One question that sorts

Attribution says where a lead came from. A signal says what it’s likely to be worth, early enough to do something about it. In a power law book, the earliest trait that predicts the above-average accounts is worth more than any other field on the form.

What makes a good one:

• It’s about the business, never the person.
• It shows up on the first form, in time to change the bidding, the lead score and who calls.
• It’s anchored to value in your book, not someone else’s. Prove that before acting on it.
• It costs the wrong lead more to give. A declared monthly sales band takes seconds for anyone, but a booked call or a document upload takes effort a tire-kicker won’t spend.
• The platforms allow it, and the law permits it.

Candidates worth testing are the size of the business (monthly sales in bands, number of locations, number of employees), its age (time in business, in bands), its presence (does it have a website already), intent (the amount or plan asked for, when they want to start), the channel itself (called, or filled in a form) and effort (finished the second step, uploaded documents). Choosing which one is the art to the science. Proving it works is the science.

Where a signal shows up decides what it can change

A signal can only change the stages from the point where it’s found onward, and an optimization built on it works at the stage where it’s applied. So the same fact about a business has a different impact depending on where in the funnel it turns up, and what gets optimized around it.

Where the signal is found	What it can change	What it can’t
On the form (conversion and MQL)	The bidding, the lead score, who calls and how fast, and every stage after	Nothing before the click
On the first call (SAL and SQL)	Routing, qualification, the offer, underwriting	The ads that already bought the lead
In underwriting (RQO)	Approval, pricing, terms	Who was acquired, or which rep worked them
In the revenue (closed won onward)	Retention, pricing, service	Everything before the sale, unless it goes back to the platforms as a later conversion

That’s why the sorting question goes on the first form. Found there, it can steer the bidding, and the bidding decides who shows up in the first place. The same answer collected by a rep on the second call can still send the lead to the right rep, but the money that bought the click is already spent.

The later a signal shows up, the closer it sits to the result, and the later and rarer it arrives. A funded account is the truest signal there is, and it comes weeks after the click, at a fraction of the volume. A form answer predicts less, and it arrives with every lead, the minute the form is sent. Choosing where to add a signal trades how much of the funnel it can move against how surely it predicts the result. Moving the bidding later in the funnel as the volume allows is the same trade, made one stage at a time.

Read every optimization from the stage where it’s applied all the way to revenue. Above that stage it shows only its cost. A step that sorts at the form lowers the form’s conversion rate, and the gain shows up in the stages below it.

Store each answer as an ordered code

The sorting question is the form’s self-assessment question, and it gets an ID of its own, its SAID, the way the form has its WFID. Those two IDs are required on every online conversion. Give each answer band a code that sorts in order, with each band adding one more 1:

Average monthly sales	Option value	Code
Under $10,000	under_10k	0000
$10,000 to $30,000	10k_to_30k	1000
$30,000 to $100,000	30k_to_100k	1100
Over $100,000	over_100k	1110

Then the line is one rule. “Starts with 11” means $30,000 a month and up. Moving the line to $10,000 is a change from 11 to 1, in one file, with nothing else touched. The answer as the visitor gave it goes into the field sales reads, and the code goes into its own field beside it, so a change to the question’s wording never breaks the rule. When the question is asked differently, it gets a new SAID with its own codes, and the map keyed on the SAID translates both into the same field.

Match on the option’s value, never the words the visitor sees. A label edited for wording, with a different dash or a currency symbol, can’t break the rule, and an answer that matches no code is reported as unknown instead of quietly failing to qualify.

The rule lives in one place, a small rules file the server reads (and the page too, where it needs to know). That’s the “one rule, in one file” in the next figure.

Where the line goes

Set the line where the bidding still gets enough to learn from. Aim it at the middle-high bands, not the very top. A line set at the top band starves the bidding of everything, including the conversions it needs to learn. The working minimum is 30 qualified conversions per campaign in 30 days. Plan for no fewer than 45, so a slow stretch doesn’t take a campaign under it, and aim for 60 a month, double the minimum. Meta’s learning phase looks for about 50 optimization events per ad set in a week, so a campaign at these levels can stay in learning longer, and run this way, campaigns have held up in practice. On Google, target ROAS on Search typically needs at least 15 conversions in the past 30 days. If the qualified conversion fires too rarely for that, consolidate first, into fewer campaigns and ad sets that each clear the minimum. Then move the line down one band, or use an earlier signal. Don’t go back to counting every form fill.

Once the campaigns are consolidated, move the line by rule, not by feel. Every lead keeps its answer code, so you can count what each line would have produced over the last 30 days. Once a month, set the line at the highest band that keeps every campaign at 45 or more qualified conversions in 30 days, and leave it there until the next month. A slow season moves it down a band, and the busy months move it back up. Log each change in the signal register, with the date, the old line, the new one and the counts behind it. A line that moves every week changes what the platform learns from every week.

What’s off-limits

Some traits would predict value and still can’t be used, by platform rule and, for protected traits, by law. Don’t use personal financial details (credit score, income, debt, net worth), health, race or ethnicity, religion, sexual orientation, political views, union membership, or any stand-in for a protected trait, no matter how well it predicts.

The platforms have said so in writing. Meta began rolling out “more proactive restrictions on custom conversions” in September 2025, blocking ones that suggest a health condition or financial status from running campaigns. Google’s customer data policy says advertisers “may not upload conversion information related to sensitive categories,” including negative financial status. On Meta, US ads for credit, housing, employment and financial products and services run in Special Ad Categories with their own targeting limits, and the financial category covers credit cards, financing, bank accounts, investment and insurance services and related financial offers, so ads for business loans fall in it and ads for payment services can too. The category comes from what the ad promotes, so no event name keeps a campaign out of it. Google limits targeting for housing, employment and consumer-finance ads in the US and Canada. And fair-lending and fair-housing law still applies, whatever a platform permits.

There’s a lesson in that. The platforms forbid optimizing on certain signals because optimizing on a signal works. If sorting didn’t change who the algorithm finds, nobody would need to ban it.

Business size is a fact about the business, not the person. Even so, check how the platform classifies your account before you rely on any signal, and check again when the platform updates its rules.

The answer never leaves your systems

The platforms never see the answers or the codes. No answers, codes or thresholds go in the conversion, the page address or the server call, and the tag manager’s data layer carries neither, because every tag in the container can read it. The conversion the platforms bid on says one thing: a qualified conversion happened. The interactive figure above shows why that matters. Put the code in the thank-you page’s address, and it goes to every tag on the page, on every page load.

The thank-you pages are for the visitor

Two thank-you pages still make sense: one for the leads that clear the line, with the next step they’ve earned, and one for everyone else. The server picks the page, and the page tells the visitor what happens next. What the page shouldn’t do is carry the conversion. Give both pages neutral addresses, and let the events the server’s answer triggers do the counting.

There are four reasons, and the first is the one most setups miss:

• The address doesn’t always reach Meta. On pixels for banking, lending, financial services and insurance, Meta’s Core Setup can restrict “custom parameters and parts of URLs after the domain.” A custom conversion on a code in the address, or on the qualified page’s path, then matches an address Meta never receives. Where a setup like that works, it’s because the tag manager reads the page in the browser and fires an event, which is the event-based build under another name.
• Meta flags custom conversions that suggest financial status. A flagged custom conversion can’t be used in new campaigns and stops receiving data. A rule written on an answer band invites the flag. The standard Lead event, fired on the server’s answer, needs no custom conversion.
• Google receives the full address, and a conversion that fires on a page load has no transaction ID to deduplicate with.
• A page load isn’t a lead. Reloads, the back button and a direct visit to the qualified page all load it again. The events fire once, for one saved lead, with its event ID.

Every source reaches the same events. The landing pages can differ by source, and so can the thank-you pages’ copy. The events can’t. A lead from Google fires Meta’s Inquiry, and its Lead when it qualifies, the same as a lead from Meta. That isn’t credit. PRO decides the source, and Meta’s reporting is a separate count: Ads Manager credits Meta with any conversion from someone who clicked a Meta ad in the last 7 days or saw or engaged with one in the last day, its default attribution setting, so a Google lead can show up there as Meta’s too. Meta’s own guidance says “you can use only matched events for ads attribution and ad delivery optimization,” so what lets an event teach the bidding is a match to a person, not a click on a Meta ad. Every qualified lead, from every source, is one more example of the person the bidding should go looking for. It keeps the counts honest too: Meta’s Lead total in Events Manager should equal every qualified submission from a visitor who didn’t opt out, whatever the source. The results Ads Manager credits to Meta’s ads are a smaller number, set by the attribution window, and they don’t tie to either count.