Share of Answer: The Metric That Replaces Your Ranking

Last week we pointed ChatGPT, Perplexity, Claude, Gemini, Grok, and AI Overviews at one consumer category and asked them the questions a real buyer asks. A hundred and ninety answers came back.

Out of that came one number: the brand we were measuring held 46.4% of the answer.

Not position three. Not page one. There is no position, and there is no page. When a model composes an answer to "what's the best X," your brand is either in the sentence or it isn't — weighted by where in the sentence it shows up. The number that describes that is a share, not a rank.

I think share of answer is the successor metric to search ranking, and I want to walk through what it actually measures and how the measurement mechanically works — not conceptually, mechanically — because this category is filling up fast with dashboards that will sell you a number, and the fastest way to evaluate any of them is to understand what an honest version of the number requires.

Full disclosure before we start: my company builds measurement tooling in exactly this space, so I have a horse in the race. I'm going to show you the machinery anyway. It's the best argument I have that the metric is real.

Why rank stops working

Search ranking was a great metric because of three assumptions it quietly stood on.

AI answers break all three. You cannot hold a "rank" in a surface that has no list.

So the right measurement doesn't come from SEO at all. It comes from a discipline older than search: share of voice. Media buyers have asked for decades, "of all the impressions in this category, what fraction is ours?" — a share, computed against a defined competitive set, robust to the fact that no single impression tells you anything. Share of answer is that instrument, rebuilt for the surface where your buyer now asks. (You'll hear the surrounding practice called GEO or AEO; how it does and doesn't resemble SEO is its own essay.)

Every clause in that sentence is load-bearing. Pull any one out and the number turns into marketing. So let me pull the definition apart the way our own pipeline executes it.

The questions are unbranded by construction

The questions that matter are the ones your buyer asks before they know you exist: what's the best X, which X is worth the money, what should a first-time buyer look for. In our pipeline, the category questions structurally cannot contain the brand's name — the templates have no slot for it. The brand appears in exactly one question in the set: a branded recognition probe, there to measure whether the engine knows you at all. The distance between "knows you" and "names you unprompted" is its own metric — the Discovery Gap — and its own essay.

This matters because the most common way to fake this measurement is to ask branded questions. Ask an engine "tell me about [your brand]" and it will perform knowledge, and the resulting score will flatter you.

The question set is frozen, and versioned

Once a question set goes active in our system, it's immutable. Editing it doesn't modify it — it mints version N+1 as a fresh draft, and every score ever produced carries the version it was measured under, "so a slate change never silently makes old and new scores incomparable," as the comment in the code puts it.

That sounds like bureaucratic fussiness. It's the whole ballgame. A share-of-answer number is only meaningful relative to a prior number from the same questions. Change the questions and quietly keep the trend line, and you've manufactured movement. When the question set legitimately has to change — categories drift, buyers' language changes — the honest move is to restart the baseline and say so on the receipt, which is exactly what the pipeline does. (Baselines and re-baselining deserve their own piece; the one-sentence version is: a trend line that survived a question change without a visible restart is lying.)

Refusals count against everyone

Every answer is stored verbatim, with its capture date — including the answers where the engine declined. Some engines, some days, respond to "which brand do you recommend" with a shrug. Those refusals stay in the denominator. The comment in our scoring module: "they legitimately dilute everyone's share."

Because think about what a refusal is from the buyer's side.

An honest share metric has to record that as a worse day for the whole category, not throw the answer away and quietly compute share over the answers that flattered somebody. The pipeline goes one step further and surfaces a per-engine answered-rate, so you can tell "this engine refuses a lot" apart from "this engine answers fine and omits you." Those are different problems with different fixes.

No model grades the models

Here's the design decision I'd defend hardest, and the first question I'd ask any vendor: the scoring loop is deterministic. There is no LLM anywhere in it.

Detection is literal word-boundary matching — the brand name and its curated aliases, matched exactly, so a brand named Dandy doesn't get credit every time an engine writes "dandelion." Prominence is a 0-to-5 ordinal computed from position.

And emphasis language only upgrades a score when it's attached to a leading mention, because — quoting the code again — a brand buried in the body while "best" appears somewhere else in the response is "descriptive language, not an endorsement of this brand." Before this scorer scored its first real run we ran it against a hand-labeled calibration set and required perfect agreement, 30 out of 30, before trusting it.

Why be this rigid?

Deterministic scoring means identical inputs produce identical outputs, forever — which also means when the weighting improves, you can re-score every stored answer for free and keep one comparable history.

There is real cost to this choice. Literal matching is dumber than a model. It needs curated help — which brings us to the denominator.

The denominator is curated, and it bites

Share is relative, so the system refuses to run without a competitive set. The actual refusal message an operator sees: "A discovery run scores share of answer against the competitor set, so an empty set has nothing to measure."

And competitive sets have a trap that nobody selling you a dashboard will bring up: consumer brands love naming themselves ordinary words. Simple. Native. Honest. Pure. A literal matcher scoring a competitor named "Simple" will find it in every answer that says "keep your routine simple" — silently inflating that competitor's share. Our fix is operator-supplied aliases and exclusion phrases plus a curation-time warning list for exactly these names, and the design comment states the tradeoff plainly: "a false warning costs the operator a glance; a missed one costs a wrong score."

Two more honesty rules round out the mechanics. The scorer refuses to score an incomplete capture — if too few answers came back, the run reports itself short rather than presenting a share computed over a partial picture; the code's phrasing is that it "never throws, never fabricates." And the prominence weighting is labeled in the source as "a first guess, NOT a spec," tuned against real specimen reads rather than defended as revealed truth. The metric is young. An honest instrument says so in its own comments.

Why this becomes the metric that allocates budget

A distribution metric matters when the organization fights over it. Rank was that number for fifteen years. Share of shelf was that number for fifty. Budget followed the metric, and the metric shaped the war.

Share of answer has the three properties that put a number in that seat.

It's competitive by construction — the denominator is your named rivals, so your share can only grow at someone's expense, which is what makes a number worth fighting over.

It's per-engine, and the engines genuinely differ. Each one is a separate machine carrying its own picture of your category. Some browse the live web and cite their sources; from those citations our pipeline aggregates a ranked list of the domains doing the recommending, 91 of them in that one run. Others answer purely from trained memory and return zero citations by design; nothing you publish today reaches them until a training cut does. Same category, structurally different fights, and a single blended score would hide exactly the information you'd act on.

And it sits at the moment of recommendation — not awareness, not traffic, the instant a machine either says your name to a ready buyer or doesn't.

This is also why the metric exists whether or not you instrument it. Operator communities I'm part of already run this sweep by hand — a couple dozen unbranded questions, pasted into the engines, monthly, tallied in a spreadsheet. The hand-run version is crude, but the instinct is exactly right. The answer is being composed either way.

What the number cannot tell you

The honest limits, because a metric pitched without them is a pitch.

So here's the buyer-education payoff, compressed. If anyone tries to sell you this number, ask five questions. What are the exact questions, and are they versioned? Who is in my denominator, and who curated it? How do refusals count? Is the scoring deterministic, or is a model grading the models? And show me the prominence weighting. An honest vendor answers all five in plain language, because an honest version of this metric is nothing but those five decisions. Hand-waving on any of them means you're buying a weather report from someone who won't show you the thermometer.

Rank was a number you could check by typing your keyword into a search bar. Share of answer is a number you have to build: questions, engines, denominators, weights, receipts. That's the cost of the new metric. Here's what it buys: the run I opened with measured a brand's entire recommendation footprint, across every engine its buyers use, against its real competitors, in a single run.

Your next customer asks a model, not a search bar. The answer is already being composed. The only question is whether anyone on your side is counting.

The measurement pipeline described here is built and running on specimen brands today. It is not yet open to customers, and nothing above claims otherwise.