AI Analytics
The honest formula for share of voice in AI search, with a worked example across 3 competitor brands and 30 prompts. Why the classic impression-share definition breaks and what replaces it.
Part of the AEO Hub — browse all 28 AEO guides.
I get asked the same question by founders about once a week now: how do I compute share of voice for ChatGPT and Perplexity the way my paid-search consultant computes impression share for Google Ads. The honest answer is that the analogue exists, the formula is one line, and the discipline around the inputs takes more pages to explain than the formula itself. This article is the long version of that conversation.
The reason it deserves more than one page is that the classic share of voice math breaks for AI in three structural ways, and a lot of vendors are reporting a clean-looking number that papers over the breaks. I will show what the classic definition was, why it breaks, what replaces it, and then walk a worked example with three competitor brands and 30 prompts that you can replicate in a spreadsheet. The example is the part most articles in this category skip, which is exactly why so many teams end up with a share of voice number they cannot defend in an executive review.
The frame I keep coming back to: share of voice in AI search is the cleanest single number for competitive position, but only when the inputs are honest. Honest inputs are a stable prompt set, an explicit competitor set, a defined window, a deterministic parser, and a re-baseline discipline tied to model updates. Skip any of those and you are reporting a vibe, not a metric.
| Spec | Value | Source |
|---|---|---|
| Year classic share-of-voice formula stabilized | 1960s-70s | Marketing literature, Ehrenberg-Bass [1] |
| Year AI share-of-voice formula started forming | 2024 | Profound, Otterly, Peec product convergence |
| Minimum prompt set for reliable SOV | ~30 prompts | Sampling math + operator practice |
| Minimum runs per prompt for reliable SOV | 5-10 | Engine non-determinism dampening |
| Recommended weekly cadence | Weekly | Engines update on weeks, not hours |
| Princeton GEO lift from citations/stats | Up to 40% | Aggarwal et al. [2] |
| US English AIO appearance rate (Q1 2026) | 13-15% | Search Engine Land [3] |
| Median GA4 Direct that is actually AI-referred | 34% | Attrifast cohort [4] |
| Per-engine RPV cohort ranking | Perplexity > Claude > ChatGPT > Gemini > AIO | Attrifast benchmark [4] |
| Loamly audited brands scoring under 20 visibility | 85.7% | Loamly research [5] |
| Recommended SOV reporting view | Per-engine primary + revenue-weighted blend | Attrifast practice |
| Re-baseline trigger | Model launches + prompt-set changes >10% + quarterly | Attrifast practice |
Share of voice originated in advertising literature as the share of total category advertising spend or impressions a brand owned in a window. The Ehrenberg-Bass tradition extended it into a predictor of market share growth (excess share of voice over market share predicts future market-share gain). In paid search, Google operationalized it as impression share: your eligible auctions where the ad appeared / total eligible auctions, computed by Google's auction log.
| Channel | Numerator | Denominator | Data source |
|---|---|---|---|
| Classic advertising | Your spend or GRPs | Total category spend or GRPs | Industry trackers (Nielsen, Kantar) |
| Paid search | Your impressions | Total eligible auctions | Google Ads platform |
| SEO (informal) | Your top-10 keywords | Total category keywords in top-10 | Ahrefs, Semrush share-of-search |
| Social listening | Your brand mentions | Total brand mentions in category | Brandwatch, Sprout Social |
| AI search | Your citations or mentions | Total citations or mentions in category | Prompt-tracker sampling |
The first four rows share a common property: there is a well-defined platform-side log of the events being counted, or at least a reasonably stable corpus you can crawl. The fifth row does not. AI search has no auction log, no impression event, no fixed inventory of "queries the engines could have shown you on." Every answer is generated per-query and the citation list is part of the generation, not a retrieval against a stable index of slots.
That structural difference is why classic impression-share math breaks. There is no stable total to put in the denominator. The reconstruction is sampling-based: define a prompt set (your treated-as-known queries), define a competitor set (your treated-as-relevant sources), and compute citation share within those bounds. The result is real and useful, but it is a sampled estimate, not a platform log.
The cleaner way to see the break is to enumerate the assumptions that hold for paid search and fail for AI:
| Assumption | Paid search impression share | AI share of voice |
|---|---|---|
| Stable inventory of events | Yes (eligible auctions) | No (per-query generation) |
| Platform-side event log | Yes (Google Ads) | No (prompt tracker reconstructs) |
| Deterministic outcomes | Mostly yes given bid | No (model sampling adds variance) |
| Unambiguous competitor set | Yes (advertisers in auction) | No (you choose) |
| Discrete event units | Yes (impression) | Mixed (citation, mention, both) |
| Constant denominator | Yes (within budget) | No (changes with prompt set) |
Five of six assumptions fail. The metric still translates conceptually, but the math has to be redesigned around sampling. That redesign is the rest of the article.
The cleanest one-line formula I use is:
Share of Voice (you) = sum of your citations across the prompt set in the window / sum of all citations across the same prompt set in the window, where the competitor set is a named list of brand domains.
That formula has six implicit choices that have to be made explicit before the number means anything. The choices are: prompt set, window, engine set, competitor set, citation-vs-mention rule, and weighting scheme.
| Choice | Standard option | Notes |
|---|---|---|
| Prompt set | 30-200 buyer-intent prompts | Must be stable across the window |
| Window | 4-12 weeks | Long enough to denoise, short enough to react |
| Engine set | ChatGPT + Perplexity (minimum) | Add Claude, Gemini, AIO based on audience |
| Competitor set | 3-7 named brand domains | Wikipedia, Reddit usually excluded |
| Citation vs mention | Report both separately | Combining produces misleading clean number |
| Weighting (blended) | Revenue-weighted by engine | Equal-weighted over-rewards low-pay engines |
Make those six choices once, write them down, and your share of voice number is reproducible. Skip any of them and you are reporting a number that cannot be defended in review.
The formula for the per-engine view is the same with the prompt set fixed to that engine's responses only. The formula for the revenue-weighted blend is:
Blended SOV = sum over engines of (per-engine SOV × per-engine revenue share of AI traffic to you).
Revenue share here is your business's revenue from each engine as a fraction of total AI-attributed revenue. The Attrifast 200-site cohort blended weights (Perplexity, Claude, ChatGPT, Gemini, AIO) sit roughly at 18% / 12% / 53% / 9% / 8%, but every business has its own weights and using the cohort numbers blindly is wrong.
The example: a B2B SaaS category with three named competitor brands (call them Brand A, Brand B, Brand C). The prompt set is 30 buyer-intent prompts split as 10 TOFU ("what is X"), 15 MOFU ("X vs Y," "best X for Y use case"), and 5 BOFU ("X pricing," "is X worth it"). Two engines tracked: ChatGPT and Perplexity. Window: 4 weeks. Each prompt run 7 times per week per engine. Competitor set explicitly bounded to A, B, C.
Total observations: 30 prompts × 7 runs × 4 weeks × 2 engines = 1,680 prompt-runs. Within each prompt-run, citations are parsed and matched against the three competitor domains. A single prompt-run might produce zero citations for any of A/B/C, or it might produce citations for one, two, or all three.
| Engine | Brand A citations | Brand B citations | Brand C citations | Total in-set citations |
|---|---|---|---|---|
| ChatGPT | 188 | 124 | 96 | 408 |
| Perplexity | 243 | 218 | 138 | 599 |
| Combined | 431 | 342 | 234 | 1,007 |
| Engine | Brand A SOV | Brand B SOV | Brand C SOV |
|---|---|---|---|
| ChatGPT | 46.1% | 30.4% | 23.5% |
| Perplexity | 40.6% | 36.4% | 23.0% |
| Equal-weighted blend | 43.3% | 33.4% | 23.3% |
Reading the per-engine view: Brand A leads on ChatGPT by a wider margin than on Perplexity, where Brand B is closer behind. Brand C is roughly even-weak across engines. The variance between engines is meaningful — Brand A's lead on ChatGPT (46.1% vs 30.4%) is structurally different from its lead on Perplexity (40.6% vs 36.4%). An equal-weighted blend hides that.
Assume revenue weights for this category are 35% ChatGPT, 45% Perplexity, with the remaining 20% from Claude/Gemini/AIO (not tracked in the example). Re-normalize: ChatGPT 44%, Perplexity 56% of the two-engine universe.
| Brand | ChatGPT SOV | Perplexity SOV | Revenue-weighted blend |
|---|---|---|---|
| Brand A | 46.1% | 40.6% | 43.0% |
| Brand B | 30.4% | 36.4% | 33.8% |
| Brand C | 23.5% | 23.0% | 23.2% |
The revenue-weighted blend favors Perplexity in this example because the category sends more revenue through Perplexity than ChatGPT. Brand B closes part of its gap to Brand A under revenue weighting because Brand B is stronger on the higher-paying engine. The strategic read: Brand B should keep doubling down on Perplexity and try to close the ChatGPT gap; Brand A should defend ChatGPT while investing in Perplexity.
Suppose you widen the competitor set to include Wikipedia, Reddit, and one large reference site that frequently appears in the citations. The total citation pool grows by another 1,150 citations across the window (these sources are cited heavily for definitional and "what do people think" prompts).
| Brand | SOV (3-brand set) | SOV (3-brand + Wikipedia/Reddit/reference) |
|---|---|---|
| Brand A | 43.3% | 20.0% |
| Brand B | 33.4% | 15.8% |
| Brand C | 23.3% | 10.9% |
| Wikipedia | — | 24.0% |
| — | 18.0% | |
| Reference site | — | 11.2% |
That second column is a different metric. It measures total citation surface area, not competitive position within the named-brand competitor set. Both are valid, both should be tracked, but the headline "share of voice" number for an executive should be the bounded competitor-set version, with the total-citation-diversity version reported alongside.
The most common reporting error in this category is using the widened denominator to report a low share of voice number and then panicking about it. Brand A's 20% in the widened view is not worse than its 43% in the bounded view — they are different metrics measuring different things.
Citation share counts cases where your domain appears in the cited sources of an answer. Mention share counts cases where your brand name appears in the answer text whether or not it is linked. These are two different shares and should be reported separately.
The reason matters because engines behave differently on the two. Claude often mentions brands without citing them (Claude conversation mode without web search rarely returns citations at all; it just narrates with brand names). ChatGPT often cites without mentioning the brand name in the prose ("source: example.com" without writing "according to example.com"). A consolidated "visibility share" that adds the two produces a clean-looking number that obscures the difference.
| Engine | Citation behavior | Mention behavior | Reporting recommendation |
|---|---|---|---|
| ChatGPT | High citation density, named sources | Variable mention density | Report both, focus on citation |
| Perplexity | High citation density, always links | High mention density | Report both, citation primary |
| Claude (chat mode) | Low to zero citations | High mention density | Mention is the headline metric |
| Claude (with web search) | Moderate citations | Moderate mentions | Report both |
| Gemini | Moderate citations | Moderate mentions | Report both |
| Google AI Overviews | Variable citations | Lower mentions | Citation primary |
A complete dashboard reports both columns. Treating them as the same metric is the most common mistake I see in vendor-supplied dashboards.
The temptation to roll five engines into a single "AI share of voice" number is strong because it produces a clean executive-ready chart. The temptation should be resisted unless the blend is weighted correctly.
The reason: engines differ in how much economic value they send. Claude conversation mode mentions your brand frequently but sends almost no clicks because there is no link to click. Perplexity links out aggressively and sends real referral sessions that convert. An equal-weighted blend treats those two as equivalent, which is structurally wrong. A revenue-weighted blend weights each engine by how much revenue it sends your business and produces a number that is at least directionally tied to dollars.
| Blending method | When to use | Risk |
|---|---|---|
| Per-engine only (no blend) | When you have meaningful per-engine business decisions | Cluttered dashboard, no headline number |
| Equal-weighted blend | Rarely | Over-rewards engines that do not pay |
| Session-share weighted | When revenue data is unavailable | Better than equal, still over-weights low-RPV engines |
| Revenue-share weighted | When you have first-party Stripe join | Best, requires attribution infrastructure |
| Citation-volume weighted | Within an engine across prompts | Useful for prompt-prioritization |
The revenue-weighted version is the one I run for the founders I help, but it requires a server-side attribution layer underneath that most teams do not have yet. The minimum honest version, when revenue data is not yet available, is to report per-engine SOV as the primary view and clearly label any blend as "session-weighted" or "equal-weighted" so the reader can interpret it correctly.
Five mistakes account for most of the bad share-of-voice numbers I have seen reported in 2025-2026. None of them is rocket science; all of them are easy to make.
| Pitfall | What goes wrong | How to fix |
|---|---|---|
| Drifting prompt set | Trend lines reset whenever a prompt is added or removed | Lock the prompt set at the start of the window |
| Ambiguous competitor set | Number swings 10-20 points on definitional change | Write the competitor set down explicitly |
| Single-run snapshots | Engine non-determinism dominates the result | Sample at least 5-7 runs per prompt per window |
| Combining citations and mentions | Produces clean number that hides engine differences | Report both separately |
| Equal-weighted engine blend | Over-rewards low-pay engines | Use revenue or session weighting |
A sixth and rarer mistake is reporting share of voice without re-baselining after a major model launch. GPT-5 and Claude 4.x both shifted citation patterns meaningfully across the cohort properties I tracked when they launched. A share of voice trend line that runs through a model launch without a baseline event is mixing two different metrics under one chart title.
The report format I run weekly for the founders I help has four sections.
| Section | Content |
|---|---|
| Prompt set state | Prompt count, engine set, competitor set, window, runs-per-prompt — printed every report |
| Per-engine SOV | Bar chart per engine for each competitor in the bounded set |
| Revenue-weighted blend | Single number per competitor, with the weights printed |
| Sensitivity + footnotes | Wider-set comparison + any baseline events in the window |
That format takes 15 minutes to read and produces a defensible scorecard for an executive review. The pieces that produce the data are a prompt tracker for the upstream half (Profound, Peec, SEOcrawl, Otterly, or Loamly) and a server-side attribution layer for the revenue weights (Attrifast or equivalent). The methodology is what stitches them together.
The two articles that complete the trilogy are what is prompt tracking for the data-collection half and AI visibility metrics and KPIs for the broader KPI context this metric sits inside. For the per-engine breakdown of ChatGPT, Perplexity, Claude, and Gemini traffic that feeds the revenue weights, the per-engine landing pages walk the attribution side.
Share of voice in AI search is your brand's citations as a share of all citations across a defined prompt set, over a defined window, on a defined engine set. It is the AI-search analogue to the classic share of voice metric, but it cannot be computed the classic way. In paid search, share of voice was impression share. In AI search there is no auction and no impression in the classic sense. The substitute is citation share: how often you appear among the cited sources when a defined prompt is asked. Across the 40 properties I have instrumented, share of voice is the single number that best captures competitive position, but only if the underlying prompt set is honest.
Three structural reasons. There is no equivalent of total impressions in AI search because every answer is generated per-query, not retrieved against fixed inventory. Brand mentions in AI answers can be unlinked, so your share is not just citation-share but mention-share too. AI engines are non-deterministic — the same prompt produces different citation sets across runs. The classic impression-share formula assumes a stable denominator and a deterministic event log, both absent. The replacement is sampling-based citation share with explicit mention treatment and competitor-set bounds.
Share of Voice (your brand) = sum of your citations across the prompt set in the window / sum of all citations across the same prompt set in the window, where citations is defined as your domain appearing in the cited-sources footnotes of an engine's answer. For mention-share, replace citations with brand-name occurrences in the answer text. For multi-engine, weight each engine by its session-share or revenue-share to your business. The base formula is simple. The hard part is the denominator.
Explicitly, in writing, and with discipline. Define your competitor set as the 3-7 named brands a real buyer would consider, and compute share of voice within that bounded set. Wikipedia, Reddit, and large reference sites typically should not be in the set even if they show up in citations, because they do not compete for your customer. Track total citation diversity separately. The risk of ambiguity is that small definitional changes swing your share by 10-20 points without anything underlying changing.
At least 30 distinct prompts covering the funnel, with each prompt run at least 5-10 times in the window. So a minimum of 150-300 prompt-run observations per engine before SOV numbers are trustworthy. Below 30 prompts the sampling error makes single-percentage-point shifts meaningless. Below 5 runs per prompt the variance from engine non-determinism dominates. The 50-200 prompts run weekly with 7 samples each is the sweet spot — large enough to denoise, small enough to manage.
Both, but report per-engine as primary and blended as a weighted average. Per-engine tells you where you are strong and losing; blended tells you total competitive position. Blend should be weighted by something economically meaningful — session share by engine, or per-engine RPV if you have a Stripe join. Equal-weighted blends treat Claude (low click-through) the same as Perplexity (high click-through), which over-rewards engines that do not pay.
Impression share is your ad shown / total eligible auctions, computed by Google against its own auction log. AI share of voice has no auction, no eligibility flag, and no platform-side log. It must be reconstructed by sampling. Impression share is deterministic given a bid; AI SOV is probabilistic given a prompt and engine state. The concept transfers; the data source does not. Anyone reporting AI SOV as if it were a platform-supplied number is overstating precision. Honest reports include prompt count, runs per prompt, and competitor set.
They are two different shares and should be reported separately. Citation share counts cases where your domain appears in cited sources. Mention share counts cases where your brand name appears in the answer text whether or not linked. Across engines, the two diverge: Claude often mentions without citing, ChatGPT often cites without naming the brand in prose. A consolidated visibility share that adds them is misleadingly clean. Track both. Citation share is closer to the click funnel; mention share is closer to brand association.
Most engines do not rank citations within an answer beyond display order, and display order does not always correspond to weight. The cleanest practice is to treat all sources equally — each gets one count regardless of position. For a position-weighted variant, weight by inverse rank (1.0, 0.5, 0.33) but report it as a separate position-weighted share, not the headline. Engines sometimes return zero citations and only a mention; those count toward mention share but not citation share.
After every major model launch (GPT-5, Claude 4.x, Gemini 3.x reshuffled citation patterns), after any prompt-set change above 10% addition or removal, and quarterly even without an obvious trigger. Engines push retrieval updates without notice; a January baseline will not be comparable to a May baseline. The discipline is to fork the dataset at each baseline event and report "share of voice (Q2 2026 baseline)" rather than running one time series through reshuffles.
A simplified but realistic three-brand category, 30 buyer-intent prompts, weekly runs across ChatGPT and Perplexity for four weeks, with citation parsing and the formula applied row by row. Brand names are illustrative but the cohort-level distribution mirrors what I see across the 200-site Attrifast benchmark for B2B SaaS with three to five visible competitors. The example produces per-engine and blended SOV, with a sensitivity analysis showing how the number shifts when the competitor set widens.
Attrifast does not run prompt tracking, so it does not measure citation share or share of voice directly. It handles the downstream half: server-side AI traffic attribution and Stripe-joined revenue per engine. The proper stack is a prompt tracker (Profound, Peec, SEOcrawl, Otterly, or Loamly) for the SOV numerator and denominator plus Attrifast for the revenue-weighting on the blended-across-engines version. Joining the two layers lets you compute revenue-weighted SOV, the most decision-useful version of the metric.
This article is the methodology walkthrough — the formula, the worked example, the standing report. For the conceptual layer covering why classic SOV is a vanity number and how Revenue Share of Voice fixes it (the metric that adds Stripe-joined revenue weighting to the mention count), see AI Share of Voice in 2026: How to Measure It, and Why Revenue Share of Voice Matters More.
For the data-collection layer this methodology sits on top of, see what is prompt tracking. For the broader KPI set, see AI visibility metrics and KPIs. For the related comparison with rank tracking, see prompt tracking vs keyword rank tracking. For the revenue layer that closes the SOV-to-dollars loop, see revenue attribution.
Attrifast joins your AI sessions to Stripe payments so your SOV can be weighted by dollars, not just impressions.
Start free trial →5-day free trial · $29/mo · cancel anytime
Discover which marketing channels bring customers so you can grow your business, fast.
Start free trial →5-day free trial · $29/mo · cancel anytime
