How to Measure AI Visibility for Property Ops Software: Notes from the First Run

Somewhere in the last two years the question a property manager types changed shape. It used to be “maintenance coordination software” into a search box, then a scroll through ten links and three ads.

Now it is a full sentence into ChatGPT or Perplexity: “What’s the best maintenance coordination software for a property management company with around 300 doors?” The answer comes back as two or three vendors. Nobody clicked anything to get there, and no analytics tool on the vendor side saw it happen.

These are notes on how we measure that moment, written after the first run of the AI Shortlist Index.

What changed when property managers started asking?

A search result is a list of pages. An answer is a list of vendors, and that difference is the whole reason old measurement stops working.

Position tracking tells you where your page sits for a search term. It tells you nothing about whether an engine that read that page included you in what a 300-door operator sees. Sessions tell you who arrived, and the buyer who took the shortlist straight to a demo request never arrived through you.

The second change is that the question carries context now. A PM does not ask for “maintenance software.” She asks for maintenance software for 300 scattered single-family doors on AppFolio, or for a 2,000-unit multifamily portfolio on Yardi with an in-house tech team.

The engine answers that specific question, and the shortlist can change with every clause. Measuring one generic query tells you almost nothing about the questions buyers ask.

The third change is that the answer is not stable. Ask twice and you may get two different lists, which is inconvenient for anyone who wants a scoreboard and is the reason a single check is worthless as evidence.

Which four numbers do we record?

The Index methodology defines the metrics in full. The short version is four figures per vendor, per category, per edition.

1. Mention rate. Of the runs where the vendor could have appeared, the share where it did, per engine and blended across the engines that answered. Prompts that put a vendor in the question (“alternatives to X”) do not count toward X’s rate, because that would hand X a mention for free.
2. Position. When the vendor appeared, where in the answer it came, counting the first vendor as 1. A vendor that always appears fourth has a different problem from one that appears first half the time and not at all the rest.
3. Consistency. Across repeated runs of the same prompt on the same engine, how often the runs agreed about whether the vendor appeared. A low figure means the engine is flipping, and a flip is not a trend.
4. The source map. Every URL the answer pointed to, normalized and classified by page type: vendor site, review platform, listicle, forum, trade publication, directory. This is the one that tells a vendor what to do next.

We also record engine spread, meaning how many measured engines included a vendor at all. Framing is stored too, as recommended, neutral or cautioned from cue words near the first mention. The four above carry the weight.

Why is one run not a result?

Most do-it-yourself measurement goes wrong at the same spot. Someone on the marketing team asks ChatGPT the question once, screenshots the answer, and takes it to the Monday meeting as “we’re not in there.”

Maybe they are not. Or maybe that run went one way and the next one goes the other.

The Index protocol is three runs per prompt per engine, each in a fresh session, spread across at least two days, with 25 prompts per category. On one engine that is 75 answers per category before anyone reads a number. The consistency metric is published next to the rank, so a rank built on a coin flip is visible as one.

One run gives you an observation. Three runs on separate days give you a measurement with an error bar you can see, and the difference matters most for vendors in the middle of the table.

What did the September 2026 edition show?

The first edition ran one engine, ChatGPT, once per prompt, across all 20 categories. Captured from the ChatGPT app in September 2026, the 25 maintenance prompts named Property Meld in 91% of runs, Lula in 65% and Latchel in 64%.

The domain ChatGPT cited most often across those answers was latchel.com, at 26 of 184 citations, with propertymeld.com and lula.life next. That is the verified set: one engine, one run per prompt, three vendors clearly ahead of the field.

Everything else on the edition’s pages is either a null or a description of what later editions will hold. Consistency needs a second run, and movement needs a prior edition.

What one run suggests

What one run per prompt suggests, not shows, is still worth writing down. Suppose vendor sites stay the most-cited domains in maintenance coordination across the full run set. Then a vendor’s own product pages may matter more than its presence in forum threads for this category.

And if the same three vendors hold the top three slots across three runs, that would be an unusually stable category. Our expectation is that they will not, and that the consistency figure will show movement. Both are hypotheses, and the next edition will test them.

The row data is on the maintenance and vendor network category page, with the engines that answered shown on it.

How do you do this yourself, badly and then well?

Badly is fine to start. Write down the ten questions your buyers ask right before they book a demo, in their words, with a portfolio size and a PMS in the sentence.

Run each one in a fresh ChatGPT session, and in Perplexity and Gemini if your buyers use them. Record who appeared, in what order, and which sources showed up under the answer. Do it again in three days and compare.

That gets you a rough mention rate and a first look at consistency, and it will already tell you something the rank tracker cannot. Where it falls apart is scale and repeatability.

You will not run 75 prompts three times a month by hand. You will start rephrasing prompts that gave bad answers without noticing. And you will have no record anyone can audit.

Well means fixing those three things: a frozen prompt set, repeatable runs in clean sessions on a schedule, and raw answers stored before anyone computes anything. That is the Index, and the AI visibility measurement guide walks through each part.

Once you know whether you appeared, the next question is why, and the answer usually sits in the pages the engine chose to read. We wrote up what an engine pulls from a vendor page separately.

More field notes from each edition are on the blog.