First party intent at events and where the signal comes from

Somebody on the commercial team asks where the intent number comes from, and the honest answer for most shows is a vendor invoice. An intent data subscription was bought, a CSV arrives weekly, and it is scored against companies that may or may not have anyone registered.

First party intent at events is sitting in the organiser's own tables the whole time, unread. Every agenda bookmark, every exhibitor page view, every buyer who came to the last three editions and scanned into the same hall each year. The work is not acquisition. It is joining what you already write down to a person record and deciding what each row is worth.

What first party data does an organiser actually hold?

More than almost any other party in a B2B purchase. The organiser sees the registration form, the agenda builder, the exhibitor directory, the search box, the badge scans, and five years of the same people doing all of it at previous editions.

Named by the table they live in, the signals worth joining are these.

• Registration answers. Declared categories, job title, company, buying role. Stated, not observed, and the trade in asking for more belongs to I10.
• Agenda and session bookmarks. Written by the schedule builder, timestamped, and mappable to categories through the session tagging you already maintain for the programme.
• Exhibitor directory events. Profile views, saved companies, brochure downloads, contact reveals.
• On-site and pre-show search. The strings buyers type are the least processed statement of intent in the building.
• Prior edition badge scans. Which halls, which stands, which categories, in which year.

CEIR's 2024 report on digital tactics, show services and DIY marketing, reported by Trade Show Executive in August 2024, found that 70 per cent of exhibitors use digital tactics offered by organisers in their attendee engagement, up ten points on the 2017 study. Those tactics run on organiser platforms, which means the behavioural exhaust from exhibitor marketing also lands in organiser tables. The organiser is the only party with the whole picture, and most of them are buying a partial one from outside.

Session bookmarks are a category vote

Bookmarks are the cleanest first party signal because a buyer spends nothing to view a page and spends a decision to save a session into a schedule they intend to keep.

Take a buyer at a food logistics show who has saved seven sessions. Three of them are tagged cold chain logistics, two are tagged warehouse automation, one is tagged customs and compliance, and one is a keynote with no category tag. Drop the untagged one. Six categorised bookmarks, so the shares are 3 divided by 6 for cold chain, which is 0.50, then 0.33 for warehouse automation and 0.17 for customs.

Now compare with a second buyer who saved fourteen sessions across nine categories, three of them cold chain. Their cold chain share is 3 divided by 14, which is 0.21, on more evidence. The first buyer is more focused, the second is more active, and a raw count would rank them equal on cold chain while a raw share would rank the first buyer nearly two and a half times higher.

The fix is to carry both numbers into the score and let the volume decide how much the share is trusted. Multiply the share by a confidence factor of n divided by n plus k, where n is the buyer's total bookmarks and k is a constant set from your own file. With k at 5, the first buyer's cold chain component is 0.50 times 6 divided by 11, which is 0.27. The second buyer's is 0.21 times 14 divided by 19, which is 0.15. The focused buyer still wins, by a margin that reflects how little either of them has told you.

Search strings deserve their own treatment in the same section, because they are the cheapest signal to collect and the one most organisers throw away. A buyer who types blast chiller into your directory search has told you a category in three words, without a taxonomy picker or a form. Log the query, the timestamp and whether it produced a click, then map the string to a category with the same mapper that handles exhibitor profiles. Queries that produce no click are the most useful rows in the table, since they name a demand your floor did not satisfy, and a category with 240 searches and 11 exhibitors is a sales brief as much as a scoring input.

Why is a missing signal not a zero?

Because you did not observe a preference against the category. You observed nothing, and those are different states that most scoring code collapses into the same number.

Hu, Koren and Volinsky made the distinction the basis of their treatment of implicit feedback at ICDM in 2008: implicit data has no genuine negative examples, so the right model separates the binary preference from a confidence in that preference, with confidence rising as the observed activity rises. A buyer who never opened the agenda builder is not a buyer with zero interest in cold chain logistics. They are a buyer you know nothing about, and the score should say so.

Practically that means the intent term needs a neutral default and a confidence weight, and the neutral default is not zero. Set the unobserved case at the population mean for that category and let the weighted score fall back on the terms you do have, which are product fit and industry fit from declared data. A buyer who registered eleven days before the show, went straight to the badge collection page and never touched the agenda should be ranked on what they declared. Scoring them at zero intent pushes them below buyers who idly browsed, and your hosted buyer programme is full of the former.

Returning visitors carry the most information

The signal nobody uses is the one with the longest history. A buyer attending their fourth consecutive edition has scanned into halls, sat in sessions and had meetings, and all of that is in the warehouse under a person identifier that may or may not be the same one they registered with this year.

Prior edition scans are strong because they cost the buyer real time. Take a returning buyer whose last edition scan history shows 11 stand scans, 6 of them in cold chain logistics companies. That is a 0.55 share, on evidence that is a year old and worth decaying, but it is evidence about behaviour on a floor rather than clicks on a page.

Blend it explicitly instead of quietly. Set this year's category intent as 0.6 times the bookmark component plus 0.4 times the prior edition scan component, and write the two weights in the config where somebody can argue with them. For the focused buyer above, if their prior edition cold chain share was 0.55 decayed to 0.44, the blended figure is 0.6 times 0.27 plus 0.4 times 0.44, which is 0.162 plus 0.176, which is 0.338.

The blend also gives you an answer for the first-time buyer, which is that the second term drops out and the first is renormalised to carry the full weight. That behaviour needs stating in the config rather than emerging by accident from a null.

The join is the hard part, and it is not a modelling problem

Every signal above is attached to whatever identifier the platform issuing it uses. The agenda builder knows a login. The directory knows a session cookie until the buyer signs in. The badge scan from 2024 knows a registration record from a system you have since replaced.

Two practical rules keep this from becoming a year of work. Join only on identifiers you can defend, which for most organisers means the registration identity plus a verified email, and record the join method on every row so a low confidence join can be excluded later. And accept coverage below 100 per cent as normal. If 62 per cent of this edition's buyers link to a prior edition record, that is the population your returning-visitor term applies to, and the other 38 per cent get the first-time treatment.

Coverage is worth reporting next to the score itself, because a match score built on a 62 per cent join is a different object from one built on a 95 per cent join, and only one of them should be described to exhibitors as behavioural.

Where this stops

First party intent is bounded by your own surfaces. A buyer researching six suppliers on their phone, reading two of them on your directory and four on Google, gives you a third of a picture and you have no way to know it is a third. The signal is unbiased about your platform and silent about everything else.

There is a second limit that shows up as soon as the score goes live. Bookmarks and views are collected under whatever consent your registration flow captured, and using them to rank commercial meeting proposals is a purpose your privacy notice either covers or does not. That question has a real answer per jurisdiction and it is worth getting before the score is built rather than after an exhibitor asks how you knew.

Start this week by counting coverage, which takes one query per source. For last edition, what share of buyers who received a proposal had at least one agenda bookmark, at least one directory event, and at least one prior edition record. If the bookmark number is under a fifth, the fastest improvement available is putting the save control where buyers already are, ahead of any modelling work, and that is the interest capture question I10 owns. Then rank whatever you do collect by acceptance lift as in I11, keep the request signal from I13 out of the first pass, and report the whole thing next to the coverage figure in your matchmaking reporting.