Relevance vs Click Prediction in Ecommerce Search
Understand relevance, predicted clicks and commercial objectives in ecommerce search. Compare ranking choices with clear constraints and testable outcomes.
Relevance and click prediction in ecommerce search
Search relevance and predicted click-through rate answer different questions. Relevance concerns whether a product satisfies the request; pCTR estimates the chance of a click in a particular context. Both can inform a ranking system, alongside purchase value and business constraints. Evaluate how they work together instead of treating them as competing architectures.
Separate the layers of the decision
Retrieval produces candidates. Eligibility controls decide which candidates are allowed for the request. Ranking chooses their order using the configured evidence and objectives. A transformer model may contribute to query understanding or retrieval while behavioural predictions contribute to ranking within the same system.
Adaptive Transformer Search and pCTR therefore describe different things. The first is a product name; the second is a model target. An ATS-versus-pCTR headline would suggest a technical choice that a buyer does not necessarily face.
Ask what the prediction represents
Predicted click-through rate should be interpreted with its context: the shopper, query, product, position and exposure assumptions used by the model. It is not the same as observed CTR, relevance, purchase probability or margin.
Clicks provide useful feedback, but presentation affects who sees and clicks a result. Primary learning-to-rank research explains this bias and methods for addressing it. Ask suppliers how they handle exposure, position effects and sparse evidence; do not assume a particular supplier ignores them.
Align the objective with the business question
Consider a simplified, entirely synthetic example. Item A has a 20% click probability, a 1% purchase probability after a click and £40 contribution per purchase. Its expected contribution per opportunity is £0.08. Item B has a 10% click probability, a 5% purchase probability after a click and the same contribution: £0.20.
The example shows why maximising clicks alone can differ from maximising expected contribution. It omits returns, interactions between items and estimation uncertainty, and it is not PA’s production ranking formula. Mandatory relevance and eligibility requirements should be satisfied before comparing those values.
Compare the configured experience
Use an offline relevance panel to inspect exact matches, constraints and severe errors. Use live experiments to estimate the chosen outcome, with guardrails for conversion, availability and shopper experience. Check new and infrequently seen products separately so an aggregate result does not conceal uneven performance.
PA’s Margin Management positioning places economic objectives after policy, eligibility and relevance. Ask the team to demonstrate the signals, controls and measurement applicable to your proposed implementation.
The buying decision is whether the system serves suitable products and improves the agreed outcome under your constraints. A label such as semantic, behavioural or AI does not answer that decision by itself.
Illustrative example
{"label":"Synthetic, not a vendor algorithm","currency":"GBP","A":{"p_click":0.2,"p_purchase_given_click":0.01,"contribution_per_purchase":40,"expected_contribution_per_opportunity":0.08},"B":{"p_click":0.1,"p_purchase_given_click":0.05,"contribution_per_purchase":40,"expected_contribution_per_opportunity":0.2}}
Sources
- Unbiased Learning-to-Rank with Biased Feedback, checked 2026-10-25
- Margin Management, checked 2026-10-25
- Constructor Ecommerce Search and Autosuggest, checked 2026-10-25
How Particular Audience does it
Frequently asked questions
Is pCTR the same as relevance?
No. It predicts a click in a specified context. A click can be influenced by visibility, imagery, familiarity and presentation as well as whether the product satisfies the query.
Does transformer search exclude behavioural ranking?
No. Retrieval architecture and ranking signals are separate design choices. A system can use transformer-derived representations and behavioural predictions together.
Should every retailer optimise contribution rather than clicks?
The objective depends on the decision and available data. Define the desired outcome and constraints, validate the inputs, and measure the trade-offs. The simplified example illustrates the distinction; it does not prescribe a universal ranking formula.
Related guides
- Vector search vs keyword search for ecommerce: Compare vector, keyword and hybrid ecommerce search. Understand exact matches, semantic recall, filters and ranking, then evaluate them on your own catalogue.