##  [Measuring AI Share of Voice for API Products: Lessons from AEO/GEO Practitioners](/articles/measuring-ai-share-voice-api-products-lessons-aeogeo-practitioners) 

By

 [Laura Vass](/users/laura.vass) 

 

 

Co-Founder. Editor, Research & Knowledge Sharing. 

Sep 08, 2026

 

AI share of voice and prompt monitoring make it possible to observe how API capabilities appear in AI-generated answers. This article looks at how practitioners establish baselines, choose and refine prompts, interpret share of voice and investigate visibility gaps. It also considers what changes when these methods are applied to API products, where the evidence can inform API documentation, competitive positioning and portfolio decisions beyond AI visibility itself.

This article draws on two panel conversations at Promptwatch’s Visible Amsterdam event: one with practitioners working inside brands and another with agencies doing AEO/GEO work for clients. Their subject was brand visibility rather than individual products, so PR, media coverage, social channels, third-party sources and consistency of brand messaging had an important place in the discussion.  
  
Pronovix comes to AI visibility through technical products. Its [Competitive API Capability Benchmark ](https://pronovix.com/#competitive-api-capability-benchmark)compares how strategically important API capabilities appear in AI-generated answers, how they connect to business solutions, and how they perform against relevant competitors.  
  
The use case discussed at the event was different, but much of the instrumentation for measuring AI share of voice was familiar: selecting and refining prompts, establishing a visibility baseline, examining sources and citations, comparing competitors, and observing what changes after an intervention. What is less transferable is deciding what to investigate in the first place. For specialized B2B integration products, that decision requires product- and market knowledge that sits outside the monitoring instrumentation itself.

## Establish an AI visibility baseline before optimizing

  
One of the more consistent themes across the two panels was establishing the current state before moving into AEO/GEO optimization. From the brand side came recommendations to understand the status quo first, get the prompt quality right, and do the less exciting work of fixing inconsistent data.  
  
Agency practitioners described a similar starting point in somewhat different terms. One [put the technical foundation and information architecture first](https://pronovix.com/blog/components-developer-portal), followed by a baseline measurement. Another described how large companies often do not have a good grip on their existing content, and that AI amplifies the inconsistencies and gaps already there.  
  
For an API program, those inconsistencies may already be familiar as problems of governance, portfolio coherence or [developer portal quality](https://pronovix.com/ai-visibility-assessment-banking-developer-portals). AI-generated answers provide another external view of how coherently the program’s API capabilities are represented.  
  
Establishing a baseline sounds simple, but it first requires deciding what deserves to be included in it. Pronovix does not start API prompt selection at the keyboard: the first step is to work with stakeholders to establish which API capabilities matter to the business and the market around them. The questions to monitor follow from that selection.

## Choosing the prompts you monitor

  
For AI visibility monitoring, the prompt set determines what will actually be observed. On the brand panel, practitioners described starting with questions known to matter, including both branded and non-branded prompts, and continuing to refine the test set rather than treating prompt quality as settled.  
  
A prompt set can easily look rigorous while measuring questions that are not particularly important. For a small and well-bounded product, constructing the test set may still be manageable. In a larger API landscape, the problem is not simply that there are more possible prompts, but that more decisions have to be made about which capabilities, business problems and competitors deserve to be represented.  
  
The most detailed example from the panels came from the agency side. One team starts with personas, core products and the problems those products solve, then follows the buyer journey from discovery through decision-making and constructs prompts around it. They also cross-check the prompt setup against user interviews. When AI visibility breaks at a particular stage of the journey, they use that as a reason to revisit the personas or the prompts they are monitoring.  
  
The prompt set is not something to establish once and leave untouched.Practitioners described refining it as they learned which questions produced useful signals, including where competitors appeared but their own brand did not.

## Reading AI share of voice beyond the score

Once the prompt set exists, AI visibility is the obvious place to start. Practitioners described following groups of prompts by topic, comparing where their brand and competitors appeared, and tracking measures such as citation share. On the agency side, citations and sentiment were also being monitored alongside traffic and other downstream signals.  
  
AI share of voice makes repeated observations across the selected prompt set comparable, including how the company appears relative to relevant competitors. But the panel conversations also kept moving beyond the score itself, into the underlying answers and sources: where and why competitors appeared and the brand did not, which sources were being cited, and where incorrect information might be coming from.  
  
This makes the underlying answers worth keeping alongside the metrics. A change in share of voice tells you that something moved; looking at the prompts, answers and sources helps investigate what moved and where to look next.  
  
There was more uncertainty when the discussion reached business impact. Agencies described looking at incoming traffic, post-purchase direct traffic and other signals, and the panel included reported examples of individual experiments that produced measurable changes. At the same time, one practitioner described the ROI of AEO/GEO as an unsolved problem.  
  
For API capability benchmarking, some of the value is less dependent on establishing a direct ROI from GEO. **Gaps found in the public API information may also affect developer discovery and onboarding independently of AI, while comparing API capabilities with relevant competitors adds external evidence to API strategy and positioning decisions.**

A prompt panel repeated over time can also show changes in which competitors appear around the capabilities being monitored, including entrants that were not part of the original comparison.  
  
There is a longer-term question around agentic discovery and implementation: whether the same gaps that make an API capability difficult to represent accurately in AI-generated answers also make it harder for AI systems to identify and work with the API.

## What do you change once you find a gap?

Finding a visibility gap does not yet tell you what caused it. In the brand panel, practitioners described looking at the sources behind AI-generated answers: which outlets were being cited, where competitors appeared and they did not, and where their own brand appeared without competitors. Old content and inconsistent information across different touchpoints also mattered.

Sometimes the intervention was outside the company’s own website. One practitioner described tracing incorrect information in an AI answer back to its external source and working with the publisher to correct it. Other brand-side interventions included PR and high-authority publications.

The agency panel brought up crawlability and information architecture, central content governance, authority signals, and problems further upstream in content creation. Their observation that AI can amplify existing content inconsistencies and gaps is useful here: [an AI visibility problem may lead back to something further upstream in the documentation](https://pronovix.com/articles/ai-success-begins-strong-developer-portal-content-strategy), structured information, product pages or other published sources, rather than to the measurement itself.

Practitioners also described small experiments as useful. A measurable change can make it easier to bring the work into the wider organization, while the discussion remained much more cautious about attributing overall business results to GEO.  
  
For an API capability, AI visibility measurement can help narrow the investigation: which capability, which question, which competitor, and which part of the published information is associated with the gap. [These gaps are not all of the same kind.](https://pronovix.com/b2b-capability-discovery-consulting) A capability can be documented but difficult to connect to the business problem it solves. What we most often see across the corporate and portal sites is a business solution that is clearly presented, while the API capabilities that support it are gated for readers in the public technical documentation. A competitor that is more consistently and accessibly represented for a capability has an advantage in AI-mediated discovery over another company with a relevant offering of its own.The same visibility measurement can surface these quite different problems.

Interpreting these gaps requires a reliable reference on both sides: what the business says the solution offers, and what the technical product actually supports and documents. Establishing whether that correspondence is sufficiently clear is part of the Pronovix investigation. From there, the appropriate change depends on what the evidence shows.

## Start with a focused AI visibility benchmark

The panels did not offer a finished AEO/GEO playbook, nor did they pretend to. Practitioners were still refining prompt sets, testing interventions and working out what their measurements could reliably tell them. Even with more visibility data available, the connection to business outcomes remained unresolved.  
  
There is nevertheless enough here for a bounded experiment: a subject with business relevance, a baseline, a considered prompt set, the underlying answers and sources alongside the aggregate measures, and some way of observing what happens after a change is made. For a small and well-defined problem, much of this is accessible with the prompt-monitoring tools now available.  
  
With a larger API landscape, more of the work lies in deciding what deserves to be investigated and interpreting what the resulting evidence means. **The Pronovix Competitive API Capability Benchmark keeps the scope deliberately focused, comparing a selected set of strategically important API capabilities with relevant competitors.** The analysis looks at how those capabilities appear in AI-generated answers, how the business solutions presented by the company connect to what can be found through its public API documentation, and what AI systems can access and interpret in the developer portal.

For API teams, this provides external evidence for conversations that otherwise tend to rely heavily on internal assessments of documentation, discoverability and portfolio priorities.

## About the practitioner discussions

This article draws on notes taken during two panel discussions at Visible Amsterdam, an AI Search event organized by Promptwatch in Amsterdam on September 3, 2026.  
  
*How Leading Brands Are Getting Ahead in AI Search* brought together Emma Gammons, Director of Global External Communications at Shutterstock; Dave Vollebregt, Marketing Manager at Treehouse; Marjolein Jongbloed, Agentic Commerce Lead at DataLab; Koen Gijsman, Digital Analyst at Zilveren Kruis; and Nathan Necciai, Growth Marketing Lead at Promptwatch.  
  
*How Agencies Are Helping Clients Win in AI Search* brought together Ilonka Karoly, Media Performance Director, Total Search at WPP Media; Iain Davenport, Director of Business Solutions and SEO at Monks; Hans van Gent, Head of SEO/GEO at Seeders; Tobias Peschke, Founder & CEO at Loud & Lexis; and Gijs De Groot, CEO & co-founder of Promptwatch.  
  
The article groups and paraphrases observations from the discussions rather than presenting a transcript.