How do I measure my brand share in AI answers?

In brief

Brand share in AI answers only means something once you split it three ways: by engine, by the type of buyer question, and by the week it was measured.
ChatGPT, Claude and Perplexity decide who to name differently, so a brand's share rarely moves the same way on all three at once.
Different buyer-question groups, such as "best for X" versus "alternatives to X," pull from different sources, so one group's share says little about another's.

This is one specific measurement inside brand intelligence: the number a buyer question produces, broken down enough to act on, rather than averaged into something that feels precise but hides which part of your presence is actually moving.

What does "brand share in AI answers" actually mean?

It's three separate counts per engine, not one score: how often you're named anywhere in an answer, how often you're recommended in the top three, and how often your own domain is cited as a source. Each count answers a different question about the same response, and a brand can score high on one and low on another at the same time.

Bottom line: if you only track one number, you can't tell which of the three actually moved, or why. The full breakdown of what each count checks lives in how to measure AI visibility.

Why does one share number for "AI" hide more than it shows?

Because ChatGPT, Claude and Perplexity decide who to name through different mechanisms, so a brand's share rarely moves the same way on all three in the same week. Perplexity runs a live search for most answers and shows numbered source links. ChatGPT blends training data with optional live search through OAI-SearchBot. Claude answers mostly from training data unless web search is switched on, and in our checks cites a source in only about half of its answers, roughly half the time there's no source list to even measure against.

Engine	How it mostly decides who to name	What that means for your share number
Perplexity	Live search on most questions, sources always shown	Share tracks your current site closely, correction reach fast
ChatGPT	Training data, plus live search when OAI-SearchBot fetches	Share mixes a slow signal with a fast one; split them if you can
Claude	Mostly training data unless web search is on	Cited is often not measurable at all; lean on Named and Recommended instead

Do this: report your share per engine before you ever average them. An average of three different mechanisms isn't a number, it's a blur.

How do I break my share down by query group?

Group your buyer questions by intent before you measure, because "best X for Y," "alternatives to X" and "cheapest X" questions each pull from a different mix of sources. A question that already names your brand, such as "what are the alternatives to [you]," creates an artificial chance to be mentioned and should sit in its own group rather than inflate a group it doesn't belong to.

Best-for-X questions lean on comparison content and reviews.
Alternatives-to-X questions already name a brand, so exclude them from that brand's own denominator.
Cheapest/for-a-budget questions lean on pricing pages and dated vendor data.
For-a-team-size or for-a-use-case questions lean on segment-specific pages, not the homepage.

Why it matters: a brand can hold a strong share in "best for enterprise" questions and a weak one in "cheapest" questions, and a single overall number would report that as "average," which describes neither reality.

How do I track my share week over week?

Run the same fixed question set on a schedule and watch for a move that holds across several weeks, not a single run. Research on how consistently language models answer the exact same prompt on repeat finds consistency rates that vary widely and are routinely well short of 100%, which is exactly why one answer, asked once, isn't a measurement.

Proof: arXiv — The Non-Determinism of LLMs: Evidence of Low Answer Consistency in Repetition Trials.

Rule of thumb: never change your question set between runs. A week-over-week share only means something if the questions, the engines and the counting rule stay identical.

What to do this week

None of this requires buying anything, and each step takes a few minutes:

Write 10-20 real buyer questions in your category and sort them into 3-4 groups by intent, not by your own brand name.
Ask each question 3 times per engine across ChatGPT, Claude and Perplexity, and log Named, Recommended and Cited separately, by engine and by group.
Run the free AI Crawler Access Check before you conclude a low Cited count is a content problem rather than a blocked fetcher.
Run the free Schema & Entity check if your Named count looks healthy but Recommended doesn't, since entity confusion often explains the gap.
Run the free check at dashboard.reflexaagent.com/audit to get all three counts computed automatically, by engine, with evidence attached to every answer.

Bottom line: a breakdown by engine and by group turns one vague share number into several specific, fixable ones.

Keep reading

Frequently asked questions

What counts as my brand share in AI answers?

The share of eligible ChatGPT, Claude and Perplexity answers to real buyer questions where you're named, recommended in the top three, or cited as a source, each tracked as its own number rather than blended into one score.

How many buyer questions do I need for a reliable number?

10-20 real buyer questions, each asked three times per engine, is enough for a first read. The Reflexa AI Visibility Index runs the same method at 100 questions and 900 observed answers per category edition.

Why does my share differ between ChatGPT, Claude and Perplexity?

Each engine decides who to name differently. Perplexity always runs a live search and cites sources; ChatGPT blends training data with optional live search; Claude answers mostly from training data unless web search is on, and cites a source in only about half of its answers in our checks.

How often should I re-measure my brand share?

Weekly, on the same fixed question set. AI answers are non-deterministic, so a single run is a snapshot, not a trend; only a sustained move across several weekly runs is worth acting on.

Does a low share mean AI is wrong about me?

Not necessarily. A low share can mean you're accurately but rarely the best fit for the questions you tested, or it can mean a crawler block, thin entity signals, or missing third-party evidence are keeping you out of answers you should win.