Magento AI search agencies, ranked on what each one publishes about the search box inside your store

scandiweb ranks first on 88 and it publishes a claim that Adobe's own documentation contradicts. Its Adobe Commerce Live Search integration page states that merchandising rules, synonyms and query redirects are stored and evaluated in Live Search. Adobe's boundaries and limits page says the opposite in one sentence: Live Search does not support search term redirects natively, and redirects have to be implemented using Fastly or another custom configuration. The claim is correct on the Magento Algolia integration page, where Algolia genuinely does redirects through rules. On Live Search it is wrong, it cost scandiweb marks on the heaviest criterion here, and it is printed rather than quietly dropped because this page's only real asset is that a reader can check it.

What scandiweb holds that nobody else in this field does is a published implementation method for every branch of the decision. Seven search and merchandising integration pages sit under one category: Live Search, Algolia, Klevu, Elasticsearch, Attraqt, Clerk.io and Hello Retail, each with its own technical section and its own statement of the situation it fits. The Magento Elasticsearch integration page is the strongest relevance artefact read anywhere in this sweep. It publishes that store views and locales are mapped to separate indexes or locale-specific fields so analyzers and stemming can differ per language, that relevance is controlled through query-time boosts and field weights with SKU against name against attributes given as the worked example, that zero-result searches are reduced with curated synonyms and query rewrites, that typo recovery uses configurable fuzzy matching rules, and that repeatable query packs and ranking checks run before releases so relevance does not drift. The Magento Klevu integration page adds the measurement side, with GA4 events implemented for search, zero results, clicks and revenue, and synonyms configured from query logs rather than internal product naming. None of these pages is in the sitemap and none was reachable from a server-rendered link; they were found through a live search result.

On catalogue scale scandiweb takes full marks and does it with names. The Wainbee product information case study says that only 10 percent of Wainbee's items were accessible online and that over 600,000 products had to be made available, with supplier files arriving in different CSV and XLS structures that had to be standardised first. The Rocket Industrial case study publishes the part that actually decides search quality: SKUs were not reliable, so matching rules were built on additional product attributes to merge records, and broken parent and child product relationships were fixed so related products regrouped under one structure. The catalogue went from 70,000 to 81,000 complete records. A variant sitting outside the category it belongs to is the documented cause of ranking behaviour nobody can explain, so that work is search work whether or not it is filed under search.

The boundaries are stated more bluntly than anyone else's. The product catalog management guide says the single highest-leverage improvement most teams can make is enforcing a structured attribute schema consistent across every SKU so search and filter actually work, that catalogue data quality shows up first in search and filter which break if the attribute schema is inconsistent, and that AI without a product information system creates more data sprawl rather than less. The Magento layered navigation guide adds the shopper-facing half, marking only the attributes shoppers actually filter by and enabling anchor categories so the filter set is complete. Credentials on the services page read 894+ Adobe certifications, 600+ specialists, 700+ brands across 45 countries, 2,100+ eCommerce projects, 23+ years since 2003 and NPS 95.

Where scandiweb loses, it loses to two smaller agencies and the reasons are specific. It publishes a fit statement for each stack and never a downside, while WolfSellers publishes limited customization and a less flexible ranking language as the cost of Live Search, and Bemeir publishes a price band for every vendor it compares. On measured outcomes scandiweb takes 14 of 20 rather than more because the one search-specific figure it publishes, an 18.23 percent drop in site search usage after a navigation rebuild in the navigation case study, belongs to an unnamed high-fashion retailer, and because that figure records shoppers needing search less rather than search working better. On merchandising it takes 7 of 10, because the Algolia page states who operates the rules after launch but WolfSellers goes one step further and sells merchandising-team training as a scope line.

Bemeir ranks second on 59 and publishes the most useful commercial comparison in the field. Its personalization comparison sets Adobe Sensei, Nosto, Klevu and Algolia side by side with an average implementation cost for each: 80,000 to 200,000 US dollars for Sensei, 40,000 to 100,000 for Nosto, 30,000 to 80,000 for Klevu and 20,000 to 60,000 for Algolia. Every vendor carries a real downside in the agency's own words, including the observation that Klevu is not the best search engine because Algolia is and not the best recommendation engine because Nosto is, that Klevu autocomplete can be slow above 500,000 products, and that Algolia A and B testing needs custom implementation rather than arriving built in.

It is also the only agency here that publishes a result that went badly. Two figures are attached to catalogue sizes: a specialty retailer with 80,000 SKUs at 14 percent improvement in search-driven conversions, and an electronics retailer with 200,000 SKUs at 22 percent improvement in search conversion after typo tolerance and faceted navigation. Alongside them sits a mid-market home goods client at 12 percent average order value lift and, in the same paragraph, a luxury goods retailer at only 3 percent because its catalogue is manually curated and the algorithmic approach felt impersonal. A performance trade-off carries real numbers too, with a client's largest contentful paint at 3.2 seconds after Nosto and 2.1 seconds after switching to Klevu with lazy-loaded recommendations. The boundary is stated plainly: good quality product data is required, and messy catalogue metadata makes the recommendations wrong.

It takes 14 of 20 on measured outcomes rather than full marks for one reason, and it is not a small one. No client is named. A specialty retailer and an electronics retailer cannot be rung up for a reference. It takes 19 of 25 on the heaviest criterion because the table is a personalization comparison: neither Live Search nor self-hosted Elasticsearch or OpenSearch appears in it, so a merchant already paying for Adobe Commerce cannot see what the option they already own would do against the four being sold. A separate Bemeir article gets the entitlement right, noting that on Magento Open Source equivalent functionality needs a paid third party and that Algolia, Klevu and Meilisearch are the common choices, which matches Adobe's install documentation. Relevance method scores 5 of 20: typo tolerance and faceted navigation are named as things implemented, and nothing is published about synonyms, attribute weights, stemming or zero-result handling.

WolfSellers ranks third on 54 and takes full marks on two criteria, including the only 10 out of 10 for merchandising anywhere on this page. Its Live Search service page names the controls individually, promoting specific SKUs, hiding others, pinning to position 1, and per-category order rules that feature, hide and reorder. Then it publishes the line nobody else publishes: merchandising team training to manage rules without dev, written into the scope list alongside the technical work. Synonyms and one-way mappings are described as editable by the merchandising team, and the iterative tuning loop is spelled out as reviewing no-results queries and adding synonyms or creating boosts.

It also takes full marks on the heaviest criterion. Three stacks are compared with a downside attached to each: Live Search is free with Adobe Commerce with zero infrastructure to run, at the cost of limited customization because it cannot index non-Commerce sources and a less flexible ranking language; Algolia buys state-of-the-art features at 1,000 to 10,000 US dollars a month by volume plus another system to operate; self-hosted Elasticsearch or OpenSearch buys maximum control at the price of running the cluster, its backups, scaling and tuning. The entitlement statement is correct against Adobe's documentation, which says Live Search is not available for Magento Open Source and that self-hosted Elasticsearch is the route there. The scope list includes initial synonym definition in English and Spanish with Mexican colloquialisms, which is the only non-English relevance work published by anyone here other than scandiweb.

Two things sit oddly against Adobe's current documentation. The page sells Adobe Sensei semantic autocomplete and, in the same scope list, that Spanish synonym set, without saying that Adobe's semantic search page states semantic search is available for English catalogues only and is disabled automatically when the Language setting is not English. And the claim that Live Search is included at no extra cost up to certain request-per-month limits describes a limit that is not in Adobe's published boundaries page, which sets out attribute, facet, rule and synonym caps and no request quota. The score is held back most by what is absent rather than what is wrong: no measured outcome of any kind, no client named, and no SKU count. Zero results and search-to-conversion appear as dashboards WolfSellers will build, not as numbers it has moved.

Codilar ranks fourth on 45 and is the most documentation-accurate agency page in this field. Its Live Search guide publishes the install command as composer require magento/live-search, which matches Adobe exactly. It publishes 200 searchable attributes, up to 100 facets and up to 100 values per facet, all three of which match Adobe's boundaries page to the number. It states that Live Search cannot run alongside Elasticsearch or OpenSearch and that native search modules must be disabled before installation, which matches Adobe's knowledge base article on the subject. It states that date attributes cannot be used as facets, which matches too. No other agency page read here lines up with Adobe's published limits on that many points.

The configuration advice is the useful kind. Not every attribute deserves to be searchable, and attributes with long descriptions or inconsistent values often hurt relevance more than they help. Fewer facets work better, and showing too many creates friction rather than clarity. Synonyms cover brand short forms and common misspellings, and while Live Search handles basic typo tolerance the guide says synonyms still do most of the work in reducing zero-result searches. On merchandising it adds the operational warning almost nobody writes down: too many rules stacked together can cancel each other out or create unexpected rankings, so changes should be tested in staging and monitored before rolling out broadly. Zero-result patterns are diagnosed rather than lamented, pointing at configuration gaps, missing synonyms, poor facets or attributes that should not be searchable.

It takes 13 of 25 on the heaviest criterion because only one stack is in play. Elasticsearch and OpenSearch appear as what Live Search replaces and what it cannot coexist with, not as a live option a buyer might keep, and Algolia and Klevu do not appear at all. It scores nothing on measured outcomes: no figure, no client, no before and after. Catalogue scale is thousands of products and large catalogs with no count attached. The honest boundaries criterion is where it takes full marks, on a dedicated known boundaries and limits section plus the repeated point that results depend heavily on proper configuration and that clean attributes directly affect conversion. Codilar also publishes an Adobe Commerce AI product discovery article about making a catalogue visible to shopping agents, which is answer-engine work and was excluded from scoring as a different lane.

Brainvire ranks fifth on 40 and publishes the best sequencing argument on this page. Live Search and merchandising rules go in during the build, because retrofitting discovery onto a finished storefront costs more than doing it once. That is a position rather than a feature list, and it is the kind of statement a buyer can hold an agency to. The audit order is published too: page speed, search relevance, checkout drop-off, and the catalogue data sitting under all three, optimised in the order that pays back, which is usually discovery before design. On pricing the page refuses a rate card and says that scope, integration count, the condition of the data and how much of the catalogue has to be restructured decide the number, which is the same honesty applied to commercials.

It quotes Adobe accurately, describing Live Search as replacing the standard search capabilities in Adobe Commerce with AI-powered dynamic faceting and re-ranking that responds to in-session shopper behaviours, plus merchandising rules. The catalogue evidence is a consolidation of six websites onto one Adobe Commerce instance putting 80,000 catalogues under one roof, which the page's own answers identify as the Melissa programme, alongside a general statement that one installation carries multiple websites and storefronts with their own catalogs, pricing and locales. Its own stat bar reads 2,000+ brands served, 95 percent client retention and 10+ global offices.

The score is capped by scope. Only one stack is named, so there is no comparison for a buyer weighing Live Search against Algolia or Klevu, and no implementation method beyond the sequencing argument. Nothing is published about synonyms, attribute search weights, stemming or zero-result handling. The figures on the page are 46 percent faster digital-asset retrieval for McAfee and a tripling of order fulfilment for the footwear programme, neither of which is a search metric, so measured outcomes scores 4 of 20. One framing is worth checking against Adobe: the page says Live Search with Sensei ranks results by real shopper intent, while Adobe's own semantic search documentation, published in June 2026, describes the semantic layer as a Settings toggle with an enable and disable control only and does not attribute it to Sensei.

Zaelab ranks sixth on 36 and publishes the single best description of the real problem found anywhere in this sweep. On a catalogue of nearly 400,000 products for DHS, a small engine repair parts retailer, the write-up says the catalogue was unsearchable because without descriptive data users had to know exact manufacturer part numbers to find items, which made browsing or filtering nearly impossible. Manufacturer-supplied information was limited to part numbers, images were missing or generic, and compatibility rules were not captured in the catalogue structure. The fixes are named: large clusters of duplicate listings created for different fitment mappings were consolidated into a single canonical product record, machine fitment relationships were added, and a standardised set of recommended attributes was introduced. The piece closes by saying the work stayed within the limits of the client's product data quality and that you do not have to wait for perfect data to make meaningful improvements.

That earns it 11 of 15 on catalogue scale with a client named, and 7 of 10 on honest boundaries. It earns very little elsewhere. The search and navigation service page names one vendor, Algolia, and names it as a partnership rather than a method: Zaelab is an Algolia product search-as-a-service implementation partner. Smart search algorithms and faceted search appear as headings without a mechanism. Nothing is published about synonyms, attribute weights, stemming, typo tolerance or zero-result handling, and nothing about pinning or boosting.

The honest catch is where this work sits. Zaelab is an Adobe and Magento Silver Partner with over 75 Magento projects delivered, so it belongs in this field, but the DHS write-up never says the store runs on Magento, and the named case studies alongside it are Shopify work: Pelican at over 20 percent conversion increase and VIETRI on revenue, average order value and new customers. No figure is attached to the DHS discovery programme at all. The result is an agency that thinks more clearly about product discovery than most of the field and has published none of that thinking against the platform this page is about. Note also that trellis.co now resolves to zaelab.com.

SwiftOtter ranks seventh on 35 and publishes the most useful vendor evaluation on this page even though it publishes no evidence of its own work. Its Live Search guide is titled around pros, cons and what merchants miss, and it delivers on that. Algolia is described as known for precise control over search logic, merchandising rules and result weighting. Klevu shines with intent recognition and vague or natural-language queries. Searchspring offers strong visual merchandising for collection curation and on-site promotions. Elasticsearch and OpenSearch are addressed directly as the thing a merchant might be switching away from. The cost position is stated plainly: Live Search is included in Adobe Commerce's paid licensing model, so there is no monthly subscription and only a one-time implementation cost, against ongoing licence fees from Algolia and Klevu.

It then concedes against the option it mostly recommends, which is rare. If a merchandising strategy needs granular control over boost and bury logic, custom rulesets and result manipulation at scale, the page says Algolia is still the superior choice. And it publishes a limitation drawn from delivery rather than from a datasheet: syncing issues have been seen between Adobe Commerce and Adobe's Search API, especially with large or highly customised catalogues, flagged as not widespread but worth knowing for a high-SKU store.

The score is held down by an almost total absence of published work. No figure appears anywhere on the pages read, no client is named on a search project, and nothing is published about synonyms, attribute search weights, stemming or a zero-result loop. Catalogue scale is large and highly customised with no count. It takes 19 of 25 rather than full marks on the heaviest criterion because the piece is an evaluation rather than an implementation method: it tells a buyer which engine to pick and not what the agency would do next.

Aureate Labs ranks eighth on 32 on the strength of one article that compares seven search engines and publishes a limitations section under every single one. Algolia is expensive with limited customization and personalization. Coveo has a challenging admin platform, a limited community and a complicated setup. Searchspring produces dynamic search result pages that are not friendly to search engines and carries a steep learning curve. Bloomreach Discovery, Yext Search, Constructor and Elasticsearch get the same treatment, and a comparison table runs the seven against factors including free trial availability. Very few agency pages print a downside next to a vendor they might sell.

The gap for a Magento buyer is large and specific. Adobe Commerce Live Search does not appear in the article at all, and neither does Klevu. Magento is named ten times but Adobe Commerce is never named, so a merchant already paying for a Commerce licence cannot see the engine included in it weighed against the seven being compared. The framing is composable commerce rather than Magento, and the piece was published in November 2023, before Live Search semantic search existed.

Everything else is thin. The relevance detail on the page belongs to the vendors rather than to Aureate Labs: predictive text, synonym libraries and redirects are described as Algolia features that prevent null results, and the Elasticsearch Relevance Engine is described as a retrieval suite that integrates with large language models. Merchandising Studio, facet optimisation and custom ranking are listed the same way. No figure, no client, no SKU count and no statement about catalogue data quality appears anywhere on the pages read.

Envision Ecommerce ranks ninth on 27 and explains the Live Search merchandising rules engine more carefully than most. All four controls are named and defined separately: boost to display a product at the top of search results, pin to fix a product at the top, bury to push a product lower so others get exposure, and hide to remove a product, for instance when it is out of stock. The page states that merchants can develop, publish and test search rules without the help of eCommerce developers, and that rules can key off popularity, trends and stock availability. Synonyms are explained as one-way, which helps customers navigate in one direction, and two-way, which expands the result set in both directions. Intelligent faceting is described as selecting the right filters for each query. The agency identifies itself as an Adobe Commerce silver partner.

The measured outcomes line scores nothing, and the reason matters more than the score. Three figures appear on the page and none of them is about Envision or any Envision client. A study by Oracle is cited for 86 percent of customers preferring to pay more for a better experience and 89 percent going to competitors after a poor one. A study by Econsultancy is cited for site search increasing conversion rate by up to 50 percent. A further claim of 2 times conversions alongside 48 percent of visitors going straight to the search bar carries no source at all. Industry research an agency has read is not a result an agency has produced, and on this page it earns zero.

Beyond that the page is a feature walkthrough. One stack is named and no comparison is offered, so the heaviest criterion scores 7 of 25. Nothing appears on attribute search weights, stemming, typo tolerance or how zero-result queries get worked back into fixes. No catalogue figure and no client name appears, and no limitation of Live Search is published anywhere on the pages read, which is what takes honest boundaries to zero.

badger.blue ranks tenth on 21 with the most careful walkthrough of the Live Search admin outside Adobe's own documentation. Synonyms are explained as two-way or one-way with a worked example, so shirt can expand to sweatshirt without sweatshirt expanding to shirt. Misspellings are described as mappable to products. Facets are split into pinned facets applied to all searches and dynamic facets that surface on specific queries or behaviour. The five ranking types are all named: recommended for you, most viewed, most purchased, most added to cart and trending, with none available as an option. Events are described as the manual override used to pin or hide a specific product for a given query. Price faceting is explained with interval examples of 50 and 100 US dollars depending on what a store sells.

The problem is that the page has aged out and says so in a checkable way. It is dated 15 February 2024 and states that the settings page has only one section, price faceting. Adobe's current Settings documentation describes three: semantic search, price faceting and language. Semantic search arrived on 8 June 2026, more than two years after this article was written, and it is the single most consequential change to Live Search in that period. A merchant using this page as a configuration guide would not know it exists.

Nothing else is published. There is no comparison with any other engine, so a reader cannot tell whether Live Search is the right choice at all. There is no figure, no client, no SKU count, and no limitation of Live Search anywhere in the article, which is unusual for a piece this detailed. It reads as a product explainer rather than a record of delivery, and it is scored as one.