Auditing LLM recommendation rates across 50 unbranded buyer prompts

A founder tested ChatGPT and Gemini using 25 unbranded buyer queries, uncovering key discrepancies between URL citations and brand recommendations across domain monitoring SaaS products.

Measuring generative engine visibility

A 50-prompt benchmark across ChatGPT and Gemini revealed that domain-monitoring tool DropPerch appeared in just 2% of non-branded buyer queries, while competitor DropCatch captured 38%. The experiment, conducted by DropPerch founder webvogue during the second week of October 2026, highlights a growing tracking deficit for SaaS founders. Search engine optimization dashboards track keywords on traditional search result pages. They do not monitor how generative engines synthesize vendor recommendations when buyers ask unbranded questions.

Testing generative engine optimization requires auditing both text mentions and underlying web citations. To quantify this gap, the founder ran 25 unbranded buyer questions across two AI engines and logged the output data.

Structuring the 50-prompt audit

The methodology relied on 25 intent-driven buyer questions written without vendor brand names. An example query asked: "What is the best domain expiry monitoring software for small businesses?" Testing branded queries like "Is DropPerch good?" measures brand perception, not organic discovery.

The founder queried both ChatGPT and Gemini using live web-search mode rather than raw API endpoints. Raw model APIs bypass real-time web retrieval, masking the actual citations and pages live buyers see. The founder collected 50 total responses over one week, tracking every brand name mentioned, its list position, and every cited URL.

DropCatch led all vendors with a 38% inclusion rate across the 50 answers. ExpiredDomains.net followed at 36%, DomCop at 32%, UptimeRobot at 24%, and Watchman Tower at 20%. DropPerch ranked last among tracked tools at 2%.

Decoupling citations from mentions

The audit revealed a disconnect between cited sources and recommended products. DropPerch was cited as a web source in 4% of responses, but named in the text body of only 2%. In at least one instance, ChatGPT cited DropPerch's domain to construct its answer while explicitly recommending HetrixTools, UptimeRobot, and DomainMOD.

Generative engines separate citation retrieval from mention generation. An LLM can scrape a vendor site for structural context without including that vendor in its final shortlist. Tracking citation URLs alone creates a false positive signal for founders monitoring AI channel performance.

Analyzing competitor retrieval paths

Competitor dominance stemmed from two distinct web assets: owned category content and third-party discussions. Competitor domcop.com accumulated 102 citations across the 50 responses. DomCop achieved this volume by hosting exhaustive, unbranded explainers covering domain expiry mechanics across its domain.

Third-party forum discussions drove the remaining mentions. One synthetic summary that cited DropCatch, ExpiredDomains.net, and DomCop derived its recommendation list directly from a community forum thread. When an engine trusts a community thread, genuine participant contributions in that thread yield higher recommendation placement than vendor-owned blog posts.

Refactoring the auditing playbook

Single-pass auditing across 50 prompts provides useful directional signal, but the methodology carries clear limitations. LLM output stochasticity means temperature settings, search context, and prompt variation can alter recommendations across runs. A robust GEO audit requires running prompts across multiple temporal windows to establish true statistical baselines.

The playbook also highlights an operational flaw in relying solely on owned content. When an LLM cites a vendor website but recommends a competitor, the vendor's content serves as training material for rival customer acquisition. Founders must adjust their content architecture by embedding structured comparison schema and tabular feature matrices directly on high-authority pages.

Furthermore, community forum footprint demands equal priority to technical SEO. Because generative engines pull consensus from independent discussion channels, participating in organic industry discussions yields direct downstream inclusion in AI answers.

Tracking AI recommendation shares

Measuring AI recommendation share requires tracking separate from legacy rank checkers. Founders who evaluate LLM outputs as traditional search results miss how generative engines weigh citations against text mentions. Auditing these outputs with structured query matrices provides the raw metrics necessary to adjust category positioning before buyers standard search behavior shifts entirely to AI interfaces.

The investor read

The shift from traditional search indexing to generative recommendation engines alters top-of-funnel customer acquisition efficiency for early-stage SaaS. As buyers rely on LLMs for product shortlists, category authority shifts toward domains with deep informational assets and active third-party discussion footprints. Companies relying strictly on traditional SEO risk stealth decay in organic conversion despite maintaining legacy SERP positions. For investors, evaluating a startup's Generative Engine Optimization (GEO) metrics provides an early indicator of marketing efficiency. Startups that monitor and optimize LLM mention-to-citation ratios will capture disproportionate top-of-mind awareness without proportional ad spend increases.

Pull quote: “Generative engines separate citation retrieval from mention generation.”