How to Measure AI Visibility: Metrics That Hold Up

Which AI visibility metrics actually hold up?

Four metrics hold up under repeated testing: visibility rate (share of runs where an engine names you), citation rate (share of those runs where it also links a source), share of voice (your slice of every brand mention in a set of runs), and sentiment read alongside accuracy (how you're described, and whether it's correct). All four are rates across many runs, not a reading from one answer.

Position — where you land in a list — does not hold up. It swings run to run on the same prompt and engine, with nothing about your brand having changed. We cover why below; the short version is measure how often you show up, not where you land. For the underlying definitions, see what AI search visibility is and why it differs from rankings.

The four metrics, one at a time — and what each one can't tell you

Visibility rate — named ÷ total runs. 14 named out of 20 runs is a 70% visibility rate. It can't tell you whether you were named well — it counts mentions, not quality.

Citation rate — cited ÷ named (pick named or total-runs as your denominator and say which). It can't tell you whether the cited page is the one you'd actually want surfaced.

Share of voice — your named count ÷ all brands' named counts, in the same runs. 12 of a 40-mention total is a 30% share. It can't tell you whether the category itself is growing, or whether a rival's share came from a prompt cluster you don't compete in.

Sentiment and accuracy — read, don't infer. Note the tone (positive/neutral/negative) per mention and check stated facts against your own. Sentiment-by-eye is subjective, so fix your own rule for "neutral" before you start.

None of the four tells you why an engine chose what it chose, or what the reader did next — a real gap no tool we're aware of, including ContextOwl's, currently closes.

Why position is the metric to drop

The clearest evidence comes from a study SparkToro ran with Gumshoe.ai, an AI-tracking vendor, published January 27, 2026. Over November–December 2025, roughly 600 volunteers ran 12 fixed brand-recommendation prompts (chef's knives, headphones, cancer-care hospitals, marketing consultants, and more) through ChatGPT, Claude and Google's AI Overviews/AI Mode, a combined 2,961 times — 60–100 runs per prompt per platform.

The result: under a 1-in-100 chance that two runs of the identical prompt returned the identical list of brands, and under a 1-in-1,000 chance of the identical order. List length varied too. But frequency held up better than order — for a given intent, the brands that showed up most often did so across a clear majority of runs, even as their order stayed close to random. (We're stating the frequency finding in general terms rather than a specific percentage, since secondary coverage of the study gives slightly different ranges and we couldn't independently recompute the raw data — see Sources below.)

Two disclosures: Gumshoe.ai sells AI-tracking software and said so in the study; ContextOwl sells one-time reports, sending each prompt once per engine — a point-in-time snapshot, exactly why this guide argues for repeated runs over one reading. "You rank #2 for this prompt" is close to meaningless alone. Ask instead: named in how many of how many runs, and cited in how many of those.

How many prompts and runs are enough?

There's no single magic number, and treating one as exact repeats the position mistake. Smoot's research on KPIs for AI visibility (14 September 2026) argues a single daily query "cannot provide a reliable estimate of true visibility" and points toward roughly seven to eight runs per prompt on a rolling two-to-four-week window rather than week-to-week snapshots. The same piece notes different AI-visibility tools measuring the same brand have produced results reported as up to 8.2x apart — largely because they define "mention" and "citation" differently, a reason to fix your own definitions before comparing numbers across tools or time.

SEOFOMO's 2026 survey (fielded 17–27 August 2026, 171 search professionals, 36 countries) found 92% now track AI visibility and citations in some form, up from 78% a year earlier, and that "reliably measuring AI visibility" was the top-cited challenge — evidence this is standard practice now, not that any one cadence is settled.

A workable manual floor: 8–10 prompts per category, run on 2 engines, repeated on at least 3 occasions spread over 2–4 weeks before drawing a conclusion. Fewer runs than that, and one unusual answer can swing your whole rate.

Keeping prompts fixed across you and your rivals

Share of voice only means something if every brand you're comparing was exposed to the identical prompts, in the same session type, over the same window.

1. Write the prompt once, reuse it verbatim. "Best project management tool for startups" and "top PM software for small teams" are different prompts, not a repeat.
2. Never put your brand name in the prompt unless you're specifically testing what an engine says when asked about your brand by name. Category and buyer-intent prompts let you and competitors compete for the same mention on equal footing.
3. Use fresh, signed-out sessions where the engine allows it. Logged-in history can bias what gets recalled. Note the date, engine, and sign-in state every time.

A simple manual routine: baseline, repeat, compare

1. Baseline. Pick 8–10 buyer-intent prompts, none naming your brand. Run each on 2 engines, in fresh sessions, and log results the same day.
2. Repeat. Re-run the identical prompt list on the identical engines 2–3 more times over the following 2–4 weeks.
3. Compare. Calculate visibility rate, citation rate, share of voice and a sentiment/accuracy count across all runs combined — not per run. Position, if you note it at all, gets a column but no conclusion.

Here's a filled-in log for a fictional brand ("Brand A") selling project-management software, against one named rival ("Rival X") that came up in the same answers:

Illustrative example with a fictional brand and rival.

From this small, illustrative log: Brand A was named in 6 of 8 runs (75% visibility rate), cited in 2 of those 6 (about 33% citation rate), and split 2 positive / 3 neutral / 1 negative on sentiment. Share of voice would need the same table run for Rival X, then Brand A's named-count divided by the combined named-count across identical runs. Position isn't in this table at all — on purpose.

How to report results so they hold up

Report a rate over a stated denominator, dated, with the prompt list and engines attached: "Named in 14 of 20 runs (70%) across 10 prompts on 2 engines, October 1–15, 2026" is a claim someone else can check and repeat. "We're #1 on AI search" is not. Bessemer Venture Partners' review of how marketers measure AI search (1 September 2026) describes teams building proxy systems instead — AI visibility scores, referral-traffic monitoring, self-reported surveys, branded-search-volume tracking — because, in their framing, "perfect GEO attribution doesn't exist." State what you measured and over what window; don't imply precision you don't have.

That also means not defending a single number quarter over quarter: model updates on the engine side can shift your rate overnight for reasons that have nothing to do with your content. Note that limitation alongside any number you report. For a broader argument against ranking-style thinking here, see presence, not position, as the right frame for AI search.

FAQ

How many runs before a visibility number means anything? Treat anything under roughly 15–20 runs as directional. Repeated-prompt research found answers vary enough that one or two runs can mislead; more runs narrow that noise but never remove it.

Why does position vary so much between identical prompts? Engines assemble answers dynamically through retrieval and ranking rather than reading from a fixed list. A controlled study of repeated identical prompts found under a 1-in-1,000 chance of the same brand order twice, even when the same brands came up.

What's the real difference between visibility rate and share of voice? Visibility rate asks "how often am I named, out of all my runs?" Share of voice asks "out of every brand mention in those same runs, what fraction is mine?" A brand can score high on one and low on the other — they answer different questions.

Can sentiment and accuracy checks be automated? Partially. Classifier scoring can flag obviously positive or negative language, but checking facts against your own (price, features, who you serve) still needs a human. Automated sentiment also tends to disagree with human readers on neutral or mixed cases.

Do I need a paid tool to run this routine? No — baseline, repeat, compare works with a spreadsheet and fresh chat sessions. A tool mainly buys scale and consistent logging, not a different metric.

What to do next

1. Write 8–10 buyer-intent prompts for your category, with no brand name in them.
2. Run each on 2 engines in fresh sessions, logging named / cited / sentiment the same day. Repeat 2–3 times over the next 2–4 weeks.
3. Calculate visibility rate, citation rate, share of voice and a sentiment/accuracy count across all runs combined, and report it with its denominator and date range — not a single score.

ContextOwl sells one-time AI visibility reports, not a monitoring subscription. See a full sample report before you run your own.