AI Visibility: What It Is and How to Measure Whether AI Recommends Your Software

AI visibility is whether an AI answer includes your software when a property manager asks. It covers how early in the answer, how consistently across repeated asks, and on the strength of which pages. This page defines the numbers the AI Shortlist Index publishes, in the same words the Index methodology uses, and shows how to produce them yourself from the public prompt library.

It also explains why a single run per prompt, including the one in our own first edition, is a status report rather than a result. If you want the measurement run for you, the monitoring retainer reports it from the same run files every edition.

What is AI visibility?

AI visibility is a count of presence inside answers. A buyer asks an assistant for a shortlist; the assistant writes two to five names; you are one of them or you are not. Measured across many prompts and many runs, that presence becomes a rate.

It is not a search ranking, a traffic figure or a sentiment score. It is also not a share of a market: a vendor with a small customer base and a well-cited marketplace listing can outscore a larger one.

Why does rank tracking miss it?

A property manager with 300 doors opens ChatGPT and asks which maintenance coordination software to use. The answer names two vendors. The manager books demos with both.

No ranking changed that Search Console can show. No session hit anyone’s analytics. The decision was made inside the answer, and the only trace of it is the answer itself.

Rank tracking measures position on a results page. Analytics measure visits. Measuring AI visibility means capturing the answer, extracting who was named, and doing that often enough that the number means something.

What is the unit of measurement?

Every Index number is built from three things: a prompt, an engine, and a run.

The prompt

A prompt is one question a property manager would ask, written in first person with a real portfolio detail, frozen per edition as part of a versioned set.

Each category has 25, across six types. Seven open shortlist questions, five “alternatives to” questions, four segment questions and three integration questions naming a PMS such as AppFolio or Yardi. Then four problem-led questions 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.

The engine

An engine is one of five capture paths in the protocol. ChatGPT is read from its app and Claude from the Anthropic API with web search. Perplexity runs on its API with the sonar model, Gemini on its API with Google Search grounding, and Google AI Overviews through a licensed SERP data provider.

All five are documented on the methodology page, and ChatGPT and Claude have runs today.

Captures from October 2026 on record the model that answered from the conversation data the ChatGPT page loads. September 2026 captures carry no model id.

The run

A run is one prompt sent to one engine in a fresh session. The protocol calls for three runs per prompt per engine, spread across at least two days, at engine default settings. A prompt paired with an engine is a cell, and a cell holds the runs for that pair.

What “named” means

A vendor is named when its canonical name or a listed alias appears in the answer text. Matching tolerates hyphen and space variants and treats common-word names like Meld or Latch case-sensitively.

Vendors named in the prompt itself, as in “alternatives to Latchel”, are recorded but excluded from Latchel’s count for that prompt. Otherwise the prompt would hand the incumbent a free mention every run.

Which metrics does the Index publish?

All metrics are per vendor, per category, per edition. Seeded mentions are excluded. Engines that did not return usable runs are excluded from denominators rather than counted as zero.

Metric	Definition
Mention rate (per engine)	Runs in which the vendor was named, divided by ok runs for that engine, over prompts where the vendor is not seeded
Mention rate (blended)	Mean of per-engine mention rates across engines that returned ok runs
Shortlist rank	Ordinal by blended mention rate; ties broken by better position score, then wider engine spread, then name. Vendors never named have no rank
Position score	Mean first-mention position across runs where the vendor was named; 1.0 means always first; lower is better
Consistency	Share of prompt-by-engine cells with at least two runs where all runs agreed, all named or none named; 1.0 is fully consistent
Engine spread	Number of ok engines, 0 to 4, with a non-zero mention rate
Source dependency	Top cited domains on the runs where the vendor was named
Movement	Prior-edition rank minus current rank; positive means moved up; null with no prior edition or no prior rank

Eight metrics, each per vendor per category per edition. Mention rate is the base; rank orders it; position, consistency and spread qualify it; source dependency explains it; movement compares editions.

Mention rate and position

Mention rate is the foundation. It answers “when the engine is asked, how often do you come up”, per engine and blended. A blended rate of 0.5 across two ok engines could be 1.0 on one and 0.0 on the other, which is why the per-engine rates are published alongside it.

Position score answers “when you come up, how early”. Being named first in a two-vendor answer and fourth in a six-vendor answer are different outcomes for a buyer who reads the first name and stops.

Consistency and spread

Consistency shows whether that number is stable or the engine is flipping. It is displayed, never hidden. A vendor with a mention rate of 0.33 and low consistency is a vendor the engine sometimes remembers, which is a different situation from one the engine reliably places third.

Engine spread tells you whether a row is one engine’s opinion or all of them. A vendor at spread 1 has a single point of failure; when that model changes, so may the row.

Source dependency, movement and sentiment

Source dependency names the pages the mentions ride on. It is the metric a vendor can act on, because it names the domains. The per-category Source Map goes further: the top 20 cited URLs and top 20 cited domains, each with page type, which engines cited it, and the vendors it supports.

When webpossible.com or another site WebPossible operates appears, it is flagged as own network and disclosed inline. Movement is the change since the last edition. It is null in a first edition, and null is a real value, not a missing one.

Sentiment is recorded too, as recommended, neutral or cautioned, from cue words within a 180-character window of the first mention. It is a coarse framing signal and is displayed as one. It is not a review score.

What is share of answer?

Share of answer is the phrase we use for mention rate read across a whole prompt set. Of all the answers a category’s prompts produced, it is the share in which you were named. The Index does not publish it as a separate column, because the blended mention rate already is that number over the 25-prompt set.

The phrase exists because it replaces share of voice, which the earlier version of this page anchored on. Share of voice was built for impressions, and there are no impressions to weight by inside an answer. An answer either named you or it did not.

The full definition, the variants and a worked calculation are in the share of answer guide. If you would rather buy the measurement than run it, the comparison of AI visibility tools sets the commercial trackers and a do-it-yourself setup side by side.

What does the September edition show?

The September 2026 edition is the Index’s first, and it is one run deep on one engine. Every Index page discloses that, and the empty cells teach the metrics better than a full table would.

The site export for 2026-09 records 25 ChatGPT runs in the maintenance and vendor network category, one per prompt, captured from the ChatGPT app in September 2026. Property Meld holds rank 1 with a mention rate of 0.91, Lula rank 2 at 0.65 and Latchel rank 3 at 0.64; each has an engine spread of 1.

Across those runs ChatGPT cited 184 URLs, and by domain the top of the list was latchel.com with 26, propertymeld.com with 23 and lula.life with 22. Each run carries the answer, its links, the untouched capture and the capture time.

Now read the nulls. Consistency is null for every vendor because a cell needs at least two runs before agreement can be measured, and each cell has one. Movement is null because there is no prior edition.

Engine spread is 1 because one engine answered, and an engine that has not answered is excluded from every denominator rather than scored as zero. A mention rate of 0.91 sounds like dominance; it is one run per prompt on one engine, and the maintenance category page says so on the table itself.

Why is one run not a result?

Engines are non-deterministic. The same prompt, sent twice in fresh sessions, produces different answers, sometimes with different vendors in them. Three runs per prompt do not remove that variance; they make it measurable, which is what the consistency metric is for.

Models drift, because the engines change models without notice. Where a capture records a model identifier it is stored per run, so month-to-month movement can be read against a model change. September’s ChatGPT captures carry none, so for that edition a row can move on a model change with nothing recorded.

The instrument matters. ChatGPT is read from its app, with memory off and the chat set to Unpersonalized. That gets the product a buyer uses at the cost of what the API disclosed: no token count, no price, and no list of the pages behind the answer.

The protocol reads the other engines through documented APIs or a licensed provider, which are repeatable and not identical to those vendors’ consumer apps.

A screenshot of an answer naming your product is an anecdote. It might be the run the engine gets right one time in three. The number you want is the one that survives three fresh sessions on two days, and the number you should distrust is any single capture, including ours.

How do you measure AI visibility yourself?

You do not need the Index to run the method. The prompts are public and the protocol is written down.

1. Take your category’s 25 prompts verbatim from the prompt library, with their ids, so anyone can reproduce what you did. Do not edit them to include your name or a feature only you have; that measures nothing.
2. Send each prompt to each engine three times, in fresh sessions, on at least two separate days. Use the APIs where you can, with web search or grounding enabled, so the answer carries its citations. Where you use a consumer app, as the Index does for ChatGPT, use a temporary chat with memory off, and record that you did so.
3. Store the full text of every response. Then record the vendors named in order of first mention and the framing around your own mention. Record every cited URL with its domain, the timestamp, and the model identifier where you can get one. If the prompt names a vendor, mark that mention as seeded and exclude it from that prompt’s counts.
4. Compute. Per engine, divide the runs that named you by the ok runs. Blend across engines that answered. Average your first-mention position over the runs that named you. For each prompt-by-engine cell, note whether the runs agreed. Count the domains on the runs that named you and sort.

That is a mention rate, a position score, a consistency figure and a source dependency list, produced the same way the Index produces them.

If that is too much to start with

If 75 runs per engine is more than you want to start with, the five-prompt checker gives a quick read of whether you appear at all. Treat it as a smoke test, not a measurement. A longer walkthrough with a spreadsheet layout is in how to measure AI visibility for property-ops software.

What are the numbers for?

For a marketing lead, the metrics answer three questions in order. Whether you are named at all, per engine; when you are, how early and how reliably; and which pages are carrying you. The third is the only one you can act on directly.

That third answer is what goes in the plan, and the engine guides explain where those pages come from. The guide to ranking in ChatGPT covers OpenAI’s web search sources and the Gemini guide covers Google Search grounding. The structured data guide covers what markup the engines’ documentation supports.

WebPossible’s practice sells the work behind those numbers, never the numbers. The monitoring retainer delivers a movement report from every edition, using the same run files the public Index publishes. The AI search optimization overview sits above this page and the engine guides.

Paying WebPossible never affects Index placement. The Index measures what the engines say, and the practice sells the work that earns a position, not the position.

Limits

Runs are US English, unpersonalized, with no account history and no location beyond “US”. Real buyers get personalized answers; the Index measures the unpersonalized baseline.

Only ChatGPT and Claude have runs. The protocol’s handling 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.

Extraction is lexical, so a vendor mentioned by a nickname not listed as an alias is missed until the review queue surfaces it. No edition publishes while review items are pending, and sentiment is cue-based.

Extraction accuracy is checked against a hand-labeled set of at least 50 responses and must reach 95 percent agreement before metrics publish. The labeled set is never edited to raise the score.

Changelog

• 2026-10-28: Reformatted with a short answer, a definition section, question headings, the self-measurement protocol as a numbered list, and an FAQ. Corrected the engine section: only ChatGPT has runs, the October 2026 captures record the model, and the AI Overviews capture is protocol rather than practice.
• 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 capture-path list, the model-drift limitation and the instrument limitation were rewritten. The 2026-09 figures are unchanged and labeled as the API capture of 3 September 2026.
• 2026-09-03: Rebuilt as the Index’s measurement pillar. Metrics redefined to match the methodology: “citation rate” became mention rate and “share of voice” became share of answer. “Agentic referrals” was dropped as unmeasurable inside an answer; sentiment kept as a coarse framing signal.
• 2026-06-10: First published as a general-purpose page on AI visibility with four metrics: citation rate, share of voice, sentiment, and agentic referrals.