Share of Answer: The AI Visibility Metric That Replaces Share of Voice

Share of answer is the share of a buyer’s prompt set on which your software is named, counted per engine over repeated runs. It takes over the job that share of voice did in paid and organic search. It is computed differently, because an answer carries no impressions and no ranked list of ten links to weight.

This guide defines the metric and its variants and works the arithmetic with the definitions the AI Shortlist Index uses. It also says what sample makes the result worth putting in front of a board. If you want the measurement run for you every edition, the monitoring retainer reports it from the same run files.

What is share of answer?

Take the 25 prompts a property manager in your category would ask. Send each one to an engine several times in fresh sessions. Share of answer is the fraction of those runs in which your name appears in the answer text, after removing the runs where the prompt itself supplied your name.

That is the whole definition. It is the same quantity the Index methodology calls mention rate.

The base figure is per engine. The denominator is the count of runs on that engine that returned a usable answer, over the prompts where you are not seeded. The numerator is the subset of those runs in which your canonical name or a listed alias appears.

A vendor named in the prompt text, as in “alternatives to Latchel”, is recorded but excluded from its own count on that prompt. Otherwise every “alternatives to” prompt would hand the incumbent a free mention on every run.

How is it different from share of voice?

Share of voice in paid search is impressions divided by the impressions you were eligible for. In organic search it is a rank-weighted sum across a keyword set, with each position given a click-through weight from a curve. Both formulas rest on quantities that exist only on a results page.

None of those quantities exists inside an answer. There is no impression log for the answers an engine produces to other people’s prompts. Position exists, but as an ordinal among two to five names in a paragraph of prose, with no published curve for how often a buyer acts on the second name.

The buyer does not have to click anything for the answer to have done its work. A property manager with 300 doors reads two names and books two demos.

There is a second difference that matters for anyone who reads the numbers. Share of voice was built on a fixed inventory, so competitors’ shares summed to one hundred. Shares of answer do not sum to anything, because an answer that names three vendors gives all three a mention on that run.

If a dashboard shows your category’s shares of answer adding up to one hundred percent, it has normalized them. The normalization has thrown away the information about how many names each answer carries.

How do you compute it from a prompt set?

The method has three parts, and the Index publishes all three so that anyone can reproduce a figure.

The prompt set is frozen and versioned

Each Index category has 25 prompts across six types. Seven open shortlist questions, five “alternatives to” prompts, four segment prompts, three integration prompts naming a property management system, four problem-led prompts with no product words, and two head-to-head comparisons. The current set is Prompt Set v1.0, approved 2026-09-02.

Every prompt is public with its id on the prompt library. Editing a prompt to include your product name or a feature only you have measures nothing.

The runs are repeated and separated

Each prompt goes to each engine three times, each in a fresh session with no shared state, spread across at least two days, at engine default settings. The model identifier is stored with every run where the engine discloses it. The engines change models without notice, and a month-to-month move has to be read against that record.

Extraction is deterministic code

A vendor is named when its canonical name or a listed alias appears in the answer. Matching tolerates a space or hyphen between name tokens and treats names that are also common English words, such as Meld or Latch, case-sensitively. Every cited URL is stored with the run.

From there the arithmetic is short.

mention_rate[e]  = runs on engine e in which the vendor is named
                   ÷ usable runs on engine e
                   (over prompts where the vendor is not seeded)

blended          = mean of mention_rate[e] over engines with usable runs

position_score   = mean first-mention ordinal over runs where named
                   (1.0 = always first; lower is better)

consistency      = prompt×engine cells where all runs agreed
                   (all named, or none named)
                   ÷ cells with at least two runs

engine_spread    = count of engines with usable runs and mention_rate[e] > 0

Share of answer is the first line. The other four qualify it, and a share reported without them is a share you cannot interpret.

What are the variants, and which does the Index publish?

Vendors and tools use the phrase loosely, and four variants do not mean the same thing.

Blended share is the mean of the per-engine shares across the engines that answered. It is a mean of rates, not a pool of runs, so an engine with no usable runs is left out instead of pulling the average toward zero.

First-named share is the fraction of usable runs in which you were the first vendor named. A buyer who reads one name and books one demo is measured here and nowhere else.

Position-weighted share gives each run where you appear a weight of one over your position, so first counts 1, second 0.5, third 0.33, divided by the usable runs. The Index does not publish it, because a single folded number hides which of presence and position moved. It publishes mention rate and position score side by side instead.

Prompt coverage is the fraction of prompts on which you appeared at least once across the runs. It is the most flattering of the four, since one appearance in three attempts counts the same as three in three. The gap between coverage and mention rate is what the consistency metric measures directly.

What does a worked example look like?

Suppose a maintenance coordination vendor, call it Vendor A, is seeded in two of the 25 prompts in its category. That leaves 23 eligible prompts, and at three runs each, 69 usable runs per engine. Suppose Google AI Overviews returned nothing that month and ships as degraded, so it is excluded from every denominator.

Engine	Usable runs	Runs with Vendor A	Per-engine share	Where Vendor A appeared
ChatGPT	69	23	0.33	first in all 23
Perplexity	69	46	0.67	second in all 46
Gemini	69	0	0.00	never
Google AI Overviews	degraded	excluded	excluded	excluded

A hypothetical vendor across 23 unseeded prompts and three engines with usable runs. Blended share 0.33, engine spread 2, position score 1.67, first-named share 0.11 across the 207 usable runs, position-weighted share 0.22.

Blended share is the mean of 0.33, 0.67 and 0.00, which is 0.33. Engine spread is 2. Position score is the mean first-mention ordinal over the 69 runs where Vendor A appeared: (23 × 1 + 46 × 2) ÷ 69, or 1.67.

First-named share is 23 ÷ 207, or 0.11. Position-weighted share is (23 × 1 + 46 × 0.5) ÷ 207, or 0.22.

Why consistency changes the story

Now hold every number in that table fixed and change only how the mentions fall across prompts. On ChatGPT, the 23 mentions could fall one per prompt across 23 different cells, or three per prompt across seven cells plus two in an eighth. The per-engine share is 0.33 either way.

The consistency is not. In the first arrangement all 23 ChatGPT cells disagree internally and, with a matching spread on Perplexity, consistency lands near 0.33. In the second, roughly 67 of the 69 cells agree and consistency lands near 0.97.

The first vendor is one an engine sometimes produces. The second is one it reliably places on about a third of the questions and never on the rest. Same share of answer, different situation, different work to do, which is why the Index shows consistency next to mention rate.

What sample size makes it defensible?

A sample is four counts multiplied together: prompts, engines, runs per cell, and days. The Index protocol is 25 prompts, three runs per prompt per engine, across at least two days, for every measured engine. That is 75 runs per engine per category.

Each count is there for a reason. Twenty-five prompts across six intent types keep one phrasing from deciding the result. A vendor that appears on every “best” prompt and no problem-led prompt has a real weakness a five-prompt set would miss.

Three runs per cell is the minimum at which consistency can be computed with a majority. Two days separates the runs from whatever the engine’s search index held on a single afternoon. Fresh sessions and engine defaults keep memory and prior conversation out of the answer.

One query is not a measurement for the same reasons in reverse. It has no denominator, so a share of 1.0 and a share of 0.0 are equally likely to be the run the engine gets wrong one time in three. It has no consistency value, and it cannot separate a change in your pages from a change in the model.

If it was run in a logged-in consumer app, it also carries personalization from an account that is probably yours. That is the one account whose answers do not matter.

How does the Index report it?

The AI Shortlist Index publishes blended mention rate as the basis for shortlist rank, with the per-engine rates alongside it. Then come position score, consistency, engine spread, source dependency, and movement since the prior edition.

The measurement framework defines each in the same words the methodology uses. Share of answer does not appear as a separate column, because the blended mention rate already is that number.

What the Index does not report matters as much. It has no click data: API answers send no traffic anywhere, and a share of answer says nothing about sessions or leads. It is not a share of a market, and being named on a third of the prompts says nothing about doors under management or revenue.

Where the September edition stands

ChatGPT and Claude are measured today. ChatGPT is captured from the ChatGPT app with memory, custom instructions and chat-history search off. Perplexity, Gemini and Google AI Overviews are in the protocol and have no runs yet, so every denominator in the September 2026 edition is a ChatGPT denominator.

The 2026-09 export records 25 ChatGPT app captures in the maintenance and vendor network category, one per prompt. On prompt mvn-best-01, the 300-door maintenance question, the answer named Property Meld first and Latchel second.

Across the category Property Meld holds rank 1 with a mention rate of 0.91, a position score of 2.43 and an engine spread of 1. Lula holds rank 2 at 0.65 and Latchel rank 3 at 0.64. Consistency is null for all of them, because a cell needs two runs before agreement can be measured.

Read as shares of answer, Property Meld sits at 0.91 on ChatGPT with a denominator of one run per prompt, not the three the protocol calls for. The first-named share is the number to watch when the second and third runs land. The maintenance category page says so on the table itself.

What the sources behind the number looked like

A run of the same prompt captured through the OpenAI API on 3 September 2026, retired from the edition and kept in the archive, consulted 124 URLs. By domain the top of the list was reddit.com with 38, appfolio.com with 16, buildium.com with 12, ir.appfolioinc.com with 5, then latchel.com and propertymeld.com with 4 each.

That single run shows the shape of the third part of the method: the pages the engine reads before it names a vendor are mostly not the vendor’s own. The app captures the edition rests on carry only the links in the answer. Across the maintenance category those went mostly to vendor sites: latchel.com 26, propertymeld.com 23 and lula.life 22 of 184 citations, with reddit.com at 6.

What does a marketing lead do with the number?

Put it in the deck as a fraction with its date, its engine and its prompt set version. “Named in 21 of 23 unseeded ChatGPT runs, September 2026 edition, Prompt Set v1.0” survives a board question. “We’re in ChatGPT” does not.

Read the metrics in order

Engine spread first, once more than one engine is measured, because a share of 0.5 built from 1.0 on one engine and 0.0 on another is two different problems.

Position and consistency second. A low share with high consistency means a fixed set of prompts where you are absent, which you can list by prompt id. A middling share with low consistency means the engine is unsure about you everywhere.

Source dependency third, because it is the only metric you can act on directly. It names the domains your mentions ride on, and the per-category Source Map ranks the top 20 domains and URLs with their page types. When those pages are a PMS marketplace, a review platform and a forum, the work is on those pages before it is on yours.

Where to go next

When the engine with the lowest share is ChatGPT, the guide to ranking in ChatGPT explains where the links in its answers come from. When your name shows up under an old brand or a spelling the alias list does not cover, the fix is in entity consistency. The markup guide covers the Organization node that ties profiles together.

Three things not to do, starting with the prompts: do not edit them to include your name, because engines echo prompts and the seeded-mention exclusion exists for that reason. Do not average across categories; the buyer asking about resident benefits is not the buyer asking about maintenance. Do not read a rise as caused by your work until the model record says the model did not change.

A comparison of the tools that claim to measure this is in what the monitoring tools measure. Paying for the monitoring retainer never affects placement, because the Index measures what the engines say and the practice sells the work.

Limits

Extraction is lexical. A vendor referred to only by a nickname not in its alias list is missed until the review queue surfaces it, and no edition publishes with review items pending. Runs are US English and unpersonalized; real buyers get personalized answers, and the Index measures the baseline.

Only ChatGPT and Claude have runs. The protocol’s treatment of the other engines, including recording an absent AI Overview as “no AI Overview” rather than as a zero, is written and untested in a published edition. Sentiment is cue-based and is a coarse framing signal, not a review score.

The metric has no causal story attached: it records who was named and which pages were cited. It does not see your markup, your profiles or your press. Any statement that a specific change moved a share is a hypothesis until it survives more than one edition with the model held constant.

What changes month to month?

The model behind each engine, which is why the identifier is stored per run where disclosed and why movement should be read against it. The prompt set, which changes only by version, with retired prompts kept in the file and every wording change logged.

The engine status, since an engine with no key or a broken API ships as degraded and drops out of the denominators without being counted as zero. The vendor list, since a name confirmed from the review queue joins the table and can take a rank without changing anyone else’s share.

Movement itself, which is null in a first edition and becomes a number only when there is a prior rank to subtract from.

Every change that can move a number is logged on the Index changelog with the edition it takes effect.

Related guides

• AI visibility: the measurement framework, the hub above this page, which defines every Index metric.
• Index methodology, the source of truth for prompt design, capture, extraction and the metric table.
• Entity consistency across LinkedIn, G2, Crunchbase and your site, for the alias and rename problems that lower a share without an engine changing its mind.
• Structured data the engines document reading, for the Organization and SoftwareApplication nodes on your own pages.
• the ChatGPT ranking guide, for the engine most vendors start with.
• AI visibility tools compared, for what the commercial trackers do and do not compute.
• Monthly monitoring, the practice offer that reports these figures per edition.

Changelog

• 2026-10-28: Reformatted with a short answer, question headings and an FAQ. Corrected Property Meld’s September position score to 2.43 from the export. The sample-size and reporting sections now say that only ChatGPT has runs and that the other engines are protocol, not data.
• 2026-09-26: The September figures now cite the 25 ChatGPT app captures per category; the API run of 3 September 2026 is described as retired from the edition and archived.
• 2026-09-19: ChatGPT capture moved from the OpenAI API to the app at chatgpt.com. The section on what the Index does not report was rewritten, and the 2026-09 figures are labeled as the API capture of 3 September 2026.
• 2026-09-04: First published.