How to Turn Ecommerce Search Data Into Better Catalogue Decisions

Start with the decision, not the dashboard

Before collecting more data, write down the decisions the business is prepared to make. Typical outcomes include adding a product, enriching attributes, changing product wording, mapping a synonym, creating a collection, adjusting ranking, correcting availability or publishing a buying guide. If a metric cannot influence one of those outcomes, it may be interesting but it is not yet operational.

Give each decision type an owner. Merchandising can assess demand and range gaps; product-data teams can repair titles, identifiers and attributes; search owners can tune matching and ranking; content teams can answer research questions; and commercial teams can decide whether demand justifies sourcing. A shared queue prevents the same query from being passed between teams without resolution.

M.I.A.I Search Intelligence is designed for query analysis, search-gap discovery, relevance review and approved feedback signals. Its purpose is to help a business understand search behaviour, gaps and opportunities across a catalogue, including zero-result reviews, customer-language mapping and prioritised catalogue improvements.

Capture enough context to explain the search journey

Store the raw query, a normalized comparison value, timestamp, market, device class, result count and session-scoped journey events. When available, connect the query to the products shown, clicks, refinements, product views, add-to-cart events and purchases. Preserve the raw wording because it contains the customer's language; use the normalized value only to group obvious variants such as capitalization, whitespace and plural forms.

Google Analytics enhanced measurement can emit a view_search_results event when a results page is shown and place the query in the search_term parameter. That is a useful starting point, but a results-page view does not prove that the results were relevant. Result count, impression order and subsequent behaviour provide the missing context.

Respect data minimisation. Search boxes sometimes receive email addresses, order numbers, phone numbers or other personal information. Exclude or redact those values before analysis, limit retention and access, and avoid sending personally identifiable information into analytics systems. The goal is to understand demand patterns, not identify an individual visitor.

• Raw query and normalized grouping value
• Market, language, device and session-scoped timestamp
• Result count and the products displayed
• Clicks, refinements, exits, add-to-cart events and purchases
• Known stock or market restrictions at the time of the search
• A privacy filter and documented retention period

Separate five different search outcomes

A single zero-results column hides several distinct problems. Create cohorts that describe what happened after the query. Start with no result, results but no click, clicked then returned and reformulated, clicked with useful commercial action, and a successful journey that ended in purchase or another agreed outcome.

Shopify's behaviour reports distinguish searches by query, searches with no clicks, searches with no results and search conversions over time. That separation matters: a query with results but no clicks suggests a relevance or presentation problem, while a query with no results may indicate missing language, missing data or missing range. A successful query is also evidence; it shows which customer wording already maps to a useful outcome.

Review trends by week and market rather than reacting to one unusual session. Account for reporting delays and bot filtering where the platform supports it. Keep the cohort definitions stable long enough to compare before and after a change.

Classify the cause before choosing the fix

For each material query group, inspect the actual result set and catalogue evidence. Assign one primary cause and any secondary contributor. Consistent labels make the queue measurable and help the business see whether its largest problem is range, data quality, language, ranking or customer guidance.

A missing-range query describes a product the business does not sell. A terminology gap occurs when the product exists but the customer's words do not appear in searchable data. An attribute gap means the product exists yet the size, compatibility, material or specification needed to retrieve it is absent or inconsistent. A relevance problem returns products, but the strongest match is buried. Availability and market gaps occur when the item exists globally but cannot be bought in the visitor's location. Guidance queries ask a question that may need comparison content or a buying guide.

Do not automatically create synonyms for every failed phrase. A synonym that connects different specifications or incompatible parts can make search look busier while making product selection less safe. Require evidence from product data or a knowledgeable reviewer before approving technical equivalence.

• Missing product or range opportunity
• Customer terminology not represented in the catalogue
• Missing or inconsistent searchable attribute
• Poor ranking or overly broad matching
• Unavailable stock, channel or market restriction
• Navigation, policy or advice question rather than product demand
• Noise, bot activity or a query that should be excluded

Prioritise by value, friction and confidence

Query frequency is useful, but it should not be the only priority rule. A frequent vague query can be less valuable than a smaller group of precise, high-intent searches for an item the business is equipped to sell. Build a simple score from demand frequency, customer friction, business relevance and confidence in the diagnosis.

Customer friction increases when a query repeatedly returns nothing, produces no clicks or causes several reformulations. Business relevance can reflect margin, strategic range, availability, market and the cost of serving the request. Confidence should be higher when the same need appears across multiple sessions, has a clear catalogue match and has been reviewed by a product specialist.

Keep the formula explainable. A reviewer should be able to see why an item is near the top of the queue and change an incorrect classification. Automated ranking can organise evidence; it should not turn a guessed synonym or unsupported demand assumption into an invisible production change.

A concrete example: technical parts searches

Imagine a parts retailer sees repeated searches for 'bottom roller Takeuchi TL12', 'TL12 track roller' and a known OEM reference. The first query returns nothing, the second returns many generic rollers with no click, and the reference finds the correct product. Treating these as three unrelated rows understates the opportunity.

The product specialist confirms that customers use bottom roller and track roller for the same relevant component in this catalogue context, and verifies which TL12 variants are compatible. The data team keeps the product's stable identifiers, adds the approved terminology, records the compatibility attribute and checks that the OEM reference remains searchable. The search owner reviews whether exact model compatibility should rank above generic text matches.

The work item contains the raw queries, affected sessions, current results, approved product evidence, proposed catalogue changes and an owner. After release, the team compares the same query cohort over a defined period: result coverage, click rate, qualified product views and add-to-cart activity. If irrelevant products begin receiving clicks, the team investigates rather than declaring success from a lower no-result count alone.

If the retailer does not stock a compatible roller, the correct outcome is different. The query becomes a range decision supported by observed demand. The business can source the item, direct customers to an honest alternative when one exists, or publish clear guidance. It should not force an incompatible product into the results.

Turn approved findings into catalogue work

Every accepted finding should create the smallest appropriate change. Terminology may belong in a searchable synonym, product title, description or structured attribute depending on its meaning. Compatibility belongs in governed product data, not only in promotional copy. A range gap belongs with buying and merchandising. A guidance question may require content linked from search results.

Preserve product and variant identity when catalogue records change. Use stable platform IDs and governed source identifiers so enrichment reaches the intended item. Titles, handles and descriptions change too often to serve as safe matching keys. Record the before value, proposed value, evidence, approver and deployment result.

Batch similar work where it improves consistency, but preview the affected products before publication. A shared synonym or attribute mapping can alter many result sets. The preview should show both intended gains and obvious collisions, especially for model names, abbreviations and technical terms.

Compare internal search with external demand carefully

Google Search Console's Performance report groups external Google Search data by queries and provides clicks, impressions, click-through rate and average position. Compare that language with internal search to find patterns, but do not merge the datasets as if they measured the same journey.

External queries describe how people encountered the site in Google Search results. Internal queries describe what visitors asked after reaching the site. A term with many Google impressions but little onsite search may already have a strong landing page. A frequent onsite query with little Google visibility may reveal navigation friction, a returning-customer task or an opportunity for a dedicated page.

Use the comparison to form a hypothesis. Then inspect the landing pages, result sets, catalogue records and commercial context before changing SEO copy or product data. Different attribution, privacy thresholds, date ranges and aggregation rules can produce apparent discrepancies that are not catalogue faults.

Run a controlled feedback loop

A useful search-intelligence process repeats on a predictable cycle. Collect a stable period of data, group and classify queries, review the highest-value evidence, approve specific changes, publish them through the responsible system and measure the same cohorts again. Keep a control period or unaffected query group where practical.

Measure several outcomes together: no-result rate, no-click rate, reformulation rate, useful product clicks, add-to-cart activity and purchase outcomes where volume is sufficient. Also track operational measures such as time to classify a gap, time to publish a correction, repeated exceptions and percentage of proposed changes rejected in review.

Avoid chasing daily noise. Seasonal demand, campaigns, stock changes and site releases all affect search behaviour. Annotate those events, compare equivalent periods and keep the original query evidence so a later reviewer can understand why the decision was made.

Common mistakes that make search data misleading

The first mistake is equating fewer zero-result searches with better discovery. Broad matching can reduce zeros while surfacing irrelevant products. The second is counting clicks as success without checking whether visitors returned immediately, refined the query or abandoned the journey.

Another mistake is rewriting product titles around every phrase. Titles still need to identify products clearly; customer language can be captured through appropriate descriptions, structured attributes and governed search mappings. Teams also lose trust when reports group unlike technical queries or silently discard rare but commercially important searches.

Finally, do not let an insight tool become an unreviewed publishing system. Search activity is evidence of customer intent, not proof of product equivalence, legal suitability, compatibility or commercial demand. Approved feedback signals and accountable owners keep the loop useful.

Search intelligence operating checklist

• Define the catalogue, ranking, content and range decisions the data can trigger.
• Capture raw queries, result context and subsequent journey events.
• Filter personal information, bots and known internal test traffic.
• Separate no-result, no-click, reformulated and successful searches.
• Classify each material query group before proposing a fix.
• Prioritise with frequency, friction, business relevance and confidence.
• Use stable product and variant identifiers for any catalogue change.
• Preview shared mappings and high-impact changes before publication.
• Keep internal site search distinct from external Google Search reporting.
• Measure the same cohorts after release and reopen work when relevance worsens.