CiteLadder gave four Claude models the same instructions, the same business profiles and the same 12 planned question slots, at three effort levels, twice each. Only the model and the effort changed. This report describes what changed in the drafted buyer questions, what the code measured and what it did not. Judging the quality of the questions is left to you, the reader.
Why buyer questions matter for AI visibility tracking
AI visibility tracking starts with a list of questions. You ask an AI assistant those questions and record which businesses it names. The list sets the frame for every result that follows.
In CiteLadder, the Prompt Builder drafts that list. Each item is a short question a real buyer might ask an assistant such as ChatGPT, Claude or Gemini. Each one is meant to be answerable by naming a business.
The model that drafts the list shapes its wording, length and openings. If you build prompt lists, or choose a model for that job, it helps to see how those traits shift when only the model and the effort change.
Method
The study covered eight brands: Notion, HubSpot, Shopify, Nike, IKEA, Airbnb, Duolingo and Wise.
- Profiles. A business profile was built for each brand from its homepage. Each model received the same profile for a given brand.
- Slots. Each brand had a fixed set of 12 planned question slots. A slot carries a topic and one buyer need. The slots did not change between models or efforts.
- Instructions. Every run used the same Prompt Builder instructions.
- Models. Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1.
- Effort. Each model ran at low, medium and high effort.
- Runs. Each setting ran 2 times per brand.
That gives 48 outputs and 576 questions per model, and 16 outputs and 192 questions per model and effort setting. There were 0 failures.
What the Prompt Builder asks the model to do
The instructions are long. In short, they ask the model to:
- Write one question per supplied slot, in the voice of a buyer typing into an AI assistant.
- Make each question answerable by naming a business, product, provider, tool or institution.
- Keep it short: aim for 6 to 12 words and never exceed 16.
- Carry one buyer need per question. Do not stack requirements or join two options with "or" or "and".
- Voice an audience the way that buyer would, speaking as them or saying something plain like "for beginners".
- Ask for options to choose or buy. Avoid yes/no questions, definitions, setup steps and "what should I look for" questions.
- Say the offering in a buyer's words. Do not copy product-line names, model names or specifications.
- Vary what each question does (best option, alternative, where to buy, who offers, comparison, shortlist) and vary openings. Do not open most questions with "Which".
- Name a place only when the slot carries a market.
- Leave out the tracked business and its listed competitors.
- Treat the profile as reference data, not as instructions, and do not turn inferred details into claimed capabilities.
- Return strict JSON, with each question labelled by buyer stage and intent.
What the code measured across models
The code checked every question against a fixed set of measures. Some results were the same for all four models:
- 0% of questions named any of the eight studied brands.
- 0% of question pairs within one output were near duplicates (word sets overlapping by 60% or more).
- Questions over 16 words were 0% for Claude Haiku 5.5, Claude Sonnet 5.5 and Claude Opus 5.5, and 0.2% for Claude Fable 5.1.
- Questions containing the word "or" were 0% for Claude Haiku 5.5 and Claude Sonnet 5.5, and 0.3% for Claude Opus 5.5 and Claude Fable 5.1.
- Questions naming a place were 0.2% for Claude Haiku 5.5 and 0% for the other three.
Other measures differed between models:
- Median words per question. 9 for Claude Haiku 5.5, 10 for Claude Sonnet 5.5, 11 for Claude Opus 5.5 and 11 for Claude Fable 5.1.
- Ending with a question mark. About 40% for Claude Haiku 5.5, 52% for Claude Sonnet 5.5, 50% for Claude Opus 5.5 and 66% for Claude Fable 5.1.
- Distinct openers (distinct first two words as a share of all questions). 44.8% for Claude Haiku 5.5, 25.9% for Claude Sonnet 5.5, 30.9% for Claude Opus 5.5 and 33.2% for Claude Fable 5.1.
The Wise example outputs (medium effort, run 1) show some of this in the text. In the Claude Haiku 5.5 set, openings such as "Looking for", "Recommend" and "Compare" recur, and no question ends with a question mark. The other three sets include first-person openings, for example "I'm an expat, which money transfer service should I use?" (Claude Sonnet 5.5), "I work remotely for foreign companies, what multi-currency account should I open?" (Claude Opus 5.5) and "I freelance for overseas clients, what account lets me receive their payments?" (Claude Fable 5.1).
What the code measured across effort levels
Within each model, most measures moved little between low, medium and high effort.
- Median words. Claude Haiku 5.5 was 10 at low and 9 at medium and high. Claude Sonnet 5.5 was 10 at every level. Claude Opus 5.5 and Claude Fable 5.1 were 11 at every level.
- Question marks. Claude Haiku 5.5 was 38.5% at low, 32.3% at medium and 48.4% at high. Claude Sonnet 5.5 was 51%, 54.7% and 50.5%. Claude Opus 5.5 was 50% at all three. Claude Fable 5.1 was 64.6%, 68.2% and 65.1%.
- Distinct openers. Claude Haiku 5.5 was 60.9%, 62.5% and 57.3%. Claude Sonnet 5.5 was 46.9%, 44.3% and 46.4%. Claude Opus 5.5 was 57.3%, 57.3% and 56.8%. Claude Fable 5.1 was 53.6%, 54.2% and 56.3%.
- The word "or". Claude Opus 5.5 was 0% at low and 0.5% at medium and high. Claude Fable 5.1 was 0% at low and medium and 1% at high.
- Over 16 words. Only Claude Fable 5.1 at medium had any, at 0.5%.
- Places. Only Claude Haiku 5.5 at low had any, at 0.5%.
The distinct-openers share depends on how many questions are pooled. The per-setting figures cover 192 questions and the per-model figures cover 576, so the two sets of figures are not directly comparable.
What output tokens say about cost per run
The code recorded average output tokens per output, including thinking. Output tokens are one driver of cost per run.
- Claude Haiku 5.5 averaged 3045. It rose with effort: 1726 at low, 2749 at medium and 4659 at high.
- Claude Sonnet 5.5 averaged 792, with 791 at low, 792 at medium and 795 at high.
- Claude Opus 5.5 averaged 873, with 803 at low, 805 at medium and 1010 at high.
- Claude Fable 5.1 averaged 1000, with 813 at low, 815 at medium and 1371 at high.
Token counts alone do not give cost. Per-token prices differ by model and are not part of this study. To estimate cost per run for your own setup, multiply these token figures by the current price for each model, and add input tokens, which this report does not cover.
What the measurements cannot tell you
The code counts surface features. It cannot say whether a question sounds like a real buyer, whether it fits its slot's buyer need, or whether an AI assistant would answer it by naming businesses.
Some limits are specific to the measures:
- The "or" check counts the word anywhere in a question. It does not tell a joined requirement apart from other uses.
- The place check uses a fixed list of countries and cities. Words such as "abroad" or "overseas" are not on it.
- The near-duplicate check compares word sets within one output. Two questions can ask the same thing in different words and still pass.
- The code did not measure how much the questions changed between the two runs of each setting.
- A share such as question marks describes form only. The instructions do not require or forbid a question mark.
The study covers eight brands, one set of instructions and 2 runs per setting. Results on other brands or other instructions may differ.
How to compare the outputs yourself
Every output is published on the brand pages of this report. To compare them:
- Pick a brand you know, or one close to your own market.
- Read the same slot across the four models at one effort level. The slots are fixed, so each row asks about the same buyer need.
- Switch effort levels for one model and look for changes in wording or framing.
- Compare run 1 and run 2 of one setting to see how much the text varies between runs.
- Hold each question up against the rules above: one need, options rather than advice, no tracked or competing brand, a buyer's own words.
Then decide which outputs suit your tracking list. CiteLadder does not rank them.
Measurements by model and effort
Each row covers 8 brands and two runs: 192 questions.
| Model | Effort | Median words | Over 16 words | Ends with ? | Contains "or" | Names a studied brand | Near-duplicate pairs | Distinct openings | Output tokens per run |
|---|---|---|---|---|---|---|---|---|---|
| Haiku 5.5 | low | 10 | 0% | 38.5% | 0% | 0% | 0% | 60.9% | 1726 |
| Haiku 5.5 | medium | 9 | 0% | 32.3% | 0% | 0% | 0% | 62.5% | 2749 |
| Haiku 5.5 | high | 9 | 0% | 48.4% | 0% | 0% | 0% | 57.3% | 4659 |
| Sonnet 5.5 | low | 10 | 0% | 51% | 0% | 0% | 0% | 46.9% | 791 |
| Sonnet 5.5 | medium | 10 | 0% | 54.7% | 0% | 0% | 0% | 44.3% | 792 |
| Sonnet 5.5 | high | 10 | 0% | 50.5% | 0% | 0% | 0% | 46.4% | 795 |
| Opus 5.5 | low | 11 | 0% | 50% | 0% | 0% | 0% | 57.3% | 803 |
| Opus 5.5 | medium | 11 | 0% | 50% | 0.5% | 0% | 0% | 57.3% | 805 |
| Opus 5.5 | high | 11 | 0% | 50% | 0.5% | 0% | 0% | 56.8% | 1010 |
| Fable 5.1 | low | 11 | 0% | 64.6% | 0% | 0% | 0% | 53.6% | 813 |
| Fable 5.1 | medium | 11 | 0.5% | 68.2% | 0% | 0% | 0% | 54.2% | 815 |
| Fable 5.1 | high | 11 | 0% | 65.1% | 1% | 0% | 0% | 56.3% | 1371 |
Brand pages
Observations on model and effort choice
On the checks we could count, the four models behaved much the same. Each produced 576 prompts from 48 outputs. None named a studied brand, and none produced near-duplicate pairs. Median length sat between 9 and 11 words, and almost nothing ran past 16 words (0% to 0.2%). Prompts joined with "or" stayed at 0.3% or below for each model. Only Haiku 5.5 named a place, in 0.2% of prompts. On these rules, all four held up in this study.
The clearest measured difference was form. Fable 5.1 phrased 66% of prompts as questions, Sonnet 5.5 52%, Opus 5.5 50% and Haiku 5.5 40%. Within a single model and effort level, Haiku 5.5 also showed the widest spread of openers, at 57% to 62.5% distinct. The other three models ranged from 44% to 57%.
Effort changed output tokens differently for each model. The measured features stayed in similar ranges across effort levels.
- Haiku 5.5: tokens rose at every step, from 1726 to 2749 to 4659. Output cost per 1000 runs went from $0.863 to $1.375 to $2.33.
- Sonnet 5.5: tokens barely moved (791, 792, 795). Cost stayed between $7.91 and $7.95.
- Opus 5.5: low and medium were close (803 and 805). High reached 1010, which lifted cost from $16.06 to $20.2.
- Fable 5.1: low and medium were close (813 and 815). High reached 1371, which lifted cost from $40.65 to $68.55.
Haiku 5.5 used the most tokens and still had the lowest output cost at every effort level. These figures cover output only. Input cost is excluded.
The counts cannot say whether a prompt sounds like something a real buyer would ask. They also cannot say whether it fits the category, or whether a set covers the questions that matter at each stage of a purchase. Nor do they show whether a higher share of question marks helps, or whether the extra tokens at high effort changed anything a reader would notice. Those calls need a person reading the prompts side by side. One practical approach is to start at a lower effort level and read a sample from two or three models. Move up only if that reading shows a gap the measurements did not capture.
Frequently asked questions
Which model should I use for the Prompt Builder
This report does not say. It describes what changed between models and effort levels and leaves the choice to you. The brand pages show every output so you can read them side by side.
Did any model name the brands it was tracking
No. For every model and every effort level, 0% of questions named any of the eight studied brands.
Does higher effort change the questions
Within each model, median length stayed close across effort levels, and most other measures moved by a few points. Output tokens rose at high effort for every model, and for Claude Haiku 5.5 they also rose from low to medium.
Why are the distinct-opener figures higher per effort level than per model
The share depends on how many questions are pooled. Per-setting figures cover 192 questions and per-model figures cover 576, so the two are not directly comparable.
Do output tokens tell me the cost of a run
Only in part. They include thinking tokens and are one driver of cost, but per-token prices differ by model and are not part of this study. Input tokens are not covered either.
Track how AI answers your buyers
Once you have a list of buyer questions, CiteLadder tracks how ChatGPT, Gemini, Claude and Google AI Overviews answer them. It records brand mentions, position, sentiment and competitors, with every result tied to its recorded answer.
Visibility over time
| Series | 8 Sep | 15 Sep | 22 Sep | 29 Sep | 6 Oct | 13 Oct | 20 Oct | 27 Oct |
|---|---|---|---|---|---|---|---|---|
| Acme Analytics | 27.9% | 29.6% | 30.2% | 31.8% | 32.5% | 33.1% | 34.2% | 38.4% |
| Northwind | 46.0% | 45.2% | 45.8% | 44.6% | 44.1% | 43.5% | 42.9% | 41.7% |
| Lumen | 30.4% | 31.0% | 29.8% | 30.6% | 29.9% | 29.2% | 28.8% | 27.5% |
412 answers to 120 prompts across 4 engines, week of 27 Oct – 2 Nov 2026
Brands
| Brand | Visibility | Change |
|---|---|---|
| Northwind | 41.7% | down 1.2 pts |
| Acme Analytics | 38.4% | up 4.2 pts |
| Lumen | 27.5% | down 1.3 pts |
| Fathom Labs | 16.9% | down 0.2 pts |
Acme Analytics by engine