Talks

The ideas behind the agent, argued in public before they shipped.

Berlin Buzzwords and MICES are where the search community tests its thinking. Four talks from the searchHub team trace how a tracking SDK became a learning loop: collect the whole journey, measure what converts, explore instead of exploit, and test at the level of the single query.

Watch the latest talk

Four talks, one loop

2018 Collect the whole discovery journey. Detect 2019 Measure what converts, not what looks relevant. Detect 2023 Explore, instead of exploiting what already sells. Excel 2024 Test every answer at the level of the query. Excel

Detect the opportunity. Apply the improvement. Measure the outcome. Learn from the result.

Four talks. Newest first.

Recordings from the community stage — no sales pitch, just the method.

03 Excel Berlin Buzzwords 2024 · Andreas Wagner

Improving search at scale with efficient query experimentation

Most teams judge a search change by relevance labels or one site-wide A/B test. Neither tells you whether a single query got better. This talk lays out how to experiment at the level of the individual intent: which metric to trust, what to randomize on, and how to decide when the data is sparse.

It is the method that now runs unattended as autoLoop: two answers to the same intent, tested live, the winner kept.

Where it lives today: reversing diminishing returns

03 Excel MICES 2023 · Andreas Wagner

Towards data-driven inspiration instead of exploitation

Optimize only on what already sells and your data narrows into an echo chamber: the ranking confirms itself, and shoppers never see what they might have bought. The talk makes the case for deliberate exploration — challenging the current result set on live traffic — and for retail knowledge alongside the numbers.

This is why searchHub challenges the incumbent answer instead of locking it in — the reason gains compound rather than plateau.

Where it lives today: reversing diminishing returns Slides

01 Detect MICES 2019 · Andreas Wagner

Measuring and optimizing findability in e-commerce search

Relevance judgements feel objective, but they carry biases and rarely predict engagement — let alone revenue. Drawing on searchHub's interaction logs, the talk shows why "relevant" results underperform, why diversity earns money, and how findability and sellability combine into one measure you can actually optimize.

It is the reasoning behind ranking by what shoppers buy, not by how closely catalog text matches the words typed.

Where it lives today: eliminating guesswork Slides (PDF)

01 Detect MICES 2018 · Pavel Penchev

searchCollector: the data a self-learning search actually needs

A self-learning system is only as good as the data it gets — and web analytics never captured the search journey: sampled sessions, no reciprocal rank, no intent. Pavel introduced searchCollector, an open-source SDK that records the full discovery trail, query to purchase, as first-party data you own.

Everything Detect does today — causal analytics across the whole journey — starts with this talk.

Where it lives today: eliminating guesswork Slides

The thread

Six years of talks. One conclusion.

A search engine is not a self-improving system. It ranks what it is given, as well as it was configured. The improvement has to come from a layer that sees the whole journey, tests its own decisions and keeps what works — on the stack you already run.

The full argumentWhy searchHub — what one agent adds to every engine. Hear it liveTrade fairs and meetups where the team speaks next. See it on your own traffic — book a demo