How to Measure AI Share of Voice

AI share of voice (SOV) is your brand's share of all brand mentions across a fixed set of prompts and engines, compared with a defined set of competitors. To measure it, choose three to six competitors, run 30-100 buyer prompts several times in each engine, count mentions (optionally weighted for recommendations and position), and divide your total by the total for all tracked brands.

Formula (unweighted): SOV = your mentions ÷ mentions of all tracked brands × 100.

On this page

• Step 1: Who belongs in the competitor set?
• Step 2: How should the prompt set be structured?
• Step 3: How do you run the scans?
• Step 4: How do you count and weight mentions?
• Step 5: How do you calculate SOV by engine and topic?
• From MentionX data: how much prompt type moves SOV
• Step 6: How do you act on the result?
• Worked example: Ledgerline's first SOV report
• How do you match brand names reliably?
• How do you know if a change is real?
• Why do some engines give fewer brands to share?
• How should SOV be reported?
• What are the common mistakes?
• How does SOV relate to other metrics?
• How do you choose weights you can defend?
• Common questions
• Sources

Engine documentation checked on 30 October 2026. "Our view" marks MentionX opinion. Ledgerline is fictional.

Step 1: Who belongs in the competitor set?

SOV only means something relative to a fixed set. Pick:

• 2-3 direct competitors you lose deals to (ask sales).
• 1-2 category leaders buyers expect to see.
• Optionally, 1 adjacent alternative (spreadsheets, an agency, a DIY route) if AI answers often suggest it.

Write down the exact brand names and common variants ("HubSpot", "HubSpot CRM"). Keep the set fixed for at least a quarter; changing it changes the denominator.

Step 2: How should the prompt set be structured?

Group prompts by topic so you can see where you're strong and weak:

Avoid brand-name prompts ("is YourBrand good?") in the SOV set; they inflate your share. Track them separately.

Step 3: How do you run the scans?

• Engines: at least ChatGPT, Perplexity, Google AI Overviews or AI Mode, and Gemini.
• Repeats: run each prompt 2-3 times per engine; answers vary.
• Conditions: fresh sessions, same location, same model or mode, logged with the date.
• Capture: every brand mentioned, in order, and whether each was recommended for the user's need.

With 40 prompts × 4 engines × 3 runs, you'll have 480 answers. That's a lot by hand, which is why most teams automate after the first baseline (manual vs automated).

Step 4: How do you count and weight mentions?

Unweighted counting treats "also consider YourBrand" the same as "YourBrand is the best fit for you". Many teams use simple weights:

Whatever you choose, document it in the report and don't change it mid-quarter.

Worked example (illustrative numbers, not data): across 120 answers, you score 60 points, Competitor A 90, B 30, C 20. Total = 200. Your weighted SOV = 60 ÷ 200 = 30%.

Step 5: How do you calculate SOV by engine and topic?

Pivot your log:

1. Rows: engine (or topic). Columns: brand. Values: sum of points.
2. Divide each brand's points by the row total.
3. Add an "answers with no tracked brand" count; a high number means engines are naming brands outside your set, so review the set.

Illustrative layout only.

From MentionX data: how much prompt type moves SOV

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,154 completed engine runs across 18 anonymised projects and six engines, called through APIs. SOV here is the tracked brand's mentions as a share of mentions of all brands in its project's competitor set, averaged across projects, by prompt intent.

Where the brand appeared in a ranked list (n=102 placements), it was first 18% of the time, second or third 14%, fourth or fifth 13%, and sixth or lower 56%.

What it means. Including brand-specific prompts lifts the all-intent SOV from 21.1 to 30.1 in this data. That gain says nothing about how buyers who don't know you will meet you. Problem prompts, where buyers describe a pain without naming a category, show the lowest share, and they're often the earliest stage of a search.

Limits. 18 projects is a small, non-random sample of B2B software brands that chose to track themselves. Competitor sets were chosen by each project, so SOV levels aren't comparable across projects. Answers come from APIs, not consumer apps. "n/a" marks figures we didn't compute separately. Perplexity, Copilot and DeepSeek figures come from a single day (26 October), so treat them as a snapshot.

Our view: report SOV on unbranded prompts as the headline, and show brand-specific SOV as a separate accuracy check.

Step 6: How do you act on the result?

Look for these patterns:

• Low SOV on one engine only. Check crawler access for that engine's bot and which sources it cites.
• Low SOV on one topic. You probably lack a page, or third-party coverage, for that use case.
• High mentions, low recommendations. Engines know you but don't think you fit; clarify who you're for on key pages and in listings.
• A competitor dominating comparisons. They may own the "X vs Y" pages engines cite. Publish an honest comparison and get onto neutral review sites.

Worked example: Ledgerline's first SOV report

Ledgerline is a fictional accounts-payable tool. Figures are illustrative. Ledgerline tracks four competitors (called A to D here) on 40 unbranded prompts in four engines, three runs each: 480 answers, of which 472 complete.

Using the weights above (mention 1, recommendation +1, first named +1):

Split by topic, Ledgerline reaches 38% on "invoice approval workflows" but 6% on "AP for construction firms", where Competitor A holds 58%. Reading those answers shows engines citing two construction-industry review pages that list Competitor A and not Ledgerline. Ledgerline also has no page about construction.

The team picks one action for the quarter: a construction use-case page with real customer workflow detail, plus an update request to one of the two review sites. They mark the publish date on the trend chart and wait two months before reading the result.

How do you match brand names reliably?

Counting is where most SOV errors creep in. Before you tally:

• Build a variant list for each brand: product names, old names, common misspellings, the parent company name if engines use it.
• Decide what counts. Does "a HubSpot-style CRM" count as a HubSpot mention? Usually not; write the rule down.
• Handle sub-brands. If a competitor has several products, decide whether you count the company or only the product that competes with you.
• Spot-check by hand. If you use a script or model to extract brands, review 10% of answers manually each month and record the error rate.

How do you know if a change is real?

Answers vary, so small moves can be noise. Practical rules:

1. Don't report a change unless it holds across two consecutive months.
2. Show the number of answers behind every percentage; 5 points of SOV on 40 answers is two mentions.
3. When a change is large, open the raw answers and read them. Often one engine changed format, or a new brand appeared and diluted everyone.
4. Mark model launches and major search updates on the trend chart.

Why do some engines give fewer brands to share?

SOV divides a pie, and the pie size differs by engine. In the same MentionX data, unbranded answers that cited sources named 1.11 brands on average, against 1.44 for answers without citations. Engines also differ in how often they mention tracked brands at all: Gemini 47.3% of answers, OpenAI 42.9%, Perplexity 31.5%, Google AI Mode 31.3%, Copilot 26.0% and DeepSeek 24.0%. A 30% share on an engine that names few brands can mean only a handful of mentions, so show the count behind every percentage.

How should SOV be reported?

• One headline number (weighted SOV, all engines), with last month and the quarter's start.
• SOV per engine and per topic.
• Mention rate and recommendation rate alongside, so SOV isn't read alone.
• Method box: prompt count, runs, engines, dates, weights, competitor set.

What are the common mistakes?

• Including branded prompts in the SOV set.
• Changing the competitor list every month.
• Reporting one run per prompt.
• Treating SOV as market share. It is your share of AI mentions in the prompts you chose, measured in the engines you chose, nothing more.
• Leaving out the method box, so no one can reproduce the number.

How does SOV relate to other metrics?

SOV is relative; mention rate is absolute. You can gain SOV because a competitor fell, without improving yourself. Always show both. See AI share of voice and AI visibility score.

How do you choose weights you can defend?

Weights turn SOV into a judgement, so keep them simple and explain them in one line. Three tests help:

1. Does each weight match a real difference in value? A recommendation for the buyer's exact need is worth more than a passing mention. First position probably is too, given how many placements sit far down lists.
2. Is it stable? If changing one weight reorders your competitors, report the unweighted figure as well.
3. Can someone else reproduce it? Write the rules so a colleague scoring the same 50 answers reaches nearly the same total.

Score negative framing separately rather than subtracting points. A negative mention still means the engine knows you, and mixing the two hides both signals.

Common questions

What is a good AI share of voice? There's no universal benchmark. It depends on how many competitors you track and how crowded the category is. Judge it against your own baseline and your closest competitors on the same prompts.

How many prompts do I need? 30-50 prompts run at least twice per engine gives a usable first read. Grow to 100+ when you report by segment or topic, because each topic needs enough answers on its own.

Should SOV include engines my buyers don't use? Weight engines by what you know of your audience, or report them separately. An average across six engines can hide a weak result in the one your buyers rely on.

Should we track more than six competitors? Only if engines regularly name them. Each extra brand shrinks everyone's share and adds matching work, so add brands when the "other brands" column shows they keep appearing.

Why did our SOV rise when nothing changed? Usually a competitor dropped, a new brand outside your set took mentions from everyone, or an engine changed answer format. Check mention rate and the "other brands" column first.

Sources

• Aggarwal et al., GEO: Generative Engine Optimization, which uses visibility metrics in generated answers as its outcome
• Google Search Central, AI features and your website, on why answers and links vary across AI features
• NIST/SEMATECH, Confidence intervals for a proportion, on how sample size limits small changes