# Six Months of Testing Contextual Intelligence: What the Numbers Actually Showed

Last week I wrote about [why we stopped trusting happy clients](/blog/six-months-testing-contextual-intelligence/) as proof that contextual intelligence works, and started testing it properly instead. I promised the real numbers this time, test by test. Here they are.

## The Clearest Result Came From a News Publisher in Asia

The clearest result came from a live news publisher in Asia, where we ran a controlled test back in January. Revenue on the previously suppressed and misclassified inventory came in **up 70% overall**. Existing buyers didn't just keep spending, they spent more once the content was properly classified, **up 70% on one SSP and 112% on another**. And it wasn't only existing demand paying more. **17 buyer networks** that had never bid on that inventory before showed up once it became readable to them. That's the part I still find the most convincing: demand that simply didn't exist for that inventory before, appearing the moment it could be understood.

## Eight Sites at Once, and Every One Moved the Same Way

More recently, we ran the same kind of test across eight sites at once for another publisher, holding back a small slice of traffic as a clean control while running enrichment on the rest. Every single site moved the same direction:

- Bid rate went up everywhere, in some cases by well over 50%.
- Win rate more than doubled on a few.
- Once you normalise for traffic volume and look at revenue per opportunity rather than raw totals, the blended lift on the previously misclassified inventory landed around 120%.

Zero sites went the other way. I didn't expect all eight sites to agree this cleanly.

## A Strong Result With Two Honest Wrinkles

A separate test, running across several demand partners side by side for a different publisher, told a similarly strong story, with a few honest wrinkles. Depending on the SSP, revenue on the affected inventory moved up somewhere **between 25% and 45%**. On the best-performing SSP, that gain didn't come evenly from everywhere, almost all of it traced back to one type of buyer: long tail and small business advertisers who simply hadn't been able to see what the page was about before.

Not every SSP in that test looked like that, and we'd rather say so than pretend otherwise.

### The One Where the Signal Wasn't Being Read

On one, the signal simply wasn't being read, revenue stayed essentially flat, no real change in either direction, until we tracked down a configuration issue and fixed it.

### The One Where Latency Ate the Gain

On another, early on, we actually saw revenue dip slightly negative, not because the technology backfired, but because of latency on how the signal was being delivered into the bid request, enough of a delay that some auctions timed out before the enrichment even arrived. Once that was resolved, that SSP fell back in line with the rest.

We're not pointing this out to lower the headline numbers. We're pointing it out because this is exactly the kind of thing that decides whether a result is real. A lot of what makes these tests trustworthy is unglamorous: checking the setup, catching a timeout, fixing a bad config, before trusting any number enough to publish it.

## Why Keyword Blocklists Keep Costing Publishers Money

It's worth explaining why this keeps happening, because it's not a fluke of one publisher's setup. Most brand safety systems still work off keyword blocklists, a list of words an algorithm scans for, with no understanding of what the page is actually saying. That approach can't tell a news report about a factory closing from a factory safety violation, or a review of a violent film from actual violence. So it does the safe thing from the platform's point of view, which is the expensive thing from the publisher's: block first, ask nothing.

> A study by ad verification firm CHEQ put that cost at roughly $2.8 billion a year for US publishers alone, inventory that was never unsafe, sitting unsold because a machine matched a word instead of reading a sentence.

## What Real Classification Changes

What real contextual classification changes isn't just the price a buyer pays for a page it can already see. It's two things at once, working together.

[Accurate IAB taxonomy classification](/solutions/bid-enrichment/) with built-in brand safety categories is what lets a demand system see what a page actually is, and say yes to inventory it would otherwise skip because a keyword matched. On the targeting side, that same understanding is what makes contextual advertising itself work, letting buyers find the right audience by what a page is actually about rather than a cookie, which is exactly why contextual has been picking up real momentum across the EU, Asia, and the US as privacy rules tighten and third party data gets harder to rely on.

**Buyers pay more for relevance when they can see it, and they show up more often when the content isn't wrongly flagged.** The numbers above are what happens when both are working at the same time, not one or the other.

## What We're Testing Next

We're continuing to test on more publishers and more markets, and we'll keep sharing what we find, including the SSPs that didn't move at all and the ones that needed fixing before they did. If your own numbers aren't showing this kind of lift yet, or an integration didn't work the way you expected, that's exactly the kind of thing we've already seen and already know how to fix. That's the real offer here: not a promise that it works everywhere instantly, but the experience of having found and solved this before.

[Talk to us](/contact/)