What the AI Engines Say About Us: the September 2026 Reading

In April 2026 an outside agency scored this site's AI visibility at 3 out of 100. That reading was taken against the old website, on a set of 32 test queries, using the agency's own scale. In September 2026 we began tracking the same question with a different tool, on a different scale, across 47 prompts and five answer surfaces. This post reports what that tool shows, and says plainly what the two numbers can and cannot be compared on.

The two readings are not the same measurement

Two things changed between April and September: the website, and the instrument. The site was rebuilt in the interim, and the tracking moved from the agency's composite to a per-engine export. So we have a baseline from April and a first real reading from September, and no defensible growth rate between them. Anyone telling you they can compute one is computing the difference between two different scales.

We are publishing the September figures rather than a comparison, because the September figures are the ones that are true.

What the September 2026 export shows

Five answer surfaces, 47 prompts, exported 18 September 2026. Mention rate is the share of tracked prompts where this firm appears at all. Average position is where the mention sits in the answer when it happens. The visibility score is the tool's own composite of the two, on its own hundred-point scale.

• ChatGPT: mentioned in 26% of tracked prompts, average position 1.8, visibility score 45.
• Perplexity: mentioned in 28%, average position 2.9, visibility score 47.
• Gemini: mentioned in 18%, average position 3.5, visibility score 36.
• Google AI Overviews: mentioned in 9%, average position 1.9, visibility score 33.
• Google AI Mode: mentioned in 6%, average position 1.4, visibility score 33.
• All surfaces together: mentioned in 17%, average position 2.5, visibility score 40.

The shape of that list is the finding. Where we appear at all, we appear early in the answer, and in the September export we rank second by visibility score on ChatGPT, of the 155 firms the tool scored there, and second on Perplexity, of 142. What varies, and varies enormously, is whether we appear at all: between six percent of prompts and twenty-eight percent, depending entirely on which engine a board happens to ask.

What is tracked, and what is not

The 47 prompts sit across five topics: lender comparison and selection, loan structuring and underwriting, construction and condominium project financing, financing tools and portals, and statewide or local association financing. They are the questions a board actually types, not keywords. The five surfaces are ChatGPT, Perplexity, Gemini, Google's AI Overviews and Google's AI Mode.

Claude is not tracked, so this post makes no claim about what Claude tells a board. An earlier version of this post reported a Claude score. There was never any data behind it.

Which of our pages get cited

The homepage does most of the work by a wide margin. Beyond it, the pages the export records being used as sources are the services page, the about page, the blog index, and eight posts: the two lender comparisons, the comparison of loans against special assessments, the guide to choosing HOA financing, the explainer on why an HOA loan is not a mortgage, the piece on how HOA loan interest rates are set, and the two posts about our application portal.

That is a short list, and it says something useful about what an answer engine does with a blog. It does not cite a site. It cites a page that answers the question asked, and eight of ours currently qualify.

Where we are weakest

Google's own surfaces. Nine percent mention rate on AI Overviews and six percent on AI Mode, against twenty-six and twenty-eight on ChatGPT and Perplexity. On AI Mode the export records no instance of our own pages being used as a source at all. That is the gap, it is measured, and naming it is more useful than a target.

What a board should take from this

Ask more than one engine, and do not treat the first answer as the market. Our own mention rate moves by nearly five times across five surfaces asked the same 47 questions. That is not a fact about us. It is a fact about how these systems assemble an answer, and it applies to every firm you are looking at. A name that does not appear on one engine has not been ruled out, and a name that appears on all five has not been vetted.

The more useful test is what the engine cites. If the answer rests on a lender's own marketing page, you are reading marketing with a summary layer on top.

What we are not publishing

A target. The previous version of this post set a year-two target on a hundred-point scale that we no longer measure against, which made it a number with nothing behind it. We will publish the next reading when there is one, dated, on the same instrument, and the comparison will be worth something because the instrument will not have moved.