Five prompt-to-citation experiments for proving AI discovery impact

Monitoring an AI answer is useful. Proving that a specific optimisation changed that answer is far more valuable.

This evidence pack sets out five reproducible, shopper-style experiments designed to test whether content and source improvements can influence how a brand is mentioned and cited in ChatGPT- and Gemini-style product discovery answers. The method is deliberately transparent: the exact prompt, the optimisation applied, the answer format and the citation outcome are all recorded.

Quadrant is built around this distinction. Its platform tracks brand and product visibility across AI environments, identifies cited sources and connects prompt-level findings with content recommendations. (projectquadrant.com)

The examples below use a controlled demonstration format. The figures are illustrative and are included to show how a robust testing framework can be documented.

Why reproducible evidence builds trust

Marketing leaders are increasingly being asked to justify investment in AI search visibility tools and AI search monitoring tools while the buying journey itself is still evolving. A dashboard can show that a brand appeared less often. On its own, it cannot show what changed or whether that change influenced product discovery.

A reproducible test creates a clearer chain of evidence:

1. Define a realistic shopper question.
2. Capture the baseline answer.
3. Make one specific content or source improvement.
4. Rerun the same question under the same conditions.
5. Compare mentions, citations and competitive presence.

This is more useful than a broad claim that a platform “improves visibility”. It gives SEO, commerce, content and analytics teams something they can inspect, repeat and challenge. Quadrant describes this operating model as a loop of asking, analysing, finding insights and executing improvements, rather than simply reporting a visibility score. (projectquadrant.com)

How the tests were run

The framework below shows how an evidence pack can be structured. In a live programme, teams should retain the full run log, including model version and cited URLs.

Test condition	Standard used
Platforms	ChatGPT-style and Gemini-style answer environments
Market	United Kingdom, English-language prompts for this demonstration
Prompt wording	Exact wording retained between baseline and rerun
Baseline	Answer captured before the documented optimisation
Optimisation	One clearly defined content or source improvement
Rerun	Same prompt, market and answer environment after implementation
Measurement	Citation presence per run, citation rate, brand mention and cited URL
Recommended repetitions	Five runs per condition, where platform access allows
Citation rate	Runs containing at least one relevant citation to the brand divided by total runs
Observation window	Record exact dates and time zone because AI answers can change

A citation should count only when the answer links to, names or otherwise identifies a relevant source that supports the product or brand statement. A brand mention without a supporting citation should be recorded separately. This distinction matters because being named and being used as a source are different discovery outcomes. (geoblog.projectquadrant.com)

Five shopper-style prompts, before and after

These five experiments cover recommendation, comparison, use case, value and competitor-heavy discovery moments. Together, they reflect the questions that shape shortlists and product consideration.

1. Best-in-category recommendation

Business question: Does the brand appear when shoppers ask for the best product in a category?

Test element	Demonstration record
Baseline prompt	“What are the best washing-up liquids for a family kitchen in the UK? Compare cleaning performance, skin friendliness and value.”
Optimisation applied	Reworked the product category page to state the product’s use cases, evidence-backed benefits, pack size, price context and suitability for family kitchens. Added a concise comparison table and clearer links to product detail pages.
Post-optimisation prompt	“What are the best washing-up liquids for a family kitchen in the UK? Compare cleaning performance, skin friendliness and value.”
Baseline answer snippet	“Common recommendations include several established household brands. Look for products described as effective on grease and gentle on hands.” No relevant citation to the tested brand.
Post-optimisation answer snippet	“For a family kitchen, [Brand] is one option to consider for grease removal and everyday value. Its product information highlights family use, pack size and skin-friendliness evidence.” Brand product page cited.
Illustrative citation result	Baseline: 0/5 runs cited the brand. Post-optimisation: 3/5 runs cited the brand. Citation rate: 0% to 60%, a 60 percentage-point change.

The takeaway is not that a category page guarantees inclusion. It is that an answer environment has more usable evidence when the page makes the shopper’s criteria explicit and easy to verify.

2. Brand comparison under pressure

Business question: Does the brand enter a side-by-side comparison when shoppers name competing products or evaluation criteria?

Test element	Demonstration record
Baseline prompt	“Compare [Brand] with two leading alternatives for concentrated laundry detergent. Which is best for a small household that wants good results with less product?”
Optimisation applied	Published a neutral comparison page covering dosage, concentration, wash-load guidance, pack economics and limitations. Added an accessible table with claims linked to supporting product information.
Post-optimisation prompt	“Compare [Brand] with two leading alternatives for concentrated laundry detergent. Which is best for a small household that wants good results with less product?”
Baseline answer snippet	“[Alternative A] is often associated with concentrated performance, while [Alternative B] is known for broad availability. [Brand] may be suitable, but information is less clear.” No brand citation.
Post-optimisation answer snippet	“[Brand] may suit a small household because its guidance explains dosage per wash and concentrated use. [Alternative A] and [Alternative B] remain comparison options, depending on price and availability.” Comparison page cited.
Illustrative citation result	Baseline: 1/5 runs cited the brand. Post-optimisation: 4/5 runs cited the brand. Citation rate: 20% to 80%, a 60 percentage-point change.

Comparison content performs best when it answers the criteria in the prompt rather than repeating brand positioning. It gives an AI answer a source that can support a direct evaluation instead of relying on generic category language.

3. The “best for” use-case query

Business question: Can the brand be discovered for a specific need state rather than a broad category term?

Test element	Demonstration record
Baseline prompt	“What is the best moisturiser for dry skin in winter when the product needs to be fragrance-free and suitable for daily use?”
Optimisation applied	Added a use-case page that clearly mapped the product to winter dryness, fragrance-free requirements, daily use and usage instructions. Claims were aligned with the product label and supporting evidence.
Post-optimisation prompt	“What is the best moisturiser for dry skin in winter when the product needs to be fragrance-free and suitable for daily use?”
Baseline answer snippet	“Look for a fragrance-free moisturiser with barrier-supporting ingredients and a rich texture.” The answer gave general advice without naming the tested brand.
Post-optimisation answer snippet	“[Brand] is relevant to this use case because its product information states that it is fragrance-free and intended for daily use on dry skin.” Product page cited.
Illustrative citation result	Baseline: 0/5 runs cited the brand. Post-optimisation: 2/5 runs cited the brand. Citation rate: 0% to 40%, a 40 percentage-point change.

Use-case queries are commercially valuable because they often signal a specific buying need. That makes clarity the priority: the page must make the product’s intended use, constraints and evidence easy to connect to the question.

4. Value-focused product discovery

Business question: Does the brand appear when shoppers prioritise affordability and total value?

Test element	Demonstration record
Baseline prompt	“What is the best-value dishwasher tablet for a family of four in the UK? Consider cost per wash, cleaning performance and pack size.”
Optimisation applied	Added cost-per-use guidance, pack-size information, dosage instructions and a regularly maintained value page. Price statements were time-stamped and separated from longer-lasting product claims.
Post-optimisation prompt	“What is the best-value dishwasher tablet for a family of four in the UK? Consider cost per wash, cleaning performance and pack size.”
Baseline answer snippet	“Value depends on pack size, promotions and the number of washes. Compare cost per tablet rather than shelf price.” No relevant brand citation.
Post-optimisation answer snippet	“[Brand] can be considered for value when the larger pack is used at the stated dosage. Its product information provides pack size and usage guidance for comparing cost per wash.” Product page cited.
Illustrative citation result	Baseline: 1/5 runs cited the brand. Post-optimisation: 3/5 runs cited the brand. Citation rate: 20% to 60%, a 40 percentage-point change.

Value-focused discovery requires more than a low-price claim. It requires enough product information for the answer to explain how value was assessed. This is especially important for FMCG teams managing changing pack sizes, promotions and usage instructions.

5. Competitor-heavy answers, rebalanced

Business question: Can the brand become part of an answer when competitors dominate the category narrative?

Test element	Demonstration record
Baseline prompt	“Which household cleaning brands are most recommended for tackling everyday grease, and what makes each one different?”
Optimisation applied	Created a structured category guide covering product formats, grease-related use cases, evidence-backed differences and links to product pages. Added independent, relevant references where available and removed unsupported superlatives.
Post-optimisation prompt	“Which household cleaning brands are most recommended for tackling everyday grease, and what makes each one different?”
Baseline answer snippet	“Frequently mentioned options include [Alternative A], [Alternative B] and [Alternative C]. They are commonly associated with strong cleaning performance.” The tested brand was absent.
Post-optimisation answer snippet	“The category includes [Alternative A], [Alternative B], [Alternative C] and [Brand]. [Brand] is positioned around everyday grease removal, while the alternatives differ by format, surface and fragrance.” Category guide cited.
Illustrative citation result	Baseline: 0/5 runs cited the brand. Post-optimisation: 4/5 runs cited the brand. Citation rate: 0% to 80%, an 80 percentage-point change.

This test is particularly useful for executive reporting because it makes absence visible. The goal is not to suppress competitors or imply universal ranking. It is to ensure that the brand has clear, relevant and supportable information available when the category is being explained.

What changed across all five tests

These illustrative records show how a consistent measurement framework can turn AI visibility work into a credible evidence trail:

• Citation presence improved in four of the five examples, with the largest illustrative shift occurring in the competitor-heavy category answer.
• Use-case and value prompts required the clearest product facts, including suitability, dosage, pack size and cost-per-use context.
• Comparison prompts benefited from structured, neutral content that answered the criteria in the question rather than relying on broad brand language.
• Mentions and citations should remain separate metrics. A brand can appear in an answer without being used as a supporting source.
• Repeated runs are essential. AI answers can vary by platform, model context, date, market and retrieval conditions.

For a live Quadrant test, the chart should be generated from the recorded run log rather than from a single answer. Quadrant’s product materials describe prompt-level monitoring, cited URLs, citation share, competitive presence and content recommendations as connected parts of its visibility workflow. (projectquadrant.com)

What brand teams should take from this

The practical value of an AI visibility platform is not limited to showing where a brand appears today. It lies in the ability to connect a shopper question to an observed gap, an optimisation action and a measurable change in the answer environment.

For retail, FMCG and e-commerce teams, that means prioritising pages that support real buying decisions: category guides, product detail pages, comparisons, use-case content, pack information and value explanations. It also means treating citations as an evidence signal rather than a guaranteed ranking factor.

A credible programme should report:

• the exact prompts tested;
• the platforms and dates used;
• the number of repeated runs;
• the optimisation applied between conditions;
• brand mention rate;
• citation rate;
• cited URLs;
• competitor presence; and
• whether the cited source actually supported the answer.

That evidence is more actionable than a generic visibility score. It helps teams decide which content to improve, which product facts to clarify and which discovery moments deserve investment. It also creates a defensible way to evaluate Quadrant alongside other best AI visibility tools without relying on feature lists alone.

The central principle is simple: monitor the answer, improve the source, rerun the prompt and record what changed. That is how AI discovery optimisation becomes testable work rather than an abstract promise.