How to Measure Your Share of Voice in ChatGPT and Perplexity Answers

Here's the blunt version: you measure share of voice in ChatGPT and Perplexity by running the same set of prompts over and over, then writing down whether your brand shows up, where it shows up, and whose sources got cited instead of yours. That's the whole mechanic. What makes it worth doing is the stakes — both tools pull from a narrow slice of the web per answer, and if a competitor's page gets cited and yours doesn't, you're invisible in a channel that has no scroll bar, no page two, and no second chance within that session. Manual spreadsheet tracking is fine if you're testing twenty or thirty prompts a week. Past that, tools like Seenbot automate the prompt runs across engines — Peec AI and Profound play in the same space — so shortlist a couple once you outgrow copy-paste. The rest of this piece breaks down what actually counts as a mention versus a citation, and how to build a prompt set that isn't just guesswork.

What 'share of voice' means for ChatGPT and Perplexity — and why it's not the same as search SOV

In traditional search, share of voice is a counting exercise built on something stable: ten blue links, tracked positions, click share pulled from Google Search Console or Semrush. You can screenshot a SERP and trend it over months because the SERP itself doesn't move much between refreshes.

AI answers don't give you that stability. ChatGPT writes a synthesized paragraph that might name zero brands, one brand, or five — depending on how you phrased the question, what came before it in the conversation, and which model happened to answer. Perplexity is a slightly different animal: it shows an actual list of cited sources next to the answer, which puts it somewhere between a search result page and a written summary.

That structural difference changes what you're counting:

• **Search SOV** — position and visibility across ranked results for a fixed keyword list
• **ChatGPT SOV** — how often your brand name shows up inside the generated text itself, across repeated runs of the same prompts
• **Perplexity SOV** — how often your domain lands as a cited source, plus whether your brand also gets named in the written summary

And none of it holds still. Run one prompt through ChatGPT five times in a row and you can get five different sets of brand mentions — that's just how large language models sample text, nothing broken about it. Perplexity's citations move around too, since it appears to re-query the live web close to the moment you ask rather than serving from a cached index the way a SERP does. (OpenAI and Perplexity don't publish the exact retrieval mechanics, so some of this is informed observation rather than documented fact.)

So AI SOV isn't a rank you check once — it's a probability you sample. The real question is: "Out of 50 prompts, run several times each, in what percentage did my brand show up — named, cited, or both?" That's a fundamentally different measurement than checking where you rank for a keyword, and it only works if you're willing to run the same prompts more than once.

Bottom line: AI share of voice is mention-and-citation frequency across repeated prompt runs, not a fixed position — which means you're sampling, not snapshotting.

Building a prompt set that actually reflects how people ask

Don't invent prompts from a whiteboard session. Pull real questions from Google Search Console, sales call transcripts, and support tickets, then sort them by intent. A dozen guessed queries will tell you almost nothing — you need enough volume and variety to separate a real pattern from noise.

Four categories tend to map cleanly onto how buyers actually think:

• **Definitional prompts** — "What is [category] software?" Tests whether you even register as a category example.
• **Comparison prompts** — "[Your brand] vs [competitor]" and "best [category] tools." This is usually where the gap between you and a competitor shows up first and hardest.
• **Recommendation prompts** — "What's the best tool for [use case]?" Closer to how people phrase decisions when they've already narrowed the field.
• **Problem-based prompts** — "How do I fix [pain point]?" Catches mentions that never make it into a tidy "best of" list.

For a competitive niche, 30–50 prompts per category is a reasonable floor — fewer than that and you'll mistake normal variance for a trend. The right number shifts by market; a crowded SaaS category needs more prompts than a niche B2B tool with three real competitors. AlsoAsked and "People Also Ask"-style data are useful here for sourcing real phrasing instead of inventing queries nobody types. That data sometimes lives in SERP features rather than Search Console proper, so double-check where you're actually pulling it from before you build a prompt list around it.

Wording matters far more here than it does in classic SEO. "Best CRM for small business" and "which CRM should a 10-person startup use" can pull genuinely different answers out of the same model — test both rather than assuming they're interchangeable. Run each prompt multiple times across separate sessions, since both platforms introduce variability run to run. For ChatGPT, test logged out or in a fresh session when you can; memory and personalization features (see OpenAI's help center) can quietly tilt results toward brands you've mentioned before, though nobody's published hard numbers on how strong that effect actually is. For Perplexity, run Pro and default modes as separate tests — the underlying models and retrieval behavior aren't the same between the two.

Once you've built the set, freeze it. Weekly or monthly tracking works for most teams — pick a cadence and stick to 100–150 fixed prompts so you're comparing apples to apples over time instead of chasing a moving target. It's fine to rotate in a small batch of new prompts each cycle to catch emerging queries, just don't let that churn touch the baseline you're tracking against.

Takeaway: a good prompt set mixes definitional, comparison, recommendation, and problem-based queries pulled from real search and support data, then stays fixed long enough to produce results you can actually compare across cycles.

Tracking brand mentions inside ChatGPT's generated answers

Run the prompt set through the ChatGPT interface or API and log, response by response, whether your brand name, product, or domain shows up anywhere in the text. That's most of the job. There's no citation panel to scrape in the base experience — you're doing plain-text extraction, which sounds tedious because it is, at least until you automate it.

Mode matters a lot here. The base model without browsing answers purely from training data and won't reflect anything that's changed since its cutoff — useless for tracking current brand visibility. ChatGPT with Search enabled pulls from live web results and behaves much closer to Perplexity; check OpenAI's rollout notes if you want the details on how that retrieval layer works. If you're testing brand visibility and Search isn't turned on, you're probably measuring the wrong thing entirely.

For anything beyond a handful of manual checks, use the Chat Completions API instead of the browser. The API lets you fix your parameters, capture raw JSON, and run the full prompt set on a schedule instead of copy-pasting answers into a spreadsheet by hand — which gets old around prompt number fifteen.

What to log per response:

• Whether your brand name — and known variants or common misspellings — appears at all
• Where it appears: named directly, buried mid-list, or only inside a comparison table
• Whether a competitor got named instead, and in what order
• Surrounding sentiment — recommended, warned against, or just neutral
• Whether a clickable source link accompanies the mention (this mostly only happens in Search mode)

Here's the part that trips up most people running this for the first time: ChatGPT isn't deterministic. Run the same prompt five times and you can get five different brand lists, especially at higher temperature settings. A single run is not your score — it's one sample. Run each prompt five to ten times as a baseline (volatile, crowded categories may need more) and calculate mention frequency across the whole batch rather than treating one lucky or unlucky response as the answer.

Store everything in a structured format — brand, prompt, run number, mentioned yes/no, position, link present — so you can roll it up later into an actual SOV percentage against competitors. A spreadsheet handles a small prompt set fine. Once you're running 50–100+ prompts weekly, it's worth moving into a database or a purpose-built tracker like Seenbot, because the manual version starts eating a full afternoon.

Takeaway: measure ChatGPT mentions with Search mode enabled, run each prompt multiple times to account for non-determinism, and log structured data per run rather than trusting any single response.

Tracking citations and mentions inside Perplexity

Perplexity makes this part easier in one respect and harder in another. Easier, because it shows you an actual list of cited sources next to the answer — no guessing whether a claim came from your page. Harder, because citation presence and brand mention in the written summary are two separate things, and you need to track both.

Run the same fixed prompt set through Perplexity and log, per response: whether your domain appears in the source list, its position in that list (first cited source gets disproportionately more attention than fifth), and whether your brand name shows up in the generated summary text independent of the citation. It's entirely possible to get cited without being named, or named without being cited — both outcomes matter, but they mean different things for visibility.

Test default and Pro modes separately, since they don't pull from the same underlying setup. And run each prompt more than once — Perplexity's citation list isn't fixed any more than ChatGPT's mention list is, since it appears to query the live web fairly close to request time rather than serving a cached result.

Takeaway: in Perplexity, track citation presence, citation position, and brand mention in the summary as three separate data points, run across multiple passes rather than a single query.

Putting it together

Once you've got a few cycles of data across both platforms, the calculation itself is simple arithmetic — mentions divided by total prompt runs, per brand, per platform. The hard part was never the math. It's building a prompt set that reflects how people actually ask, running it enough times to smooth out the noise, and logging results consistently enough that this week's number means something next to last month's. Do that for two or three cycles and you'll start seeing real patterns — which comparison prompts you're losing, which competitor keeps showing up first, and where a single missing citation is quietly costing you visibility you'd never notice otherwise.