An unbranded buying prompt asks an AI engine for help choosing, such as a comparison or alternatives question, without naming your brand. Unbranded prompts are the only kind worth tracking for visibility, because a prompt that names your brand supplies the name itself, so a mention shows nothing about whether the engine would bring you up.
- Penlid's methodology says its prompts are deliberately unbranded, drawn from the comparison, alternatives and buyer's-guide questions a real customer types (Penlid, Methodology, read 30 October 2026).
- Penlid's free AI visibility check asks one engine, Claude with live web search, four buying questions, so one result rests on four samples (Penlid, AI visibility check tool page, read 30 October 2026).
- OpenAI says that when ChatGPT search partners with other search providers, it typically rewrites your query into one or more targeted queries (OpenAI, OpenAI Help Center, read 30 October 2026).
- An unbranded rate of 3 answers in 15 has a 95% Wilson interval of 7.0% to 45.2% (our arithmetic with the method in Penlid's methodology, read 30 October 2026).
What is an unbranded buying prompt?
An unbranded buying prompt is a question someone asks an AI engine while choosing what to buy, with your brand name left out. In this guide, a prompt that names a competitor still counts as unbranded, because the name being measured is the one that is absent. "Buying prompt" is our term, not an industry standard.
Penlid's methodology says Prompts are deliberately unbranded.
Penlid asks the comparison, alternatives and buyer's-guide questions a real customer types
, and the table sorts those kinds from the branded ones it leaves out.
| Prompt | Kind | Names your brand? | Tracked for visibility? |
|---|---|---|---|
| What is the best invoicing tool for a freelancer? | Buyer's guide | No | Yes |
| Is a spreadsheet or invoicing software better for a freelancer? | Comparison | No | Yes |
| What are the alternatives to Rivalco? | Alternatives | No, it names a rival | Yes |
| Is Acme any good? | Branded question | Yes | No |
| Which is better for a freelancer, Acme or Rivalco? | Branded comparison | Yes | No |
Why does a branded prompt tell you nothing about visibility?
A branded prompt tells you nothing about visibility because your name is already in the question, so a mention in the answer shows only that the engine answered what it was asked. Penlid's methodology gives "is Acme any good?" as its example of a question that measures nothing. Visibility is the case where a buyer who has not named you asks for help and the engine offers you anyway, and only an unbranded prompt can test that. An unbranded answer also shows which competitors are named instead of you.
Two more effects are our reasoning, not published results, and the example below works both through. A branded prompt in a tracked set lifts the average whatever the engine thinks of you, and it mutes change: with half the prompts branded, the blended rate moves half as far as the rate that matters.
A third reason has outside support, with a caveat. A preprint reports that ChatGPT writes brands it knows into its searches
. If so, an unbranded prompt also tests whether ChatGPT reaches for your name before anyone supplies it, which is our reading of the finding. It is one preprint, its finding is stated for ChatGPT, and we could not confirm that it has been peer reviewed.
What do AI engines do with a prompt before they answer?
Some engines turn a prompt into other searches first: OpenAI says that when ChatGPT search partners with other search providers it typically rewrites your query into one or more targeted queries
, and Google says AI Overviews and AI Mode may use query fan-out, which its optimization guide defines as A set of concurrent, related queries generated by the model
to fetch more search results.
Bing's AI Performance help page makes a related point: its grounding queries are short, grouped phrases, described as the key phrases the AI used when retrieving content that was cited in an answer, and They are not full user questions or prompts.
In our judgment that is a reason to write your own prompts as a buyer would phrase them and let the engine choose its searches, since, if the preprint above is right about ChatGPT, which brands it puts into them is part of what you measure. Answers also vary from run to run, so one ChatGPT answer is one draw from many; our guide to checking whether ChatGPT recommends your brand develops that point.
How do you write unbranded buying prompts?
Write each prompt as a buyer would phrase it before knowing your name, keep your brand out of it, and keep the wording fixed. The rules below are our reasoning, except where a source is named.
- Keep every brand cue out. A product or feature name only you use, or your domain, makes a prompt branded in effect.
- Use real buyer wording. Sales calls and support questions show how buyers phrase a need; add the buyer's situation, not just a category keyword.
- Do not reword between runs. A preprint says
Answers can vary across runs, prompts, and time, making one-off observations unreliable.
A reworded prompt is a new prompt. - Show every prompt. Penlid's methodology says
We show you every prompt in full so you can judge the sample yourself.
Penlid's free check also reports being recommended and being listed as a source separately, because its page calls them different outcomes.
Penlid's free AI visibility check reads your homepage to work out what you sell, asks four buying questions and shows you the exact prompts.
How many answers does an unbranded prompt need?
An unbranded prompt needs as many answers as the precision you want, because each answer is one sample and the samples differ. Penlid's methodology says it samples each tracked prompt repeatedly on a schedule and reports each rate with a 95% Wilson interval. The table gives its three published examples of a 30% rate, plus the two ends of the free check's four answers, which we computed with the formula in the example.
| Result | Answers | 95% interval | From |
|---|---|---|---|
| 0 of 4 | 4 | 0.0% to 49.0% | Our arithmetic |
| 4 of 4 | 4 | 51.0% to 100.0% | Our arithmetic |
| 3 of 10 | 10 | 10.8% to 60.3% | Penlid's methodology |
| 30 of 100 | 100 | 21.9% to 39.6% | Penlid's methodology |
| 300 of 1,000 | 1,000 | 27.2% to 32.9% | Penlid's methodology |
The free check's page says four samples is four samples and we won't pretend otherwise
. Four answers separate only the extremes: 0 of 4 and 4 of 4 do not overlap, but every other pair of results does, so Penlid's overlap rule would call them no significant change. How many answers a given width takes is covered in our guide to measuring AI visibility.
Is a branded prompt ever worth asking?
A branded prompt is worth asking for a different question: what an engine says about you once someone asks about you by name. That is a check on accuracy and reputation, not a measure of visibility, so in our judgment those prompts belong in a separate list and never in a visibility rate.
Example (illustrative numbers)
These numbers are invented and are not data from any site. Acme tracks six prompts on one engine, three branded and three unbranded, and asks each prompt 5 times. An answer is a hit when it names Acme, so there are 6 × 5 = 30 answers.
| Prompts | Answers | Answers naming Acme | Rate |
|---|---|---|---|
| Branded | 3 × 5 = 15 | 15 | Set by the question |
| Unbranded | 3 × 5 = 15 | 3 | 3 ÷ 15 = 20% |
| All six | 30 | 18 | 18 ÷ 30 = 60% |
The blend hides the answer. In this example, branded prompts name Acme in every answer because the question does, so they add 15 hits whatever the engine thinks of Acme. The blended rate is 60%, which is 60 − 20 = 40 points above the unbranded rate, and 15 ÷ 18 = 83% of its hits come from questions that supplied the name.
The blend mutes change. The 15 branded hits never move, so only the unbranded hits can. If they double from 3 to 6, the unbranded rate goes from 3 ÷ 15 = 20% to 6 ÷ 15 = 40%, a 20-point change, while the blended rate goes from 18 ÷ 30 = 60% to 21 ÷ 30 = 70%, a 10-point change.
The interval. For the unbranded 3 of 15, p = 3 ÷ 15 = 0.20, z = 1.96 and z² = 3.84. The denominator is 1 + 3.84 ÷ 15 = 1.256. The center is (0.20 + 3.84 ÷ 30) ÷ 1.256 = 0.328 ÷ 1.256 = 0.261. The half-width is 1.96 × √(0.20 × 0.80 ÷ 15 + 3.84 ÷ 900) ÷ 1.256 = 1.96 × √(0.01067 + 0.00427) ÷ 1.256 = 1.96 × 0.1222 ÷ 1.256 = 0.191. The interval runs from 0.261 − 0.191 = 0.070 to 0.261 + 0.191 = 0.452, or 7.0% to 45.2%.
The honest report is 20%, plausibly 7.0% to 45.2%, not 60%. Fifteen answers are also too few to see a large rise. By the same formula, 11 of 15 gives 48.0% to 89.1%, the first result whose interval clears 45.2%, so Penlid's overlap rule would report no significant change for any rise from 3 hits to fewer than 11.
What this cannot tell you
Unbranded prompts show only what an engine answered to the prompts you chose, and Penlid's methodology says There is no neutral prompt set.
A different list of buyer questions gives a different rate.
- They cannot say why you were or were not named. Penlid's methodology lists your content, a model update, a change in the engine's retrieval and a competitor's PR win as possible causes (see whether AI citations can be attributed to a content change).
- The interval covers sampling noise for the prompts you asked. It does not cover which prompts you picked, personalization, or a model change inside the window, which Penlid reads as two measurements (see why AI visibility moves when you change nothing).
- They say nothing about clicks, visits or sales, because a mention in an answer is not a visit.
- As its page reads on 30 October 2026, Penlid's free check samples one engine, Claude with live web search, so it says nothing about ChatGPT, Perplexity, Gemini or Google AI Overviews.
Questions people also ask
What is the difference between branded and unbranded prompts?
A branded prompt contains your company or product name, for example a question about whether it is any good, while an unbranded prompt asks for a category, a comparison or alternatives and leaves the name out. Because an engine answering a branded prompt starts from your name, only the unbranded kind shows whether you come up when a buyer does not know you yet.
What is a prompt in AI visibility tracking?
In AI visibility tracking, a prompt is a fixed question you put to an AI engine again and again to see which brands it names. It is the starting point, not the whole search: where ChatGPT search partners with outside search providers, OpenAI says it typically turns a question into one or more targeted queries, so the searches behind an answer can differ from your wording.
How do you track AI brand visibility?
Write a fixed list of unbranded buying prompts, ask each one many times on every engine you cover, record "recommended" and "cited" as separate results, and report each rate with an interval and the model version behind it. Penlid's methodology describes the sampling, interval and model-version steps, and our guide to measuring AI visibility explains them.
Is there a free AI brand visibility checker?
Penlid's free AI visibility check is one: it takes a domain name, needs no account, asks Claude with live web search four buying questions and shows who gets recommended, with 95% Wilson intervals. On 30 October 2026 it samples one engine, not four, and its results name the engines that actually answered.
Sources
- Searching the web with ChatGPT, OpenAI Help Center, OpenAI, read 30 October 2026.
ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers.
- AI features and your website, Google Search Central documentation, Google, read 30 October 2026.
- Optimizing your website for generative AI features on Google Search, Google Search Central documentation, Google, read 30 October 2026.
A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query.
- AI Performance, Bing Webmaster Tools Help, Microsoft, read 30 October 2026.
They are not full user questions or prompts.
- What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), arXiv preprint, read 30 October 2026.
ChatGPT writes brands it knows into its searches
- Don't Measure Once: Measuring Visibility in AI Search (GEO), arXiv preprint, read 30 October 2026.
Answers can vary across runs, prompts, and time, making one-off observations unreliable.
- Methodology: how we measure, and what we can't, Penlid, read 30 October 2026.
Prompts are deliberately unbranded.
We ask the comparison, alternatives and buyer's-guide questions a real customer types.
There is no neutral prompt set. We show you every prompt in full so you can judge the sample yourself.
- Free AI Visibility Check, Penlid, read 30 October 2026.
four samples is four samples and we won't pretend otherwise