One lead arrived from an AI answer, not two, and it still proves nothing

Correction, 5 August 2026. The enquiry below is real and the channel is right. What it was about was wrong. We wrote that a US commercial real estate company found REN.PH, our Philippine property data platform, through ChatGPT. Going back to ask what they had typed produced a different story: they asked which firms do real estate AI search visibility, and the answer named RealEstateSEO.ph, the service brand we run beside the platform. They reached us through the parent consultancy from there.

So an engine recommended a service. It did not cite a dataset. Those are different claims, and the one we published is the one that made our own architecture look like the cause. The count is unchanged at one enquiry and one lead. What changed is that only the lead is evidence about the data platform, which is the thing this post spends its second half arguing for.

The rest of the post is left as written on 2 August, with the sentences that carried the wrong attribution corrected in place and marked here rather than silently.

Correction, 2 August 2026. The first version of this post was titled “Two leads arrived from AI answers” and treated both enquiries below as the same kind of evidence. They are not. One was traced to ChatGPT by asking the buyer. The other came through Google Search and was never asked anything. Counting it as an AI referral was the loose attribution this firm sells against, published on our own site, which is worse than getting it wrong about a client. The post is rewritten rather than deleted, because a correction that removes the original is not a correction.

Two smaller things went with it. The original described the ChatGPT enquiry as having mentioned the channel “in passing, not in response to a question we had designed to be answerable”. That was written loosely: we did ask them and they answered. The distinction is not cosmetic, because the counting rule at the bottom of this post turns on it, so the sloppier of the two descriptions should not be the one left standing. And the original had no counting rule at all, which is the gap the Google referral walked through.

In June 2026 a commercial real estate company in the United States asked ChatGPT which firms do real estate AI search visibility, was pointed at RealEstateSEO.ph, and reached us through the parent consultancy. We know that because we asked the buyer and they told us. The engine named a service. REN.PH was not the artifact it cited.

Three weeks later a proptech company in Zurich found the same platform through Google Search and made contact on LinkedIn. That is organic search. Google may well have been showing them an AI Overview at the time, and that is exactly the problem: an AI Overview click and an ordinary organic click are the same referrer, so there is nothing to trace and no honest way to claim one.

So the count is one enquiry, not two.

What the lead capture recorded

Separately from those two, REN.PH publishes Philippine zonal values as pages down to barangay level, and lead capture on them is source-tracked at first touch.

Between 14 May and 1 August 2026, a 79-day window, those pages produced 133 leads. Exactly one carried a chatgpt.com referrer, and it landed on a barangay-level page rather than on the homepage. Data collected 1 August 2026.

That single lead is a better data point than the enquiry we asked about, for one reason: nobody had to remember anything. A header recorded it.

What is wrong with all of it

These are floors, not counts. A share of all leads arrive with no referrer at all, and the share is worse for AI sessions, because mobile app webviews strip the header. We have not measured either share on our own traffic, and we are not going to put somebody else’s percentage here to fill the gap: a figure we cannot stand behind is exactly what this post is a correction about. So the bullet is a direction and not an adjustment. The true numbers are not lower than 1 and 1, we cannot tell you how much higher they are, and a firm that reported those two as totals would be doing the thing this site spends two sections refusing to do.

One and one are not a sample. Any business operating long enough accumulates a handful of inbound enquiries with an unusual origin. Two traced events in one quarter are consistent with a real mechanism and equally consistent with coincidence, and nothing about the size of the number distinguishes those.

Different audiences. The 133 are consumers looking at Philippine property valuations. The buyers this firm sells to are proptech and CRE software companies. Same mechanism, a dataset published as pages and found by people asking questions, but a market that has not been tested.

The query is a recollection, not a record. We went back and asked what they had typed. What came back is the shape of the question, which firms do real estate AI search visibility, rather than the string itself. That recollection is what surfaced the correction at the top of this post, so it earned its keep. It is still somebody remembering a search from two months earlier, which means we can approximate the query and we cannot reproduce the session.

We are the interested party. We now sell a service premised on this mechanism. That is exactly the position in which people find patterns in small numbers, and it is the position we were in when we wrote “two”.

Survivorship. We know about the ones who made contact. We know nothing about how many were shown a competitor instead, which is the number that would actually tell us whether the mechanism favours us or is simply noisy.

Why we still think it is a mechanism

Not because of the count. Because of what REN.PH is.

It is a national property dataset holding 234,337 zonal value rows and 25,264 broker records across 37,660 barangays, counted 8 June 2026. It was built with entity architecture, provenance tracking, verification, and freshness systems in place, because a dataset of that shape does not function without them. It happens to be exactly the structure a retrieval system can read and quote: specific figures, stated provenance, dated records, resolvable entities.

Those are record counts. This post said “more than 60,000 verified nodes” until 5 August 2026, which counted published pages rather than records.

The one referrer we caught supports that in a small, specific way. It landed on a barangay-level page, not the homepage. A model that cites this dataset cites the page that answered the question, which is the argument for publishing data as pages rather than as a product tour.

We did not build it to be cited. It was built to be correct, and being cited appears to be downstream of being correct in a machine-readable way. That is a hypothesis with a plausible mechanism behind it, which is a better starting position than a correlation with none. It is still a hypothesis.

What we are doing about it

Three things, none of them expensive, all of them running now.

Reconstruct the first query. Done, and it cost us the better version of the story. The contact remembered the shape of the question and not the string, which was enough to establish that the answer named our service brand rather than the data platform. Running that reconstructed query against the engines on a schedule, and publishing the result whichever way it goes, is the next step.

Capture discovery source on everything. Every inquiry form across every property we operate now asks how the buyer found us, referrers are logged at first touch rather than last, and the question gets asked on every sales call. This costs effectively nothing and compounds automatically. Every AI-sourced enquiry from here is a data point we actually recorded rather than one we reconstructed from memory.

Measure ourselves. martenfield.com went into the same tracking pipeline we sell, on day one, against its own query set: proptech AI visibility, GEO agency proptech, how do proptech companies get cited by ChatGPT, AI visibility audit commercial real estate.

We expect the early results to show we do not appear. We are going to publish them anyway, because a firm that publishes its own bad baseline and then the improvement is running the most credible case study available to it, and it is the one client whose permission we do not need.

The threshold

We will publish the running count of AI-sourced enquiries once it clears five, and we will publish it whether or not it supports the thesis.

The count has a definition now, which it did not have when this post first went up, and the missing definition is how a Google referral got into it. An enquiry counts on one of two conditions: a referrer header named an AI tool, or we asked the buyer and they named one. A guess offered without being asked does not count. A Google referral does not count, whatever anyone believes sent them.

There is a version of this business that grows by repeating the origin story with more confidence each time, and quietly rounds two different things up into one number while doing it. It would probably work for a while. It is not the one we are interested in running, because the entire product is telling companies the truth about numbers they would rather were higher, and you cannot sell that from a position of having flattered your own.

So: one traced enquiry, one traced lead, and two corrections. Ask us again in a quarter, and we will have either more of them or an explanation of why not.