Branded vs. Unbranded Queries in AI Search Tracking

A single AI-visibility number that averages every prompt you track mixes questions carrying your name with questions that do not, and the first kind pulls the figure up on its own. Keep the two sets apart as separate samples, fix the engine, language, region and question version before any comparison, and report a branded rate and an unbranded rate side by side. This is the measurement discipline behind aeo work, and it changes what the number you show a client means. As of November 2026, that split has an official equivalent in Google Search Console, not just a private spreadsheet rule.

What the split asks of you:

A prompt naming your brand and a prompt describing a buyer's problem answer different questions, so they cannot share a denominator.
Brand-name questions lift any mention rate by themselves, which makes a rising blended score a poor sign that new buyers are finding you.
Google now offers the same cut natively for organic search, so the segmentation is a published reporting convention.
A tracking plan needs a fixed question version, engine, language and region before you compare two periods or two competitors.

What counts as a branded or unbranded prompt

Google Search Central announced its branded queries filter in 2025 and later opened it to eligible sites in an update, in a post by Michael Huzman. The post defines a branded query as one carrying the brand name, variations or misspellings of it, and brand-related products or services. Classification comes from an internal, AI-assisted system rather than a keyword list, some queries can be misidentified, and the filter has no effect on how Google ranks anything.

That definition transfers to answer engines with one adjustment. In traditional search the unit is a typed query; in AI search the unit you track is a prompt written down and submitted to a named engine. A branded prompt carries your brand word or a product word. An unbranded prompt describes a category, a problem or a use case and never names you.

The two types belong to different moments of buying. Someone who already knows the name asks a branded prompt to compare it, check it or confirm a choice. An unbranded prompt comes from someone still shopping across the category, who has not picked a brand at all.

A third case sits outside both lists and deserves its own label: a prompt naming a competitor but not you. "Alternatives to [rival]" is not unbranded in the strict sense, because it names a brand, and it is not yours. Folding those prompts into the unbranded set is one of the quiet ways a report goes wrong.

The blended-score trap

A blended visibility rate, the kind of figure most aeo dashboards put on the home screen, is the count of answers that mention your brand divided by all answers in the set. The HubSpot article of 9 November 2026, written by Ramona Sukhraj, works through the arithmetic of a mixed set: the branded and unbranded rates sit far apart, the merged figure lands somewhere between them, and adding branded prompts moves the merged figure without touching discovery. The article also proposes a branded-to-unbranded ratio for comparing the two sets over time.

A blended figure is easy to keep improving and hard to learn from. The only rate a new buyer's question produces is the unbranded one, and it barely moves when the prompt list grows on the branded side.

The denominator has to stay fixed for a second reason. Answers vary between runs of the same question, so a rate built on one round per prompt is noisy. HubSpot's documentation states that its AEO tool runs tracked prompts daily and that responses change over time, which is why it asks teams to review several days or weeks before judging a trend. In Singapore the practical version is to treat the weekly review as the reading unit and the daily run as raw material.

A divided tray holds one tall pile of counters and one shallow pile, with a flat strip laid across the top.

Setting up two separate prompt sets

The setup stage decides whether later comparisons hold. Six steps, in this order.

Deduplicate the library. Merge prompts asking the same thing in different words, and keep one canonical wording per question.
Label each prompt by buying moment, using a vocabulary you can keep: category, problem, comparison, brand. Add a fifth label for prompts that name a competitor.
Attach one buyer profile and one market to each prompt, so a Singapore prompt does not get averaged against a regional one.
Assign questions to the sample they belong to. Brand-oriented questions form one set, issue-oriented questions another, and the two never share a denominator.
Fix the engine list. ChatGPT, Gemini and Perplexity are the defaults in HubSpot's AEO tool, which covers fewer engines than a consumer might use; premium engines are opt-in through a beta.
Fix the cadence and the question version before the first report, so the second period compares like with like.

Two counting rules from the platform level explain why sample size matters more than prompt count. HubSpot AEO lists 25 prompts on its standard plan, and Marketing Hub Enterprise lists 5,000 answers and 50 prompts. Daily runs on a fixed list produce far more answers than weekly runs of the same list, so a plan splitting prompts into two sets without checking answer capacity ends with a thin denominator on each side. Generated additions come in batches of 8-15 prompts at a time, which is one more way a list can outgrow the answers behind it.

HubSpot's product page prices AEO at $50/month, or $45/month when paid annually, including 25 prompts with more purchasable at any time. For a Singapore team paying in SGD, that is a starting reference for the tool budget, not a market benchmark.

Reading the two rates without fooling yourself

Report each set with its own denominator, then compare the pair. The HubSpot article suggests a branded-to-unbranded ratio, and a high ratio says people find you mostly after they already know the name; a ratio falling over several weeks is the better signal that discovery is improving.

The four numbers carry most of the story:

Number	How it is built	What a weak value suggests
Branded mention rate	answers mentioning you ÷ valid branded answers	the engine describes you wrongly or has too few facts
Unbranded mention rate	answers mentioning you ÷ valid unbranded answers	the category conversation does not include you
Owned citation share	answers citing your site ÷ valid answers in the set	other domains own the sources the engine leans on
Prompt coverage	prompts where you appear at least once ÷ tracked prompts	the list is broader than your content covers

Three habits keep the reading honest. Count each answer once, even when it repeats. Write the denominator next to every percentage, and say whether it counts answers or prompts. Keep "not mentioned" separate from "described negatively", because an answer that leaves you out is silence, not a verdict.

Answers that fail to return form a fifth category, and most aeo reports bury it in a footnote. A run can come back empty, and an empty run dropped from the denominator quietly raises the rate. Log failure counts per period and per engine.

A person sorts printed answer sheets into three trays on a table, leaving a small pile of blank sheets aside.

What is geo, and where does the branded split sit in it

Generative engine optimization is the practice of shaping what an AI engine retrieves, how it phrases a description and when it cites a brand in an answer to a buyer's question. It works on the sources the engine reads and the wording of the brand's own pages, so the result shows up inside the answer rather than as a position on a results page. None of the pages in the search results for this topic defined it as a discipline that connects to branded and unbranded tracking; that definition belongs here because the split is the measurement half of the same work.

The split sits inside the workflow at the tracking stage, before any content decision. When the unbranded rate is weak, ask which third-party pages and content types the engine cites for those prompts, because those are the sources a brand has to get into or answer better. When the branded rate is weak while the unbranded rate holds, the problem is usually description and facts rather than reach: the engine knows the name but says something thin or wrong. Two different fixes, two different sets of evidence.

Singapore adds a workflow consideration rather than a vocabulary one. Buyers here ask in English, often mix in Chinese-language product names, and compare regional options across a smaller pool of local sources. Running generative engine optimization singapore work on a global English prompt set alone will miss the comparison questions local buyers actually type, and the unbranded side will read lower than the brand's real standing in this market. For a wider view of how AI visibility differs from ranking, see [AEO vs. Search Rankings: AI Visibility Basics](https://www.topkee.com.sg/dynamic/ai-visibility-vs-search-rankings-is-your-agency-measuring).

What Topkee does with brand-oriented and issue-oriented samples

Topkee's GEO service is built around this distinction in its reporting rather than its package structure. Prompts Tracking compiles real buyer questions into a list and samples them on the agreed AI engines to see whether the brand appears and how it is described, and the service states that monitoring should revolve around real purchase situations rather than only tracking the brand name. Its aeo work therefore starts from the comparison questions a client was actually asked, not from a generic brand-word list.

The service treats the branded and unbranded cut as a measurement control. Issue-oriented and brand-oriented samples are kept apart, explicitly to avoid manufacturing lift by changing questions or inserting the brand name. Each monitoring report records the monitoring period, question version, engine, sample size and failure count, so a later period can be compared with an earlier one without guessing what changed.

Counting rules in the service guard the denominator. The tier counts are questions active at once, not monthly query volumes; one question run on several engines counts as sample volume rather than a new question; cross-language variants or alternative phrasings are managed as separate questions; and a replaced question keeps its version.

The sample report that GEO Insight produces fixes one engine, mode, region, language and question version, and shows the layout a client receives. It is a simulated screenshot used to show the report structure, not a client result, and the rounds in it illustrate the example rather than a contracted sampling frequency. The same monitoring console is where competitor and citation comparisons happen on the same question set, which is how a brand sees its own rate next to a rival's without switching questions partway through.

A wide monitor displays two mirrored result panels side by side on a desk, each with the same settings row.

A weekly review checklist

Run this once a week, with the same question set as the previous week.

Pull the branded and unbranded rates separately and note each denominator.
Compare against the last four weeks, not yesterday, because single runs fluctuate.
List the prompts that returned no usable answer and the engines that failed them.
Open five to ten captured answers and check how the brand is described, not only whether it appears.
Note the domains cited for the unbranded prompts where the brand was absent.
Record changes to the prompt library since the last review, with versions.
Write one decision next to each finding, with an owner and a date to look again.

A weekly pass of this kind is also the cheapest way to keep an aeo report from drifting between periods. If the reading points to a content gap rather than a prompt problem, the closed-loop view in AI-Powered Closed-Loop Marketing Boosts ROI shows how visibility findings connect back to production.

FAQ

What is the difference between a branded and an unbranded query in AI search?

A branded query contains the brand name, a variation or misspelling of it, or a brand-related product, following the definition Google Search Central published with its Search Console filter. An unbranded query names no brand and describes a category, a problem or a use case instead.

Why can't I use one blended visibility score?

Because brand-name prompts almost always produce a mention, while unbranded prompts have to earn one. Blending them raises the average with prompts that were never a discovery test, so the number can climb while the share of new buyers who meet the brand stays flat.

How many answers do I need before a rate means something?

Enough that a single answer does not move it much, and enough rounds that natural variation is visible. A published tool limit is not a coverage guarantee: HubSpot AEO lists 25 prompts on its standard plan, and that capacity has to serve both sets if you split the list.

Should a prompt that names a competitor count as unbranded?

No. It names a brand, just not yours, so it belongs in its own label. Keeping competitor-named prompts inside the unbranded set distorts both rates and hides the questions where a rival is being compared directly.

When is the split urgent rather than nice to have?

When the visibility number is used to judge whether new buyers are discovering the brand. If the figure only needs to describe how engines talk about a known brand, the branded set alone answers that; the unbranded set is what carries the discovery question.