An estimate, and we say so

Built to be useful, not to look precise

The purpose of a wealth estimate in prospecting is to answer a triage question: is this household plausibly in the range where the conversation I want to have makes sense. That question is well served by a range and poorly served by false precision. A tool that reports $14,382,000 has not measured anything more finely than one that reports $10M to $20M, it has simply chosen to sound authoritative about the same underlying guess.

So the output is four ranges rather than one number, and the four are genuinely different questions. Net worth is the headline. Total assets is the gross picture before whatever is owed against it. Investible assets is the one advisors actually care about, since a large fraction of many balance sheets is a house nobody is going to liquidate. Income speaks to ongoing capacity rather than accumulated capacity, which is the more relevant figure for a recurring gift or a savings relationship.

Each is returned as a minimum and a maximum in dollars, and the schema enforces that. The model cannot answer with a single figure even if it wanted to.

• Net worth: the headline range, assets net of what is likely owed
• Total assets: the gross picture, including the property itself
• Investible assets: the portion that could plausibly be deployed
• Income: ongoing earning capacity rather than accumulated capacity

The whole input list

It is worth publishing the input list in full, because the honest way to judge an estimate is to know what it was allowed to see. Three groups of signals are assembled and passed to the model, and nothing else is.

From the linked parcel: the property address, the county's assessed value, the recorded owner, and the use type, read out of the cached property record. From residential records, where a lookup returned them: the person's current home address and up to five other properties they are recorded as owning. From the profile itself: name, title, city, state, country, organization name, domain, and website, LinkedIn URL, employment history, education history, and skills.

That is the entire input surface. There is no credit data, no tax return, no brokerage feed, no filing of holdings, and no purchased consumer wealth score. If a signal is not in the list above, it did not influence the estimate.

• Parcel: address, assessed value, recorded owner, use type
• Residential records: current home address, up to five other owned properties
• Person: title, organization, employment and education history, skills, location
• Nothing else, and specifically no account, holdings, or filing data

The most solid input, with a catch

Property is the input with the firmest footing, because it is a public record rather than an inference. Somebody's name is on a deed for a parcel a county has valued, and that establishes a floor: whatever else is true, this household controls an asset of roughly that size.

The catches are worth stating plainly. Assessed value is a tax figure on a county cycle and can drift from market value in either direction. Leverage is invisible, so an assessment tells you what the asset is worth and nothing about how much of it the household actually owns. And a single property is a weak sample of a balance sheet.

This is why the second property matters so much more than the first. When residential records show the same person recorded on additional properties, the estimate moves harder than the added value alone would justify, because multi-property ownership is one of the few observable signals that reliably separates accumulated wealth from a large mortgage. Up to five of those are passed in alongside the linked parcel.

What a career implies

Career data is the inferential half of the estimate and it is where the reasoning actually happens. A title carries information: the compensation structure of a managing director, a founder, a partner, or a surgeon differs in kind and not just in degree from a salaried role, and equity-heavy industries accumulate differently from ones that pay in cash.

Duration matters as much as level, because wealth is cumulative. Twenty years of senior earnings is a fundamentally different balance sheet from the second year of a large salary, even where the current number is identical, and employment history is passed in whole rather than as a current title so that trajectory is visible.

The honest weaknesses live here too. A retired executive with no current employer looks thin on paper and is often the wealthiest person on the street. A founder of a private company has an asset nobody can see the size of. Inherited wealth leaves almost no career trace at all. The model is instructed to stay conservative, which means these cases tend to be underestimated rather than overestimated, and that is the preferable direction for a triage tool: a prospect you look at more closely than necessary costs you a few minutes, while one you dismiss wrongly costs you the relationship.

One structured call, after the profile exists

Mechanically, this is a single structured generation. The assembled context goes to Gemini 3.1 Pro through OpenRouter with medium reasoning effort, and the response is forced to match a schema of four objects each carrying a minimum and a maximum in dollars. The instruction asks explicitly for conservative ranges. Because the schema is enforced at the call, there is no parsing step where a malformed answer could be quietly reinterpreted.

It runs asynchronously, after the profile has been written, which is why a new profile appears with an estimating state before its number lands. The prospect list polls while an estimate is pending and fills the chip in when it completes. Each profile carries an explicit status of pending, completed, or failed, and a timestamp for when its estimate was produced, so a stale or missing estimate is visible as such rather than reading as a zero.

The estimate is produced once, at profile creation. It is a snapshot of what was knowable then, not a subscription that revalues quarterly, and the timestamp on the profile is the honest way to tell how old it is.

The width of the range is information

A range is not a hedge, it is a reading. When the bounds sit close together, the underlying signals corroborated each other: a clear title at an identifiable organization, a long employment history, a well-assessed property, maybe a second one. When the bounds are far apart, the signals were thin or in tension, and the right response is to look at the evidence rather than the number.

Two other markers on the same profile help calibrate. The confidence score, from 0 to 100, describes how well the identification itself held up, along with a count of how many independent sources agreed. That score is about who this person is, not about how much they have, and it deserves reading first: an estimate attached to a shaky identification is not a wealth question at all, it is an identity question.

The written research summary is the third piece. It states how far the research run actually got, and where the profile was assembled from fewer sources it says so. Taken together, the three tell you whether you are looking at a well-evidenced household or a plausible sketch, which is a distinction most wealth screening products deliberately blur.

• Narrow range: corroborating signals across property, career, and records
• Wide range: thin or conflicting inputs, worth opening the evidence
• Confidence score and source count: about the identification, not the wealth
• Research summary: states how far the run got and what it could not confirm

On the list, the profile, and the export

In the prospect list each row carries a net worth range chip, or an animated estimating state while one is being produced, which makes the list scannable in the way that actually matters when you are deciding whose morning you are going to occupy. The list can also be filtered by estimate status, so you can separate the profiles that have a completed estimate from the ones still pending or failed.

On the profile itself, all four ranges appear together in the wealth section, above the contact details and beside the research summary that explains where the underlying facts came from. On the map, a parcel linked to a saved profile shows the same figures in its drawer, which is what turns the map from a lookup surface into a view of the ground you have already worked.

In a CSV export, the wealth estimates column group carries the net worth and income ranges alongside the rest of the profile, so the figure survives the handoff into whatever system your team actually runs the pipeline in.