What columns should an AI share-of-voice tracker include?

A defensible AI SoV tracker needs nine columns: Prompt text, Engine name, Run date, Brand mentioned (Y/N), Position in response (first named / secondary / not mentioned), Sentiment (positive / neutral / negative), Competitor mentions (comma-separated brand names), Cited URL (the source link the engine attributed), and Notes for hallucinations or unusual response behavior. These nine columns capture the data you need to compute SoV, audit sentiment, and spot citation gaps without bloating the sheet.

What is the share-of-voice formula for AI search?

AI SoV = (your brand mention count) / (total mentions across your defined competitor set, including your own brand) x 100. For example, if you ran 100 prompts and your brand appeared in 28, Competitor A in 41, and Competitor B in 19, your SoV is 28 / (28 + 41 + 19) = 31.1%. The denominator must always include your own mentions, not just rivals.

How many runs per prompt do I need to get a reliable reading?

Run each prompt a minimum of five times per engine before drawing a conclusion. Answer engines are probabilistic: a single response is one sample from a distribution. Five runs give you a stable citation rate for that prompt. Use 10 runs for high-entropy prompts (open-ended, opinion-heavy, or low-consensus topics) and re-run on a weekly or fortnightly cadence to track drift.

Should I track the same prompts every week?

Yes, with periodic pruning. Consistent prompts let you spot drift over time. Retire a prompt when it no longer reflects how real buyers phrase the question, when your SoV has saturated near 100% on that prompt for several weeks, or when it duplicates another prompt in your set. Invest those run credits in prompts where the answer is still contested.

Which AI engines should I include in the tracker?

Start with ChatGPT, Perplexity, Google AI Overviews, Gemini, and Microsoft Copilot. Those five cover the majority of AI-assisted research and purchase journeys. Add Google AI Mode and Meta AI once the core five are running smoothly. Track each engine as a separate row for the same prompt so you can see per-engine SoV and detect the divergence that makes multi-engine tracking necessary.

Can I automate this tracker instead of doing it manually?

Manual tracking works for 20 to 50 prompts across two or three engines. Beyond that, dedicated platforms handle prompt execution, run averaging, and SoV computation automatically. Peec AI, Otterly.AI, Profound, and Temso all run prompts on a schedule, export raw response data, and surface SoV dashboards. The spreadsheet template is the right starting point for understanding the methodology; automation is the right next step once your prompt set grows.

The Share-of-Voice Tracker Template: A Spreadsheet to Measure Your AI Brand Mentions vs Competitors Across 5 Engines

Last updated July 2026

Knowing whether ChatGPT mentions your brand is a start. Knowing your brand appears in 28% of category prompts while your top competitor appears in 41% is a number you can act on. That gap is what an AI share-of-voice tracker makes visible.

The problem is that most teams either skip the tracker entirely or build one that collapses under scrutiny. The column choices look fine until someone asks why the numbers bounce week to week. The answer is almost always the same: a single run per prompt, no averaging, and a formula that counts raw mentions rather than a defined competitor set.

This post gives you the full column schema, the correct SoV formula, a worked example with three competitors, and the run-averaging method that makes the numbers defensible.

The 9-column schema

Copy this structure into any spreadsheet. Each row is one prompt run on one engine on one date.

#	Column name	What to record	Example value
1	Prompt	The exact question text, verbatim	”What is the best AI visibility tool for B2B SaaS?“
2	Engine	The AI platform queried	ChatGPT
3	Run date	ISO date (YYYY-MM-DD)	2026-07-16
4	Brand mentioned (Y/N)	Was your brand named in the response?	Y
5	Position	Where your brand appeared: First named / Secondary / Not mentioned	Secondary
6	Sentiment	How your brand was described: Positive / Neutral / Negative	Positive
7	Competitor mentions	Comma-separated list of rival brands named in the same response	Profound, Otterly.AI
8	Cited URL	The source URL the engine linked to or cited, if any	https://yoursite.com/category-page
9	Notes	Flag hallucinations, wrong pricing, outdated claims, or unusually long or short responses	Hallucination: incorrect pricing stated

A few field notes:

Position matters because “first named” brands carry disproportionate influence in reader recall. Track it separately from the Y/N flag.
Competitor mentions should use the exact brand name as it appears in the response, not your internal shorthand. This makes COUNTIF formulas work correctly.
Cited URL is often blank for conversational engines and populated for retrieval-augmented ones (Perplexity, Google AI Overviews). Record it when it appears.
Notes is the column analysts skip and then regret. A note flagging “ChatGPT stated our pricing as $99/mo; correct price is $29/mo” is exactly the kind of accuracy gap that needs tracking and fixing.

The SoV formula

AI share of voice = (your brand mention count / total mention count across the competitor set) x 100

The denominator includes your brand plus every competitor in your defined set. It does not use the total number of prompts run. It counts how often any brand in your competitive universe was mentioned, then measures your slice of that total.

SoV = (YourMentions / (YourMentions + CompetitorA + CompetitorB + ... + CompetitorN)) × 100

This is the formula used by dedicated platforms like Peec AI, Profound, and Otterly.AI. When you build your own tracker, using the same formula keeps your numbers comparable to what those platforms report.

Worked example: 3-competitor set

You run 100 prompts across five engines (20 per engine). After averaging five runs per prompt (so 500 raw runs total per engine, but 100 averaged results), you tally:

Brand	Mentions across 100 averaged prompts	SoV
Your brand	28	31.1%
Competitor A	41	45.6%
Competitor B	19	21.1%
Competitor C	2	2.2%
Total	90	100%

Note that 10 of the 100 prompts produced no brand mention at all. Those 10 prompts are not in the denominator. Your SoV is 28 / 90, not 28 / 100.

This matters. If a category prompt like “how do I improve my search visibility?” returns a generic answer with no brand names, including it in the denominator would deflate everyone’s SoV artificially. Exclude no-mention prompts from the SoV calculation but keep them in the sheet. A rising share of “no-mention” prompts for your category is its own signal.

The run-averaging method

Answer engines are probabilistic. Ask ChatGPT the same question five times and you may get five different brand mentions. A single run is a data point. Five runs are a measurement.

The practical workflow:

Run each prompt five times in a single session (same day, same engine, same account state).
Record each run as a separate row in the sheet.
Compute the citation rate for that prompt: mentions across five runs divided by five. A brand mentioned in three of five runs has a 60% citation rate for that prompt.
Use the per-prompt citation rates, not the raw counts, when you compute weekly SoV.
Track the citation rate over time. A prompt moving from 40% to 70% over four weeks is the kind of drift that warrants investigation: either the model’s training updated, a new piece of content earned a citation, or a competitor lost ground.

The methodology note that makes your numbers defensible: state the run count and the date range every time you report SoV. “Our AI SoV for the category prompt set was 31.1% (5-run average per prompt, 100 prompts, week of 14 July 2026, ChatGPT + Perplexity + Google AI Overviews + Gemini + Microsoft Copilot)” is a number a stakeholder can interrogate. “Our AI SoV is 31%” is not.

How to structure your prompt set

The tracker schema handles the recording. The prompt set determines what you measure.

Aim for 30 to 100 prompts per measurement cycle. Fewer than 30 gives you a noisy SoV figure that swings too hard on individual prompt changes. More than 100 is hard to maintain manually (and the right answer at that scale is automation).

Organize prompts into intent clusters:

Category evaluation (“what is the best tool for X”): captures top-of-funnel discovery.
Feature comparison (“which tool does X better”): captures mid-funnel consideration.
Pricing or objection (“is X worth the cost”): captures bottom-of-funnel decision.
Problem-first (“how do I fix X”): captures prompts that do not name your category but where you want to be mentioned.

Track each cluster’s SoV separately. A brand that dominates category-evaluation prompts but disappears from problem-first prompts has a specific gap to close: content that answers the problem in a way that leads to a category recommendation.

Per-engine SoV: why it matters

Run each prompt against each engine you track. Compute SoV by engine, not just in aggregate.

Research from Profound (July 2025) found that only about 11% of domains are cited by both ChatGPT and Perplexity, based on analysis of 100,000 prompts run across both platforms. The engines draw from largely different source pools. Your aggregate SoV can look healthy while you are invisible on one engine that your buyers use most.

The per-engine column in your tracker makes this split visible. If your SoV on ChatGPT is 40% but on Google AI Overviews it is 8%, those are two different problems requiring two different fixes.

Tracker maintenance: a short checklist

Consistent tracking is more valuable than a perfect tracker you abandon. Keep it simple:

Run prompts on a fixed day each week (or fortnight). Tuesday morning works well because it avoids weekend model-update noise.
Archive old runs, do not delete them. Historical data is what lets you draw a trendline.
Retire prompts that no longer reflect how buyers phrase the question. Add replacements when you spot new phrasing in sales calls or support tickets.
Flag any week where you changed the prompt wording. Even minor edits (adding a question mark, changing tense) can shift citation rates and break your trendline.
Review the Notes column monthly. Patterns in hallucinations (wrong pricing, wrong feature set, wrong company size) indicate source material that needs updating.

When to use a dedicated tool instead

The spreadsheet template is the right place to start. It forces you to understand the methodology before you hand it to software.

Once your prompt set grows beyond 50 to 100 prompts, or you track more than three engines, manual tracking becomes a bottleneck. That is when purpose-built platforms earn their keep.

Peec AI runs prompts on a daily schedule across 9+ engines including DeepSeek and Grok, exports raw response data, and surfaces per-engine SoV without manual COUNTIF formulas. Agencies use it because it carries unlimited user seats on every plan.

Otterly.AI offers structured GEO audit guidance alongside prompt-level citation tracking across six platforms. Its Standard plan starts at $189/mo and includes competitive SoV benchmarking, which the Lite tier does not.

Profound is the specialist pick for teams that need the deepest citation intelligence. Its 680-million-citation dataset and Prompt Volumes feature (which surfaces which questions real users are asking AI engines) go well beyond what a spreadsheet can replicate. Entry for a full-featured programme is $399/mo.

Temso is the all-in-one option for teams that want monitoring and an action plan in one product. It tracks share of voice, brand mentions, citations, and sentiment across eight AI engines from $89/mo, and converts visibility gaps into a prioritised fix queue inside the same subscription.

Semrush added AI Overviews tracking to its platform in 2025. It is a reasonable starting point for teams already inside the Semrush ecosystem, though its AI tracking is narrower than dedicated AEO platforms.

The full ranked list of AI visibility tools covers all major options with pricing and engine coverage side by side.

The methodology note

Every SoV number you share should carry three pieces of metadata: the run count per prompt, the prompt set size, and the engine list. Without those, stakeholders cannot judge whether the number is a snapshot from a handful of runs on one engine or a stable average from a rigorous measurement programme.

The formula and schema in this post match the methodology used by the major AI visibility platforms. When you graduate from this spreadsheet to a dedicated tool, your historical data will still be interpretable and your SoV figures will be comparable across both approaches.

For the full measurement methodology used to score tools on this site, see /methodology. For term definitions, see /glossary.

Start with the nine columns, run five times per prompt, apply the formula, and add the methodology note to every report. That is the complete foundation for a defensible AI share-of-voice programme.

If you want the measurement handled automatically across eight engines from day one, Temso’s $89/mo plan runs the prompts, averages the runs, and surfaces the SoV dashboard without a spreadsheet in sight. Peec AI and Otterly.AI are strong alternatives if agency multi-seat access or structured GEO audit guidance matters more than all-in-one execution.