How to measure AI share of voice: the practitioner's method.

AI share of voice (AI SoV) is the percentage of AI-generated answers that name your brand across a defined set of buying-intent prompts. It is the primary metric for measuring your brand’s presence inside ChatGPT, Perplexity, Claude, and Gemini — and it is not tracked by Google Search Console, GA4, or any rank-tracker you are currently running.

The gap matters. You can sit at position 3 for your main keyword and still be invisible in AI answers, because AI answers are not built from SERP rankings. They are built from a thin layer of trusted sources: a handful of comparison sites, a few independent blogs, and your own domain when it answers the question directly. The list of sources an LLM retrieves for “best accounting software for small businesses” is not the same list that determines your Google ranking.

I measured this directly. In June 2026 I ran a share-of-voice study for the Finnish ecommerce platform market — 6 models, 6 buying-intent prompts, 36 answers. The total API cost was $2.56. This article is the methodology behind that study, made replicable for any market.

What AI share of voice measures — and what it doesn’t

AI SoV measures one thing: the percentage of AI-generated answers, across a defined prompt set, that include your brand by name.

If you run 36 answers and your brand appears in 28 of them, your AI SoV is 78%. If a competitor appears in 35 of those same 36 answers, their SoV is 97%. The gap between those numbers is your target — and unlike a Google ranking gap, it has a knowable cause: a small set of intermediary sources the model retrieves that don’t mention you yet.

What AI SoV does not measure: click-through rate, revenue influence, or sentiment. Those are downstream questions. SoV is the top-of-funnel signal — are you in the shortlist the buyer sees before they consider clicking anything? Generative engine optimization is the practice of getting into that shortlist; AI share of voice is how you measure whether the practice is working.

The metric becomes legible to a CFO for the same reason share of search does: “we appear in 78% of AI answers for our category” is a business position, not an SEO metric. It competes on the same language as market share and share of wallet. Teams that are still measuring this as an SEO metric miss the stakeholder conversation where it matters most.

The 5 metrics that matter

A single SoV number misses the structure of the data. Run a full prompt study and you have five metrics worth tracking separately.

1. Share of voice. Percentage of answers naming your brand across the full prompt set. The headline number — useful for benchmarking against competitors run through the same prompts in the same session.

2. First-mention rate. In the answers that name you, how often does your brand appear first? First position in an AI answer is the strongest proxy for the model’s top recommendation — the equivalent of the featured snippet position in classic SEO. A brand with 78% SoV but 5% first-mention rate is being mentioned as an afterthought, not a recommendation.

3. Citation share. What percentage of the source URLs cited across all answers point to your domain? You can appear in an answer without being cited as a source. Citation share measures whether the model treats your domain as an authority — a source it retrieved — or just a subject it knows about from training data. These are meaningfully different positions.

4. Prompt coverage. Across your defined prompt set, what percentage of prompts trigger any citation of your domain at all? Coverage below 50% on your highest-intent queries is the clearest signal that the model is answering your buyers without you in the answer. A single prompt with 0% coverage is often more actionable than a weak average SoV.

5. Model-tier split. The same brand can have dramatically different SoV across model tiers. In the Finnish ecommerce study, Finqu had a SoV of 67% on Claude and Perplexity — and 0% on OpenAI’s two models. Not one mention across 12 answers. A buyer using ChatGPT never saw Finqu. A buyer using Perplexity saw it in most answers. If your buyers cluster around one assistant, that model’s number is the only one that matters for prioritization.

How to design the prompt set

The prompt set is the study design. Get it wrong and the data doesn’t answer the question you’re actually asking.

Six to ten prompts covers most categories. The minimum viable set has five types:

Generic best-of. “What is the best [category] for a small business in 2026?” This establishes the baseline SoV on the most common query shape — the question buyers ask before they know what they want.

Budget variant. “What is the most affordable [category] for a startup?” Budget models truncate the candidate list to well-known names. You need to know if you’re on the short list when the buyer is price-sensitive.

Niche variant. A specific use case: “What [category] is best for a freelancer?” or “What [category] is best for an e-commerce business?” Niche queries surface different brands than generic ones. If your product has a strong niche fit, this is the prompt that should show your highest SoV.

Head-to-head. “Company A vs Company B vs Company C — which should I choose?” Use your two main competitors as the reference points. This is where the model reveals its ranked preference most clearly. It’s also where you’ll catch the model actively deprioritizing you relative to named rivals.

B2B or enterprise variant. Enterprise queries often return a different set than SMB queries. If your buyers are procurement teams, this prompt frequently matters more than the generic one — and the model often draws from different source material.

Run all prompts in the same language your buyers use. In the Finnish ecommerce study, all six prompts were in Finnish — because that is what Finnish founders type. Language choice changes which retrieval set the model pulls from. Mixing English and Finnish prompts would measure two different populations.

How to run it: the DataForSEO API

The DataForSEO AI Optimization API gives access to live LLM responses with web search enabled across OpenAI, Anthropic, Google, and Perplexity models. It is the cheapest way to run this study at scale without a SaaS subscription.

The endpoint is /v3/ai_apis/openai/{llm_type}/llm_responses/live. You pass a prompt, a model name, and "web_search": true — except for Perplexity, which is web-native and rejects that field. Each call returns the full answer text plus annotated citation URLs.

For 36 answers (6 models × 6 prompts), the model breakdown in June 2026 was: gpt-4.1 at roughly $0.13/call, claude-opus-4-5 at $0.21, gemini-2.0-flash and sonar-pro at $0.02–$0.04. Total: $2.56. The budget models (gpt-4.1-mini, claude-haiku-4-5) cost under $0.02 per call each.

Two practical notes. First: include at least one budget-tier model per provider. The budget tiers are what most users are actually on — and in the Finnish study, they behaved measurably differently from flagship models, both in which brands they named and in how many candidates they listed. Second: store each response as a raw JSON file and skip if the file already exists. This makes the script idempotent — you can re-run it after an interruption without re-spending money on calls that already completed.

The raw response for each call contains the answer text and, for web-search-enabled models, an array of annotated citation objects with URLs. Both are worth storing.

How to parse the output: two outputs from one run

Each response gives you two types of extractable data. Treat them as separate datasets.

Brand mentions. Search each answer’s text for each brand name using case-insensitive regex with word-boundary matching. Record whether each brand appears per answer (binary — once per answer for SoV counting). This gives you the share-of-voice table: how many of the 36 answers named each brand.

Cited URLs. Extract the domain from each citation URL and count by domain across all 36 answers. Normalize for www. prefixes. This gives you the citation frequency table: how often each external source appeared.

The SoV table answers: who appears in AI answers?

The citation table answers: who does AI trust as a source?

The citation table is the output most teams ignore. In the Finnish ecommerce study, one comparison site, tyokaluvertailu.fi, was cited 59 times across 36 answers. A solo blog, sampsavainio.fi, was cited 20 times — more than most platforms’ own domains combined. A payments provider, Paytrail, appeared 9 times for questions that were never about payments, because its platform compatibility pages list every ecommerce tool and the model treats those lists as ground truth.

The citation table is a target list. In the Finnish ecommerce category, six sources drove the majority of brand visibility. Being present and accurately described on those six sources has more impact on AI SoV than any on-page optimization you can do on your own domain. This is what tracking AI citations looks like in practice — the intermediary layer is thin, knowable, and actionable.

How to act on the data

The study produces three outputs with immediate tactical implications.

Your SoV number per model. This tells you where you stand in the specific models your buyers use. 0% on gpt-4.1 requires a different fix than 0% on sonar-pro, because the retrieval sets differ. If your buyers skew toward ChatGPT, OpenAI model numbers are the ones that drive decisions. The aggregate average hides this.

The model-tier split. If you appear on flagship models and not budget models, your problem is likely content recency: budget models fall back on training data more readily than models with aggressive live-web retrieval. Getting cited by recently-indexed comparison sites closes this gap faster than anything else. If you appear on budget models and not flagships, the problem is more likely entity recognition — the flagship model doesn’t have a confident mental model of what you do.

The citation domain ranking. This is the most actionable output. List the top 10 cited domains. For each one, check whether your brand appears and is correctly described. A domain cited 40+ times that doesn’t mention you is your first task. A domain that mentions you incorrectly — wrong pricing tier, wrong category, outdated features — is your second.

The fix is not on-page SEO in the classic sense. You don’t need more backlinks to your own domain. You need a placement on the intermediary sites the model already retrieves. This is PR logic applied to AI visibility, and it is the practical implementation of entity-based SEO: make the model confident about who you are by being present where it looks.

Sequence the work this way. Start with the citation domains that are already cited heavily and don’t mention you — those are the fastest wins. Then work on citation domains that mention you inaccurately. Last, target new domains that aren’t in the current top-10 but rank well for your buying-intent queries — getting them to cite you expands the model’s retrieval surface for your brand.

When to use a SaaS tool instead

The DataForSEO approach makes sense for audits, campaign snapshots, and non-standard markets. Three situations where a dedicated SaaS tool is the better choice.

100+ prompt sets. Manual JSON parsing and tracking across hundreds of answers requires a structured dashboard. Profound, SE Ranking’s AI visibility tracker, and LLMrefs all provide this. The engineering overhead of a custom pipeline above roughly 50 prompts per week stops being worth it.

Continuous competitive benchmarking. If you need to track 5 competitors weekly and monitor for answer changes over time, a SaaS subscription costs less than the engineering work. Profound and Ahrefs Brand Radar run continuously and surface changes between runs — the model named a new competitor, a previously-cited source dropped out, your SoV shifted 10 points. That surveillance function is hard to replicate manually.

Google AI Overview coverage. The DataForSEO API does not return Google AI Overview data. AI Overviews are a separate retrieval surface — different from ChatGPT, different from Perplexity — and they reach a different audience (passive Google search users, not active AI assistant users). SE Ranking and Profound both cover AI Overviews alongside ChatGPT and Perplexity. If your buyers are primarily Google users, the API approach alone misses that surface.

The Ahrefs AI visibility checker is a free entry point — it runs your domain against a small standard query set and returns a citation score. It’s useful for orientation but doesn’t provide model-tier split or the citation domain breakdown you need to act. Neil Patel’s Ubersuggest offers a similar free tool. Both are useful for a sanity check before deciding whether to run a custom study. Neither replaces the custom run for your specific category and buying-intent prompts.

How to maintain the measurement over time

Re-run monthly. Model versions ship on a rolling basis. Retrieval sets change as new content gets indexed. Comparison sites update their rankings. The Finnish ecommerce study would look different today than it did six weeks ago — that’s not a flaw, it’s the nature of live-web retrieval. A snapshot that’s 90 days old is historical data, not a current measurement.

Keep the prompt set identical between runs. Prompt phrasing changes the answer, and you need a stable baseline to measure change. If you want to test a new prompt, add it alongside the existing ones rather than replacing them. Label it in your analysis so you don’t compare its first run against older runs.

The signals worth watching across runs: first-mention rate shifts (did you move from mentioned-somewhere to mentioned-first?), model-tier convergence (are budget models starting to include you?), and new citation domains entering the top 10. These mark the difference between “this is working” and “this is drift.”

The entire infrastructure is a Python script, a DataForSEO account, and a spreadsheet. The monthly cost is the same $2–5 as the initial run. The reason most teams have no idea what AI recommends about them is not cost or complexity — it’s that they are still looking at the wrong dashboard.

If you want to see what this looks like across real software categories, I have run the same methodology against four markets: Finnish ecommerce platforms, CRM software, accounting software, and project management software. The patterns — one tool dominating first-picks, SoV and recommendation position diverging sharply, a handful of intermediary sources driving almost all AI retrieval — repeat across every category.