Case study: how a B2B SaaS tripled ChatGPT citations in 90 days

Meta description: Learn how a B2B SaaS improved AI visibility in 90 days by tracking citations, sentiment, and share of voice across major LLMs.

Slug: case-study-tripled-chatgpt-citations-90-days

Can a B2B SaaS really triple ChatGPT citations in 90 days? This example is a structured case study template, not a published, independently audited result. To make a “tripled citations” claim defensible, you need (1) a repeatable scan methodology, (2) a verifiable dataset of prompt runs, and (3) a results table you can share internally or publicly.

Below is a practical, step-by-step blueprint you can use to run that kind of 90-day experiment with LLM Monitor as the tracking layer.

> How to make this real (recommended): If you publish the outcome, include a link to your case study page (or attach an appendix) with: the prompt library, scan cadence, scoring rules, and a results table by model.

---

What the 90-day result actually means

In this case study format, “tripled citations” means: for a fixed set of prompts, the brand’s citation frequency at day 90 is ~3× the baseline at day 0.

Two important clarifications:

1. It’s not “every model did the same thing.” You should report lifts by engine (ChatGPT, Gemini, Claude, Perplexity, and Google AI Overviews) because citation behavior differs. 2. A single spike is not enough. The strongest evidence is a repeated lift across multiple prompts and multiple intent categories (discovery, comparison, decision).

What you should report (so it’s traceable)

For each model, publish a small table like this (fill with your numbers):

Where the lift typically shows up first

For B2B SaaS, the earliest improvements usually appear in:

• Comparison prompts (models are more likely to cite a vendor when asked to weigh options)
• Decision prompts (models often cite “best fit” tools when the prompt includes constraints like team size, compliance, or workflow)

---

Why AI citations matter before the click

AI engines increasingly act like a front door: they summarize, recommend, and sometimes include citations or source links directly in the answer.

For a B2B SaaS, that changes the funnel:

• Discovery becomes “who gets named” inside the answer. - Consideration becomes “who gets compared fairly.”
• Shortlisting becomes “who appears as a viable option under constraints.”

If your brand is missing or consistently described negatively, you don’t just lose traffic, you lose inclusion.

The practical business implication

Even if you don’t track downstream conversions in the first week, you should track leading indicators that correlate with inclusion:

• citation frequency
• mention rate
• position within ranked lists
• sentiment of the mention
• share of voice versus named competitors

---

The measurement setup for a 90-day test

A credible test is built on fixed inputs and explicit scoring rules. Don’t rely on “eyeballing” answers.

Step 1: Define the prompt library (fixed, versioned)

• Create 15, 30 prompts.
• Tag each prompt with intent: discovery, comparison, decision.
• Version prompts (e.g., `v1`, `v2`) and freeze the version during the 90-day window.

Concrete prompt examples (use your own, but keep structure stable):

• Discovery: “What are the best tools for monitoring LLM outputs for a B2B product team?”
• Comparison: “Compare Brand A vs Brand B for enterprise compliance and audit trails.”
• Decision: “Which vendor is best for a mid-market SaaS team that needs SOC 2 evidence and citation tracking?”

Step 2: Run the same scan cadence

A common cadence:

• Baseline: day 0 (and optionally day 7 to confirm stability)
• Midpoint: day 30
• End: day 90

If you can afford it, run weekly. If not, keep it consistent and report the cadence.

Step 3: Compute citation frequency (explicitly)

Because “citation” can mean different things across engines, you must define it.

A traceable definition you can use:

• Extract all brand mentions and source/citation references from the model output. - Count a citation event when the output includes your brand name and a citation/source indicator (if the engine provides one). - If the engine doesn’t provide explicit citations, you can still score mention events separately, but don’t label them “citations.”

Citation frequency can then be computed as:

• `citation_frequency = (total citation events across all prompts) / (number of prompts)`

Step 4: Compute mention rate

• `mention_rate = (number of prompts where brand appears at least once) / (total prompts)`

Step 5: Position scoring (how to avoid ambiguity)

Position depends on output format. Use a consistent rule, for example:

• If the model provides a ranked list, record the rank where your brand first appears. - If it provides a paragraph recommendation, record a position bucket (e.g., early/middle/late) based on sentence order.

Step 6: Sentiment scoring (define the rubric)

Sentiment should be scored using a rubric you can reproduce.

Example rubric:

• Positive: explicitly recommends, praises, or highlights benefits (e.g., “best,” “strong,” “reliable,” “helps with compliance”). - Neutral: factual mention without evaluative language. - Negative: warns, criticizes, or flags limitations (e.g., “doesn’t,” “limited,” “risk,” “hard to”).

If you use an LLM to classify sentiment, document it and keep the classifier stable.

Step 7: Share of Voice (SoV) definition

Define SoV as a ratio against competitors:

• Choose a competitor set (e.g., 3, 8 named rivals that appear in your prompt answers). - `share_of_voice = your_citation_events / (your_citation_events + competitor_citation_events)`

> Source/methodology note: If you publish this methodology, cite the exact documentation for your extraction and scoring approach (e.g., your internal scoring spec, or the LLM Monitor documentation page describing brand mention, citation tracking, sentiment. And competitor benchmarking).

Where LLM Monitor fits in

Use LLM Monitor as the tracking layer so your baseline and weekly comparisons don’t drift.

In practice, you’ll typically:

• configure the brand + competitor entities you want tracked
• set up prompt runs across engines
• review brand mention + citation + sentiment + SoV outputs in one place

LLM Monitor is especially useful when you need the same scoring rules week over week, so you can focus on prompt and source improvements rather than manual spreadsheets.

---

The prompt library that makes the test defensible

A prompt library is not just “a list of questions.” It’s a measurement instrument.

Build prompts to isolate variables

Instead of asking only broad questions, design prompts that test specific inclusion drivers:

• Evidence prompts: “What proof points support claims about auditability?”
• Workflow prompts: “How does the tool integrate into an existing review process?”
• Constraint prompts: “For a SOC 2, focused team, what features matter most?”

These prompts tend to produce outputs where citation/mention behavior is more consistent.

Keep wording stable, but add controlled variants

To reduce template repetition without breaking comparability:

• keep the core structure constant
• vary only one element per prompt set (e.g., compliance requirement vs integration requirement)

Example controlled variants:

• Same intent, different constraint: “SOC 2 evidence” vs “GDPR retention”

How to handle citation parsing across engines

Different engines format citations differently. To keep scoring consistent:

• normalize brand names (e.g., “LLM Monitor” vs “LLMmonitor”)
• strip punctuation and unify casing before matching
• define what counts as a “brand mention” (exact match vs fuzzy match)

If you’re using LLM Monitor, configure entity matching rules so the tracking layer applies the same normalization each run.

---

The 90-day workflow that can lift citations

The goal is to improve the probability that models will (a) recognize your brand and (b) cite or recommend it when prompted.

This is usually not one “magic” change. It’s a sequence of improvements that make your content easier to retrieve, easier to compare, and easier to trust.

Week 1, 2: Establish baseline and confirm stability

1. Run the prompt library across all target engines. 2. Record baseline metrics: citation frequency, mention rate, position, sentiment, and SoV. 3. Run a second baseline scan (e.g., day 7) to confirm the baseline isn’t an anomaly.

Concrete checks (do these before changing anything):

• Are your brand mentions being detected consistently (no missing due to naming variants)?
• Are competitors being tracked with the same entity rules?
• Are any prompts producing “no answer” or refusal responses that skew results?

Week 3, 6: Improve source material (what to change, specifically)

Instead of “update pages,” make targeted edits that map to how models justify recommendations.

A practical checklist for B2B SaaS pages that models reuse:

1. Comparison-ready sections

• Add a “Why we’re different” block with 3, 6 crisp bullets. - Include a short “best for” section that matches your decision prompts.

2. Evidence and auditability language

• Add explicit statements that correspond to compliance prompts (e.g., retention, audit logs, access controls). - Where possible, link to verifiable artifacts (e.g., security documentation pages).

3. FAQ that mirrors prompt phrasing

• Create FAQs that use the same language buyers use (e.g., “citation tracking,” “brand mentions,” “sentiment analysis,” “competitor benchmarking”). - Avoid generic FAQs; answer with specifics.

4. Integration and workflow clarity

• Add “how it works” diagrams or step-by-step descriptions. - Include integration pages that match common tool ecosystems.

5. Reduce ambiguity in naming

• Ensure your brand name appears consistently on key pages. - Add a short “About” snippet that includes the exact brand name and product name.

How to connect content changes to metrics:

• After each content batch, rescan a subset of prompts that directly map to the edited pages (e.g., comparison prompts that mention compliance).
• Use LLM Monitor to verify whether the lift appears in the intended prompt category, not just overall.

Week 7, 10: Rescan and verify the lift is real

Rescan the same prompt library (same versions). Then:

• compare baseline vs day 90 by model
• compare by intent (discovery vs comparison vs decision)
• check whether sentiment improved or merely shifted mention volume

Concrete “lift validation” rules:

• Require improvement in at least two intent categories (e.g., comparison + decision).
• Require improvement in at least two engines (to avoid a single-engine artifact).
• Confirm that position improved (not just more mentions).

Week 11, 13: Lock in the gains with prompt and content refinements

If the lift is partial, don’t restart from scratch. Iterate:

• If citation frequency improved but sentiment didn’t: adjust content to address recurring negative themes. - If mention rate improved but position didn’t: strengthen “best for” and comparison blocks so your brand appears earlier in ranked outputs.

Use LLM Monitor again to keep the tracking consistent as you iterate.

---

FAQ: common questions about measuring AI citations

How to publish results without overclaiming

If you want to claim “tripled citations,” include:

• the prompt library (or an appendix)
• scan cadence and date range
• the exact definitions for citation frequency, mention rate, position, sentiment, and SoV
• a results table by model and intent
• any limitations (e.g., engine formatting differences, citation availability)

Close with a sharp, non-summary statement:

The fastest way to improve AI visibility is to treat it like measurement-driven product work, then prove it with repeatable scans, not one-off screenshots.