides by answers returned, not answers requested.
• The single worst error is listed by name, with the wrong fact and the correct one.
• Each platform has its own row. No cross-platform average is presented as a measurement.
• Retrieval state (citations shown or not) is noted per answer.
• Peer benchmarks use like-sized brands only.
• The prompt set version matches last period's, or the change is disclosed.
• No rank or position "inside Google AI" is reported. Google exposes impressions only, so there's nothing to check such a number against.

Good

• Branded errors are few, and each one traces back to something you can find: an old pricing page, a conflicting about page, a third-party listing.
• Unbranded mention rate is stable across repeated runs and sits in the same range as your like-sized peers.
• When you're absent from an unbranded answer, the cited sources tell you which sites the engine trusted. That points to a fix, covered in finding prompts where competitors appear and you don't.

Bad

• Visibility jumps because someone added branded prompts.
• A wrong acquisition claim or wrong price disappears inside a healthy average.
• Unbranded answers name you, but in the wrong category. That is a presence result hiding an accuracy problem. See what to do when AI puts you in the wrong category.
• Branded errors appear only in answers with no citations. That often points to stale training data, which nobody can change this quarter.

What to do with each

For branded errors in grounded answers, state the correct fact plainly on a page crawlers can fetch, in the HTML before JavaScript runs, and remove contradictions across your own site. Then check that OAI-SearchBot, PerplexityBot and Claude-SearchBot are allowed in robots.txt and not blocked by your firewall. These are the levers with evidence behind them. None of them guarantees what a model says.

Run both sets monthly as a baseline. That's often enough to catch drift and rare enough to keep the prompt set stable.

Rerun the branded set within a few weeks of any fact change on your site, such as a new price or a renamed product. That shows whether grounded answers picked it up. Ungrounded answers won't change until a new model with a later cutoff ships.

Rerun both sets when a platform launches or switches its default model. Note the date in your report so a shift isn't misread as your doing.

Change the prompt set at most once a quarter, and version it. If you add prompts, report the old set and the new set side by side for one period so the trend line survives. For platform-specific setup, see reproducible ChatGPT prompt tracking.