Tracking Share of Voice in Generative AI Search

Standard SEO platforms like Semrush and Ahrefs are built to track keyword rankings and backlinks. They tell you where your page appears in a list of ten blue links. They cannot, however, tell you if your content's substance is being used as a source inside a Google SGE or Perplexity answer. Their crawlers analyze SERP features but lack the ability to parse the synthesized text of a generative answer and attribute it back to specific source documents.

Consider a B2B tech company that publishes a detailed guide on cloud cost optimization. An Ahrefs report shows the page ranks #4 for its target keyword. This data is misleading. A user searching that keyword gets an AI-generated summary at the top of the page. That summary might cite a competitor's blog post, not the company's guide. The marketing team has no visibility into this; they only see a high ranking and assume success, while the AI engine is actively promoting a competitor's viewpoint using their target keyword.

This isn't a missing feature; it's a fundamental architectural mismatch. SEO tools are designed to analyze a static web of documents. AI search engines create a dynamic, derivative work in real time by synthesizing information. To track share of voice in this environment, a system must analyze the language model's output, not just the ranked inputs. Existing tools are built on a crawler-centric model that is blind to the content of the AI-generated answer itself.

The consequence is operating without accurate feedback. You cannot know which content formats earn citations, what questions drive traffic through AI answers, or how your messaging compares to competitors at the point of search. Optimizing for blue links is becoming an outdated strategy when the real competition is for citation slots within the answer engine.

We built a four-stage AEO pipeline to solve this problem for our own content. The system's first stage, the Queue Builder, acts as a discovery engine. Python scripts run 24/7, scanning sources like Google PAA, Reddit, and industry forums to find questions that signal a high potential for AI citation. Each discovered question is scored based on search intent and competitive gaps before being added to a processing queue in our Supabase database.

The core of the system is the Generate and Validate stages. A queued item is sent to the Claude API with a low temperature setting (0.3) to generate a fact-focused, citation-ready article. To avoid publishing duplicate content, we use a pgvector index in Supabase to perform a trigram Jaccard similarity check; any draft scoring above 0.72 is rejected. A subsequent 8-point quality gate uses the Gemini Pro API to verify data accuracy and check for filler content before approving a page for publication.

When a page passes validation with a score of 88 or higher, the Publish stage executes. This is an atomic operation that updates the database, invalidates the Vercel ISR cache, and submits the URL to indexing services via IndexNow. The entire process, from a question being pulled from the queue to the new page being live and indexed, completes in under 2 seconds. This architecture allows us to produce 75-200 highly targeted pages per day, each designed specifically to be cited by AI search engines.