Intent signals for matchmaking and which ones are worth scoring

The intent term in the match score was worth 0.22 of the total. The function behind it returned 0.5 for every buyer who had not sent a meeting request, and 0.9 for everyone who had. That was the whole definition. Somebody had spent a morning arguing about whether 0.22 was too high, and nobody had spent ten minutes on what the number was measuring.

Intent signals for matchmaking are the part of a fit score most likely to appear on a slide and least likely to be defined anywhere in code. The weight is a policy argument that can run forever. What feeds the weight is a data question with an answer sitting in last edition's proposal log, and the answer takes an afternoon.

What actually counts as an intent signal here?

An intent signal is a recorded behaviour that changes the probability a proposed meeting gets accepted. That definition is deliberately narrow, and it rules out most of what gets called intent in a sales meeting.

Stated preferences are not intent signals under that test. A buyer who ticked three categories on a registration form has told you something about fit, and that belongs in the product and industry terms where it can be scored against the taxonomy. Intent is about the buyer moving towards a specific supplier or a specific category before anyone proposed anything.

The candidates worth putting on the list are small in number and all of them are things an event platform already writes to a table.

• Saved or shortlisted exhibitors. The buyer put a company on a list without being asked to.
• Exhibitor profile views. Weaker, higher volume, and contaminated by whatever the platform was promoting on its home screen that week.
• Category page views and search terms. Directional towards a category, silent about a company.
• Agenda bookmarks in a category. A session saved is a topic the buyer is spending show hours on.
• Sent meeting requests. The loudest of the lot, and the one that comes with a feedback problem of its own.

Each one arrives with a timestamp, which matters later. None of them is worth anything until you know what it does to acceptance.

Ranking candidate signals by lift

Lift is the ratio of the rate under the signal to the rate without conditioning on it. Provost and Fawcett set it out plainly in Data Science for Business in 2013 as how much more prevalent a pattern is than chance would give you, and it is the right frame here because the base rate is high enough that a raw percentage tells you almost nothing.

Take one edition. 4,120 proposals were shown to buyers and 1,566 were accepted, so the base acceptance rate is 1,566 divided by 4,120, which is 0.380.

Now condition on each signal as it stood at the moment of proposal.

Proposals where the buyer had already saved that exhibitor: 240 of them, 168 accepted. That is 0.700, and 0.700 divided by 0.380 is a lift of 1.84.

Proposals where the buyer had viewed that exhibitor's profile at least once: 1,340 of them, 592 accepted. That is 0.442, a lift of 1.16.

Proposals where the buyer had opened a marketing email naming the category: 2,010 of them, 800 accepted. That is 0.398, a lift of 1.05.

Three signals, three very different stories, and only one of them is large. Before believing any of it, put an error band on each. The standard error on a proportion of 0.70 across 240 observations is the square root of 0.70 times 0.30 divided by 240, which is the square root of 0.000875, which is 0.030. So the saved-exhibitor rate is 70 per cent give or take about 3 points, and the gap to the 38 per cent base rate is more than ten standard errors. That signal is real.

The email open runs the other way. Its standard error across 2,010 observations is the square root of 0.38 times 0.62 divided by 2,010, which is 0.011. The gap between 39.8 per cent and 38.0 per cent is 1.8 points, which is 1.6 standard errors, and 1.6 standard errors on a measure you chose after looking at the data is not a finding. Leave it out of the score and stop arguing about it.

Coverage decides which lift matters

A high lift on a rare signal moves very few meetings. This is where most intent scoring goes wrong, because the signal everyone gets excited about is the one that fires for 6 per cent of proposals.

The saved-exhibitor signal covers 240 of 4,120 proposals, which is 5.8 per cent. Its excess acceptance rate is 0.700 minus 0.380, or 0.320, so across those 240 proposals it is associated with about 77 accepted meetings above base. The profile-view signal covers 1,340 proposals, which is 32.5 per cent, with an excess rate of 0.062, so it is associated with about 83 meetings above base. The weak signal on the wide base does slightly more work than the strong signal on the narrow one.

That arithmetic should change what you build next. Raising coverage of the strong signal is worth more than adding a fourth weak signal, and coverage is a product decision: put the save control in front of buyers earlier, on the category page as well as the company page, and the 5.8 per cent moves.

It also tells you how to combine them. Feeding all three into the intent term as a sum lets the widest signal dominate by volume. Feeding only the strongest lets the term sit at its default for two thirds of proposals. A capped sum, where each signal contributes its own lift-derived amount and the term is bounded at 1.0, keeps both properties, and the cap is what stops one buyer's clicking habit from swamping the five sub-scores that make up the score, which is I1's subject.

What does third party intent data do that your own data cannot?

Third party intent is the category most likely to be bought before the first party version is measured, so it is worth being specific about what it contains.

Bombora's own published documentation, in its 2026 co-operative page and its Company Surge user guide, describes a co-operative of more than 5,500 B2B media sites contributed by over 200 publishers and brands, and scores an account's consumption of a topic against that same account's historical baseline, from 0 to 100, with 60 and above treated as a statistically significant spike. The design is careful and the baseline framing is the right one, because raw volume just finds large companies.

Two properties of that design decide where it fits in a match score. It is account level, so it cannot tell you which of the four people from that account is walking your floor on Tuesday. And it is topic level against a taxonomy that is not yours, so mapping surge topics onto your own category tree is a piece of work with its own error rate.

My view is that third party intent belongs in an account level term or as a tie-breaker between two otherwise equal candidates, and that using it inside a person level intent term makes the score harder to explain for very little gain. There is also a coverage problem nobody advertises. A 40-person distributor in northern Italy generates almost no consumption events in a co-operative of English language B2B publishers, so the signal is missing exactly where your international buyer mix needs it most, and missing signals are not neutral in a weighted score.

Decay, and the signal that was true in March

A saved exhibitor from fourteen months ago is a different fact from a saved exhibitor from Tuesday. Most implementations treat them identically because the query is easier to write that way.

An exponential decay is enough, and it needs one number: a half life. Set it at 30 days and a signal from 30 days ago counts half, one from 60 days ago counts a quarter, one from 90 days counts an eighth. Whether 30 is right is testable on the same file. Split the saved-exhibitor proposals by the age of the save at proposal time and recompute acceptance. If saves under 14 days old accept at 76 per cent and saves over 90 days old accept at 58 per cent against the 38 per cent base, decay is real and you can fit the half life to roughly match that curve.

Recency also protects you from a specific failure. Registration for most B2B shows opens months before the floor plan is final, so a buyer's early browsing reflects an exhibitor list that has since changed. Without decay, the score keeps promoting companies the buyer looked at when the alternatives were not yet listed.

Where this stops

Lift is an association, and the whole method is one step away from confusing a signal with a cause. A buyer who saves eight exhibitors is more engaged than a buyer who saves none, and that buyer would accept more meetings whether or not the save button existed. Some of the 1.84 is the signal and some of it is the person, and this file cannot separate them.

The honest way to find out is to hold back a random slice of proposals from the intent term entirely for one edition, then compare acceptance across the two groups. That is a real experiment, it costs a little acceptance in the holdout, and it is the only version of this that supports a causal claim.

The second limit is the loop. Once the intent term drives which proposals get shown, tomorrow's log records acceptance among the proposals that today's score chose, and the measured lift on any signal drifts upward for reasons that have nothing to do with buyers. That problem has a shape and a literature and it belongs to meeting requests treated as intent in I13.

Start this week with the proposal log. One row per proposal, the outcome, and one column per candidate signal as it stood at proposal time, which means reconstructing from timestamps and not from today's state. Compute the base rate, then the rate under each signal, then the coverage of each. Anything with a lift inside two standard errors comes out of the score, and the signals your registration and agenda data already hold in I12 are where the replacements come from, feeding the same matchmaking reporting your team already reads.