Measuring paid search for trade shows beyond the platform conversion count

The search platform says the campaign produced 2,039 registrations. The registration file for the same period holds 1,610 rows that can be traced to any paid search click at all. Somebody has to take one of those numbers to a budget meeting.

Choosing between them is the smaller problem. The larger one is that the single figure, whichever you pick, is an average across three activities that have almost nothing in common except the auction they run in.

Three spend groups, split at campaign level

Paid search for trade shows contains at least three distinct businesses.

Brand defence buys your own show name, its abbreviations, its misspellings and the phrases people use when they already know you exist. The traffic is people who were coming anyway, most of the time.

Category demand buys the sector terms, product terms and problem terms a buyer might search without knowing your show exists. This is the only group that can genuinely find new audience.

Competitor terms buys the names of rival shows. The traffic knows the category and has already chosen somebody else.

Split them into separate campaigns rather than separate ad groups or keyword labels. Terms migrate, match types broaden, and a reporting split that depends on somebody maintaining a keyword label will be wrong within an edition. A campaign level split survives.

Anything that cannot be classified goes into a fourth group named unclassified, and the size of that group is a data quality measure in its own right. If it exceeds a few per cent of spend, the account structure is the problem to fix before any of the analysis below is worth running.

What does the arithmetic look like per group?

Report each group against registrations in the file, joined on the click identifier captured at the form, not against the conversion count the platform reports. Three groups, one edition.

Brand defence cost 42,000 and produced 1,050 registrations at 40.00 each. Category demand cost 118,000 for 885 registrations at 133.33 each. Competitor terms cost 26,000 for 104 registrations at 250.00 each. Click to registration rates fall the same way: 5 per cent on brand, 3 per cent on category, 2 per cent on competitor terms.

Presented as one line, the account shows 186,000 of spend, 2,039 registrations and a cost per registration of 91.22. That blended figure describes none of the three groups and hides the fact that the cheapest group is also the one least likely to be adding anything.

The brand defence problem, which the evidence is unusually clear about

Blake, Nosko and Tadelis ran large scale field experiments at eBay and published them in Econometrica in 2015. Their headline result is that brand keyword ads had no measurable short term benefit, because the paid click had a near perfect substitute sitting immediately below it in the organic results. For non brand keywords they found new and infrequent users were positively influenced, while frequent users, whose behaviour the ads did not change, accounted for most of the advertising expense, so average returns were negative.

Apply that to the table. If 90 per cent of the 1,050 brand registrations would have arrived through the organic listing anyway, the incremental count is 105 and the incremental cost per registration is 42,000 divided by 105, which is 400. The cheapest line in the account becomes the most expensive one, and nothing in the platform report would ever have told you.

The counter argument is real and also evidence based. Simonov, Nosko and Rao, in Marketing Science in 2018, ran experiments on Bing across thousands of brands and found a positive causal effect of focal brand ads of 1 to 4 per cent when no competitor was present, with larger brands seeing smaller effects. For brands that faced competition on their own brand terms and chose not to advertise, competitors took 18 to 42 per cent of the clicks.

So brand defence is close to worthless in an empty auction and materially valuable in a contested one. Whether brand search works in general is the wrong unit of analysis. The answerable question is whether anybody is bidding on your show name this month. That is checkable in an afternoon with an auction insights report, and it should be rechecked every edition, because a competing show launching in your category changes the answer.

What you do about it beyond checking is an experiment rather than a report, and that experiment has its own design problems which belong to it and not here.

What the platform conversion count is doing differently

The 2,039 the platform reported and the 2,039 registrations implied by the table are not the same population, and I have written them identically on purpose because that is how the confusion starts.

The platform counts a conversion when its tag fires inside its window, on its rules. That includes modelled conversions where consent was declined, view through credit where the rules allow it, and conversions attributed on a window that may not match yours. It has no visibility of your registration table, so it cannot know that 214 of those form completions were duplicates, staff, or people who abandoned payment on a paid pass.

Your file counts rows in a table. It is auditable and it reconciles to the badge count.

For the cost per registration figures above, the file is the correct denominator, because the finance question is what the show spent to produce the audience it actually had. The platform figure keeps its job as an optimisation signal, since the bidding algorithm needs a conversion signal fast enough to act on and your file cannot supply one on a two hour cycle. Both numbers survive. They report to different people, and walking one down to the other line by line is the exercise that stops the argument recurring every quarter.

Match types leak the split you just built

The three group split holds only as long as the terms stay in the campaign they were put in, and broad match is designed to stop that happening.

The mechanism is ordinary. A category campaign bidding on a sector term serves against a query containing your show name, the click gets billed to category demand, and a registration that was going to happen anyway is now credited to the group you are trying to judge on new audience. It runs the other way too, with a brand campaign picking up generic queries that contain a word from the show title.

The check is the search terms report, read at the query level rather than the keyword level, at least twice per campaign. Count the spend in the category campaigns that served against queries containing the show name, and either exclude those queries from the category campaign or accept the leakage and report it. On accounts where nobody has looked, that figure is often 10 to 20 per cent of category spend, which is enough to move the cost per registration comparison that the whole three group split exists to support.

Negative keyword lists maintained per group are the standard fix and they need an owner and a review date. A list built once at launch and never revisited will be wrong by the second edition, because the queries move even when your keywords do not.

Why does category demand get cut first?

The group most likely to be cut in a bad year is category demand, because it has the worst cost per registration in the table. That decision is usually wrong.

Category terms are the only paid search group that reaches somebody who does not know the show exists. At 133.33 per registration it looks poor against brand defence at 40.00, and it looks reasonable against the fully loaded cost of finding an equivalent new registrant through any other prospecting channel.

The number worth adding to the table for category demand is the first time registrant share. If 68 per cent of the 885 category registrations are people with no prior registration in the portfolio, that is 602 new people, and the cost per new registrant is 118,000 over 602, which is 196.01. Do the same for brand defence, where the first time share is usually far lower, and the comparison inverts again.

Competitor terms at 250.00 per registration are harder to defend on cost alone. They earn their place when a rival show cancels, moves, or changes dates, at which point the term list you maintained all year is the fastest audience acquisition asset you have. Keep the group, keep it small, and judge it on the two months where it matters rather than on an annual average.

Where this stops

Every figure above rests on the join between a paid click and a registration row. If the click identifier does not reach the registration record, the registration lands in the file as direct or organic and the paid group is undercounted, which makes cost per registration look worse for exactly the campaigns you are trying to judge.

Cost per registration also says nothing about who registered. A category campaign bringing students at 90 each and a brand campaign bringing specifiers with signing authority at 200 each are not comparable, and the ratio alone will push you towards the cheaper and worse of the two.

The deepest limit is that none of this measures incrementality. A cost per registration is an accounting statement about registrations that happened next to some spend. The experiments cited above exist because the accounting statement and the causal question give different answers often enough to matter. Where paid social differs in this respect, because the placement and the objective change the price, is a separate matter, and the wider picture sits on the acquisition and attribution pillar.

Split your search account into the three campaign groups this week, if it is not split already, and report the last edition's spend and file matched registrations for each. The first thing to look at is the click to registration rate on brand terms against category terms, because the gap between them is the size of the argument you are about to have.