What is Share of AI Answer?

Share of AI Answer is the percentage of tracked question–engine pairs where an AI assistant (ChatGPT, Claude, Gemini, Perplexity or Grok) names or cites your brand in its answer. The formula is cited results ÷ completed results × 100. If you track 40 buyer questions across 5 engines, that is 200 answers per run; if your brand appears in 68 of them, your Share of AI Answer is 34%.

Last updated 2026-07-28

What is the exact formula for Share of AI Answer?

Share of AI Answer = (number of results where the brand was detected ÷ number of results that completed without an error) × 100, computed per run. A "result" is one prompt sent to one engine — so a 40-prompt set across 5 engines produces 200 results per run, and each one is a single yes/no on whether your brand appeared.

Two details in that formula do real work. First, failed calls are excluded from the denominator, not counted as misses. If Gemini rate-limits you on 12 of 40 prompts, those 12 are dropped from both sides of the ratio — counting a timeout as "not cited" would make an outage look like a visibility collapse. Second, the score is computed per engine and then overall, because a single blended number hides the thing you can act on: which engine you are losing.

How does a tracker decide that a brand was "cited"?

A brand counts as cited when either its domain appears among the answer's structured citations, or its name appears in the answer text. Both signals count, because an assistant can recommend a product by name without linking to it — and can link to a page about you without ever naming you.

AI Visibility Tracker resolves this with two separate checks against a brand profile you define (names plus domains). The URL check normalises each cited link to a hostname, strips a leading www., and matches the domain or any subdomain of it. The text check is case-insensitive and diacritic-insensitive (so "Café Nero" matches "cafe nero"), anchored to word boundaries so "Notion" does not match "notional", and tolerant of a trailing "s" or "'s" and of hyphen/space variation, so "Answer Engine" matches "answer-engine".

Everything cited that is not one of your domains is recorded as a competitor domain on that result. That is what makes the loss data useful rather than merely discouraging: for every prompt you lose, you get the list of domains that won it. See AI visibility metrics that matter for how to read that list.

Why must every engine run with web search enabled?

Because an ungrounded model invents its citations. Ask a model without web access which tools lead a category and it will answer confidently from training data that is months stale, and it will frequently produce URLs that look plausible and do not exist. Score those answers and your Share of AI Answer is not a weak measurement — it is a measurement of a different, imaginary web.

This is the single most common way AI visibility numbers get quietly corrupted, and it is invisible in the output: a hallucinated citation looks exactly like a real one in a spreadsheet. AI Visibility Tracker runs every engine in its web-grounded mode without an option to turn that off, and reads citations only from the provider's structured metadata — never by regex-scraping URLs out of prose. The full mechanism per engine is in why web-grounded measurement is non-negotiable.

What counts as a good Share of AI Answer score?

There is no universal benchmark, and anyone quoting one is selling something. A niche B2B tool cited on 30% of its category questions may be dominating a small field; a consumer brand at 30% may be losing badly. The number that matters is your own trend and your own gap.

Read it this way: movements of a few percentage points between daily runs are normal sampling noise, because grounded answers are stochastic — the same prompt genuinely can cite you today and not tomorrow. Sustained direction across three or four weeks is signal. A per-engine spread (say 50% on ChatGPT and 10% on Gemini) is more actionable than the headline number, because it tells you whether your problem is classic search visibility or something engine-specific.

How often should Share of AI Answer be measured?

Daily on a small core prompt set, weekly on the full set. Daily sampling is what turns a stochastic answer surface into a readable trend; a full sweep of every prompt every day mostly buys you noise and API spend.

AI Visibility Tracker builds this in with two prompt tiers. Core prompts — typically the 10 to 15 questions you must not lose — run on the nightly schedule. Extended prompts only run in a weekly full sweep. Trends are kept comparable within their own series, so a Monday full sweep does not create a false spike against Sunday's core run.

What are the most common ways this metric gets measured wrong?

The four failures below account for nearly every AI visibility number that turns out to mean nothing. Each one produces a plausible-looking chart.

• Spot-checking by hand. One prompt, asked once, in one chat window, is an anecdote. Answers vary run to run; a single sample cannot distinguish a real change from ordinary variance.
• Measuring without grounding. You get the model's memory instead of the answer surface your buyers actually see — including citations to pages that do not exist.
• Tracking only branded prompts. Most buyer research never mentions any brand name. "Best CRM for small agencies" is where discovery happens, and a brand-name-only prompt set will report excellent health while you are invisible in the questions that matter.
• Changing the prompt set constantly. The prompt set is the yardstick. Add a prompt and your denominator changes; rewrite one and its history is not comparable. Fix the set, then hold it — a disciplined prompt set held for six months is worth more than a sophisticated dashboard over a shifting one.
• Changing the model. Measuring ChatGPT with a cheap mini model because it costs less means you are no longer measuring what ChatGPT users see. Measurement models should be fixed and consumer-representative even though that costs more than the alternative.

How is Share of AI Answer different from share of voice?

Same idea, radically different arithmetic. Share of voice measures your slice of a surface where everyone gets some slice — ad impressions, search results, press coverage. Share of AI Answer measures presence in a surface that names three or four options and stops.

That is the substantive change. Being the fourth organic result on a search page still earned you clicks; being the fourth-best candidate for an AI answer that names three earns you nothing. AI answers are winner-take-most, which is why absence — not ranking — is the thing to measure. The strategic response is answer engine optimization; the measurement discipline is this page.