They searched for something specific. You showed bestsellers.

Declared

The query, filters, and spelling variants.

Connected

Past purchases or category affinity that should re-weight results.

Example eligibility rule

Re-rank when the query contains a recognized product, job, or attribute and produces at least 12 relevant results. Use purchase history only for authenticated customers with at least two prior category interactions. Keep default ranking for ambiguous, navigational, or very broad queries.

Query language is specific (a size, a job, a competitor) but merchandising still sorts by bestsellers. The customer has to filter their way back to what they asked for.

Avoid this assumption

Better search is personalization. You do not need a "personalization platform" to rank the right 12 products first.

Search is one of the clearest places customers state their intent. Acting on that signal is more useful than adding personalized decoration elsewhere.

Reach

How many eligible people can you reach in the chosen channel and time period?

Cost

What will content, delivery, QA and ongoing maintenance cost?

Return

What incremental outcome would cover that cost, and what is it worth?

Planning inputs, not measured results. Estimate reach × incremental outcome rate × value per outcome, then subtract cost. Use a range and a consistent time period.

Example format: Recommendations

Re-rank the first page around the query and known affinity.

Compare lighter and deeper responses

Choose the lightest response that solves the need.

Assist: Offer useful help

Keep the current layout, but explain the ranking: "Because you searched for waterproof hiking shoes, waterproof trail options appear first."

Adapt: Change a meaningful part

Re-rank the first page around the query and known affinity.

Transform: Tailor the journey

A guided results template: matching items, a missing-filter prompt, and a related-need module.

What to watch out for

• Overfitting to last click and creating a filter bubble.
• No holdout, so you never know if ranking actually paid.
• Ignoring null searches, which are often the highest-intent misses.

Control

Default ranking and merchandising.

Treatment

Intent-weighted ranking for the same query.

Success measure

Search-to-product (or article) engagement, then conversion.

Guardrail

Zero-result rate and filter usage. Spikes mean you hid too much.

Win or lose, record what worked, for whom and why. What you learn becomes the next signal.

• Bloomreach Ecommerce Search AI-driven ecommerce search and discovery platform (Loomi AI) that personalizes ranking, recommendations, and merchandising from first-party shopper data.
• Algolia Hosted search infrastructure with rule-based merchandising and an opt-in machine-learning personalization layer.
• Adobe Target Uses machine learning (a Random Forest model plus a multi-armed bandit) to automatically select the best-performing, marketer-defined experience for each visitor, rebuilding its model roughly every 24 hours.
• Kameleoon Audiences Experimentation platform with an AI Predictive Targeting feature that scores each visitor's likelihood to convert (0-100) and lets you target or exclude by that score.
• Algonomy Retail personalization suite (search, recs, dynamic listing pages, content/styling, social proof, shopping agent) built on the earlier RichRelevance/Manthan merger. Now operating under the ADA brand after ADA acquired Algonomy in July 2026.
• Athos Commerce Athos Commerce is the January 2025 holding brand created when Searchspring and Klevu combined under PSG ownership, alongside Intelligent Reach. Three separate products with separate contracts and pricing, not one merged platform.

They said why they came. You showed the homepage. Show them that next, not everything.

Search came back with nothing they can use. Show a close match for what they typed.

They filtered until there was nothing left. Clear the filter that emptied the grid. Show what comes back.