An AI visibility score describes the prompt set behind it before it describes your market. A prompt that names your brand almost always gets an answer that names it. It tells you nothing about your competitive position, yet it counts like every other prompt. Once part of the set works this way, the figure you take to the executive team reflects your drafting choices as much as your standing.
Three decisions correct the bias: take those prompts out of the market score, build the set around intents rather than phrasings, and add up counters when you roll up, never average percentages.
One limit remains. As of October 27, 2026, no study reviewed here tests the same prompt with and without the brand name, and none gives the real distribution of questions people ask engines. The orders of magnitude below come from observational comparisons and arithmetic.
How much does a branded prompt inflate the score?
By several points from about 9% of the set, and the bias does not hit everyone equally. Pratyush Kumar, co-founder and shareholder of the vendor Ranqo, published a preprint in June 2026 on his platform's data: 102 brands and 102,025 answers from five engines, March to May 2026. On each brand's first run, ChatGPT names it on 94.2% of prompts containing its name and 22.1% of the others. Gemini goes from 94.0% to 18.7%, Perplexity from 98.5% to 23.9%.
Branded prompts make up only about 9% of responses. That is enough to lift ChatGPT from 22.1% to 28.6%. The split by brand stature matters most:
| Brands, by stature | Prompts without the name | All prompts | Gap | Relative gap |
|---|---|---|---|---|
| Best known (11) | 72.9% | 73.5% | +0.6 pt | +1% |
| Mid-tier (36) | 43.6% | 52.2% | +8.6 pts | +20% |
| Least known (55) | 11.4% | 17.3% | +5.9 pts | +52% |
Source: Kumar, Ranqo, first run, March-May 2026. Relative gap: author's calculation.
A well-known brand reads about right. A little-known one sees its score inflated by half, and may believe it is established in a market where it remains marginal. Take the result for what it is: a comparison across one vendor's customers, with no test of the same prompt with and without the name.
Arithmetic does the rest. Assume 95% presence on branded prompts, close to the floor Kumar reports:
| True presence on generic questions | 10% branded prompts | 20% | 30% |
|---|---|---|---|
| 5% | 14% | 23% | 32% |
| 20% | 27.5% | 35% | 42.5% |
| 50% | 54.5% | 59% | 63.5% |
Arithmetic illustration, author's calculation.
Going from 10% to 20% branded prompts adds 7.5 points to a brand whose presence has not moved. That is about the margin on a 100-prompt set (±7.8 points at 20%, see the full calculation), so a mere change of prompts can pass for a win.
A score computed only on answers naming a brand does not escape this, because those answers enter both numerator and denominator. Illustrative example, fictional brand: on 80 generic prompts, 50 answers name a brand and 10 of them name Alba & Co, or 20%. Add 20 branded prompts, so 20 answers naming a brand, 19 of them Alba & Co, and the score becomes 29 out of 70, or 41.4%.
What is a branded prompt?
A branded prompt is a prompt whose text gives the engine the name of the brand being measured, one of its products or one of its aliases. The real category is wider than the name alone:
- The head-to-head. "Alba & Co or Maison Lin for a couple?" sets the candidates. In Kumar's data, comparison prompts show 74.6% presence against 22.9% for discovery prompts.
- The named rival. "Alternatives to Maison Lin" does not cite Alba & Co, but it defines the market from a competitor and measures a different question.
- House vocabulary. A technology name or a phrase your marketing coined steers the engine to your pages without saying your name. This is reasoning rather than a published measurement, and only a reader from outside the brand tends to catch it.
These prompts also expose you to a false signal: an engine that returns the prompt as is appears to name the brand, as the article on false signals explains.
Should you delete branded prompts?
No: take them out of the market score and read them separately. They answer a real question, namely what engines tell someone who already knows you. Kumar's point is that coverage on these prompts confirms the brand exists for the engine, without measuring how it stands against competitors.
| Market set | Brand set | |
|---|---|---|
| Question | Are you cited when nobody names you? | What is said about you when you are named? |
| Useful metric | Presence, share of voice, gap to competitors | Accuracy, sentiment, sources used |
| Decision | Budget, editorial priorities | Fixing errors, reputation |
The brand set feeds the work on engine errors. The two sets produce two figures, never one. There is no correct mix between them, since the real share of such questions is unknown.
How does a prompt set get biased without naming anyone?
Three ways: the mix of intents, the slope of the wording, and picking prompts after the fact.
The mix of intents. In Kumar's data, presence ranges from 22.9% on discovery prompts to 15.3% on expert questions, 11.4% on use cases and 10.8% on a problem to solve. Semrush, with Kevin Indig (June 2026, 115 prompts, 14 countries, four engines), measures something else, the share of domains present in an answer whose brand is also named, and finds the same slope: 43.3% on comparative queries, 18% on informational ones. A set packed with "best mattress" can show double the score of one built on concrete problems, with neither being false.
The slope of the wording. Lin and co-authors (preprint revised February 2025) show that synonyms picked by optimization can shift by up to 78% the likelihood that a model mentions a target concept, such as a brand. This is a lab attack rather than ordinary drafting, but it establishes that words carry weight. The everyday forms show up on a read-through:
| Slanted wording | What it imposes | Neutral rewording |
|---|---|---|
| Why are organic mattresses better? | The conclusion | How does an organic mattress differ from a standard one? |
| Which mattress brands should I avoid? | The negative pole | How do I compare mattress brands? |
| Is a $900 mattress worth it? | A reference price | What budget should I plan for a good mattress? |
Illustrative examples.
Neutral does not mean realistic: your customers ask slanted questions too. The goal is a set balanced across poles, where each slant is a deliberate choice.
Picking after the fact. Removing the prompts where you are absent because they "aren't relevant" manufactures a gain. Write the set before seeing results, and date every change.
Then there is the buyer the prompt assumes. Will Jack and co-authors, affiliated with Unusual (May 2026, 2,000 runs on three model configurations), prefix the same prompts ("best CRM software", for instance) with 10 buyer profiles. Similarity of recommended brands falls by 0.12 to 0.20 on a 0 to 1 scale. Leaders hold, at about 80% consistency, while mid-market brands see up to 75% of recommendations change. As in Kumar's points, the middle of the market moves most, and a set written for a single profile describes it poorly.
How do you build a prompt set that measures the market?
Intents first, wording second. SparkToro (January 2026) collected 142 prompts written by people for one need, choosing headphones for a family member who travels: mean semantic similarity was only 0.081. Yet Bose, Sony, Sennheiser and Apple appear in 55 to 77% of the 994 responses, while other needs (gaming, podcasting) return very different answers. One category, and a co-author who sells tracking, but the lesson holds: intent shapes the answer.
- One intent, one topic. Start from customers' words: sales calls, support tickets, on-site search, Search Console queries. Avoid the product page, which speaks like you.
- About 100 prompts, at least 10 topics. This benchmark comes from the margin calculation: at 100 prompts, going from 5 to 10 topics gains up to 3.7 points of margin, depending on how similar prompts are assumed to be. A shorter list of 30 to 50 questions still suits a first decision audit: it shows who appears, not movement. For a large brand, add a level above topics rather than piling on rewordings.
- Several registers per topic. Short keyword ("mattress couple back pain"), direct request, full question, chatty, situated ("I've slept badly since moving house..."). Nobody knows the real proportions of these registers, so vary them without claiming to reproduce reality.
- The customer's language. The prompt's language sets the answer's language: on a project set to French, the Nanga team saw 50 English prompts return 50 English answers. Chen and co-authors (September 2025) also note that engines differ in their stability across languages. Adapt each set rather than translating it, and test language separately.
- Tags from day one. Brand, competitor, buying stage, profile. Without them, none of the checks below is possible.
Why add up counters instead of averaging scores?
Because an average of percentages ignores what each percentage stands on. Take a score computed on answers naming a brand:
| Topic | Answers | Name a brand | Name Alba & Co | Score |
|---|---|---|---|---|
| A | 40 | 4 | 2 | 50% |
| B | 40 | 36 | 3 | 8.3% |
| C | 40 | 0 | 0 | Not measured |
Illustrative example, fictional brand.
The average of A and B is 29.2%. Count C as zero and it drops to 19.4%. The counters give 5 answers out of 40, or 12.5%. Three figures for the same answers, and only one respects its base. Measured against all answers, the count is 5 out of 120, or 4.2%: a different metric, just as valid, as long as you say which. An answer naming no brand is not a failure of your brand, and a parent topic should read exactly what its sub-topics read.
Adding is not neutral either: each topic weighs according to its number of prompts. With 40% presence on your strong topic and 10% on your weak one, a 60/40 split shows 28% and a 30/70 split shows 19% (illustrative arithmetic). The split is therefore a business decision. Set it by the economic weight of each subject, write it down, and only compare sets with the same structure.
Which figure belongs in front of the executive team?
The market score, alone, with its definition. Branded prompts name you and not your competitors, so they distort the comparison first, which is what matters most for a budget. Measure competitors on the same set and read the gap.
The day you split the sets, the figure drops. That is not lost visibility, it is a series break: show both calculations for a quarter and note the date. Better to explain a drop once than to defend a figure that fails the first question.
How do you clean up an existing prompt set?
- Reread the set. Tag every prompt containing your name, an alias, a product or a competitor, then look for house vocabulary.
- Recompute two scores. Market and brand, from the counters. The gap gives you the size of the bias.
- Remove each topic in turn. If one topic moves the score by more than its margin, that topic is what you are reading.
- Run the set against your competitors. If you lead everywhere while the market says otherwise, suspect the set before celebrating.
- Freeze and date. Every change to the set goes in a log, next to the curves.
These rules wear down as a prompt set lives: every prompt added can reopen the bias. Nanga's GEO module enforces them at the moment the set changes. When you add prompts that name your brand, a warning notes that they raise the AI Visibility Index and that they are better kept in a separate project to read neutral questions. "Analyse bias" scores each wording from -3 (risk) to +3 (benefit), names the marker it found and offers a neutral rewording, which is never applied without your say. In the filter bar, one click includes a tag and a second excludes it.
Topics sit on three levels (category, topic, sub-topic), and a parent topic recomputes its figures from its sub-topics' counters. Tracking is weekly by default, daily as an option.
To see your market score, separated from your branded prompts: request a demo.
Frequently asked questions
What is a branded prompt?
A prompt whose text gives the engine the name of the brand being measured, one of its products or one of its aliases. The answer almost always repeats it: 94% to 100% depending on the engine, according to Pratyush Kumar's June 2026 preprint.
Do branded prompts really lift the score?
Yes. In the Ranqo data published by Pratyush Kumar, about 9% branded prompts move little-known brands from 11.4% to 17.3% and mid-tier brands from 43.6% to 52.2%. The best known barely move.
What share of branded prompts is acceptable in a market score?
None. There is no correct dose, because no source reviewed as of October 27, 2026 gives the real share of such questions among users. Keep them in a separate set, used to read what engines say about you.
How many prompts does it take to measure the market?
To follow a trend, about 100 prompts across at least 10 topics is a reasoned benchmark, not a standard. A list of 30 to 50 questions is enough for a first audit: it shows who appears, but it cannot show movement.
Can you average the scores of several topics?
No. An average gives a topic with 4 usable answers the same weight as one with 36, and reads a topic where no brand is named as a zero. Add up the answers naming your brand and the base, then divide once.
Sources:
- Pratyush Kumar, Ranqo, Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines, preprint, June 2026.
- Semrush, with Kevin Indig, Why 62% of AI citations don't lead to brand mentions, June 2026.
- Lin et al., LLM Whisperer: An Inconspicuous Attack to Bias LLM Responses, preprint, June 2024, revised February 2025.
- Jack, Lehman, Maloney and Xu, Unusual, Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat, preprint, May 2026.
- SparkToro, NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility, January 2026.
- Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas, Generative Engine Optimization: How to Dominate AI Search, preprint, September 2025.