How to track share-of-voice across AI engines

The four signals, explained

RankRush tracks four distinct signals, reported separately for each engine because each assistant has its own citation behavior. Aggregating them into one number hides where the real gap is.

• Citations — how often an assistant names your domain by name in an answer for the prompts that matter to your category.
• Mentions — how often you're referenced by brand name even when the assistant doesn't link to you directly.
• Share-of-voice — your slice of the citation set for a given prompt versus the named competitors in your category, tracked per assistant.
• Competitor delta — which competitors are picking up citations you aren't, and on which prompts, so you can see exactly where the gap is.

Step 1 — Pick the prompts that matter to your category

Share-of-voice is only meaningful against a fixed set of prompts. Choose the high-intent questions a buyer in your category would actually ask an assistant — comparison questions, "best tool for X" questions, and problem-framed questions. Keep the set stable; this is the cohort you'll re-scan against every time, so a changing prompt list makes the delta uninterpretable.

Step 2 — Run a baseline scan across all four engines

ChatGPT, Claude, Perplexity, and Gemini are the four engines most B2B audiences use today, and they behave differently — a brand can be visible on one and invisible on the others. Run your baseline across all four so you know where you stand on each, not just on average. The baseline is the zero point every later re-scan is measured against.

Step 3 — Read the four signals separately, per engine

Look at each signal per engine rather than as a blended score. You might have strong mentions on Perplexity but zero citations on Claude — those are different problems with different fixes. Pay special attention to competitor delta: the prompts where a competitor is cited and you are not are your highest-leverage targets, because absence on an AI answer isn't neutral — the buyer walks away with the competitor's recommendation and there's no second result to click.

Step 4 — Ship a fix, then re-scan on the same prompts

Use your audit's prioritized fix list to ship one structural change — see how to fix a failing audit node. Then re-scan against the same prompt cohort. Holding the prompts constant is what makes the comparison valid; if you change the prompts and the fix at the same time, you can't tell which moved the number.

Give the engines time. Models refresh their indexes on their own cadence. Wait roughly 10–14 days after shipping a fix before reading too much into a re-scan, and expect each engine to move on its own schedule.

Step 5 — Attribute the delta and set a cadence

Because the prompt set is fixed and only one variable changed, the difference between scans is attributable to the fix you shipped — not guessed at. Set a recurring scan (weekly while you're actively shipping, monthly once your visibility is stable) and read the diff each cycle. That recurring, prompt-stable measurement is the entire point: improvements become traceable to specific actions.

Quotas are per plan. AI-visibility scans, website audits, and node re-checks are metered separately on every tier. If you're iterating fast, lean on the cheaper node re-checks for single-fix verification and reserve full visibility scans for periodic share-of-voice snapshots.

Next steps

Most share-of-voice gaps trace back to how your pages are structured for extraction. If you're not being cited where competitors are, start with how to structure a page so AI engines can cite it. To establish your baseline now, run a free RankRush audit.