How AI Product Recommendation Engines Work: 9 Platforms for Online Stores

An AI product recommendation engine is not a black box. It is a ranking function that scores your catalogue against one shopper, on one page load, and it does that with whichever of six mechanisms the vendor chose to implement. Understanding which mechanism a vendor actually runs tells you more than any feature grid, because it tells you what happens when your data is thin.

A 200-SKU jewellery store and a 60,000-SKU electronics retailer need different recommendation systems. Vendor names need checking too: Klevu and Dynamic Yield no longer operate as independent vendors with their own pricing, despite appearing that way in some 2026 shortlists.

Why AI product recommendation engines are not interchangeable

They are sold on a common promise and priced on incompatible models, so the decision is not “which model is smartest”. It is three questions, in this order. How much interaction data do I actually generate per product per month? Which mechanism survives my weakest data? And what does this cost at my volume, then at double the volume?

If your answer to the first question is “a few hundred interactions per SKU per month”, you want an engine whose cold-start behaviour is content-based and whose pricing does not punish you for having a small catalogue. If you have tens of thousands of sessions a month and a catalogue that rotates, you can afford collaborative filtering and a request-priced platform. This article gives you the mechanism, then the numbers, then nine platforms judged on both.

Pricing and performance evidence

Prices, limits and features are from vendor documentation checked between 12 and 16 October 2026. No trials were run to measure latency or ranking quality. Sales-led products without public prices require a quote; case-study uplift figures are vendor-reported.

The six mechanisms: what these engines actually do

Every platform in this article is running some combination of these. They differ mostly in which one they lead with, and the lead choice tells you what breaks first.

Content-based filtering: reads the product, not the shopper

Content-based filtering recommends products whose attributes resemble what you are already looking at. It compares item properties, which vendors call catalogue fields: material, colour, category, price band, brand, size availability.

• Data it needs: your catalogue. No behaviour required.
• What it breaks on: thin or inconsistent catalogue data. If your descriptions are empty and your categories are a single “Shoes” node, the similarity signal collapses.
• Retail output: a shopper on a navy merino cardigan is shown three other navy merino cardigans in the same price band.

Recombee implements this as distinct named models rather than one blended thing. Its documentation lists recombee:similar-properties as recommending similar items based on item properties such as numerical values, sets and texts, and recombee:visually-similar as a separate model that recommends on visual similarity “using deep-learning neural networks to process the images provided through the image properties”. Keeping those apart is good design, and it is the difference between a store that works on day one and one that does not.

Collaborative filtering: learns from what people did

Collaborative filtering recommends products because other shoppers with similar behaviour bought them. The classic formulation is matrix factorisation over user-item interactions; modern versions generally learn dense item embeddings from that same signal.

• Data it needs: a lot of it. Views, add-to-carts, purchases, and consistent identity across sessions.
• What it breaks on: new products, long-tail products, sparse traffic and anonymous sessions.
• Retail output: a repeat customer who always buys technical outerwear is routed to the best-selling waterproof shell, not the highest-rated one.

The academic foundation is solid and old. Neural Collaborative Filtering (arXiv 1708.05031, submitted August 2017) and SASRec, Self-Attentive Sequential Recommendation (arXiv 1808.09781, 2018), are still the standard citations, and both are more than seven years old. Treat them as settled background, not news.

Hybrid ranking: the only sane production default

Hybrid ranking blends collaborative and content signals, then applies merchandising rules on top. Almost every platform here runs some version of it.

• Data it needs: catalogue plus whatever behaviour exists.
• What it breaks on: nothing gracefully. It is the fallback-tolerant option.
• Retail output: the engine proposes the top ten, then rules push in-stock items and suppress discontinued SKUs.

The rule layer is what makes hybrid deployable in commerce. Recombee exposes it as first-class objects, with filters deciding which items may appear, boosters deciding which are preferred, constraints limiting how often items sharing a condition may appear, and a query language called ReQL that composes them. Algolia sells the same idea as Rules, capping them at 10 per index on the free and Grow plans and 10,000 per index on Grow Plus and Elevate.

Session-based and real-time behavioural signals

Session-based recommendation predicts the next click from the current anonymous session, with no login and no history. This is the mechanism that matters most for first-time visitors and for traffic from paid social, and it is quietly the most valuable one for most mid-sized stores.

• Data it needs: an event stream, not a historical database. This is the cheapest useful signal.
• What it breaks on: event loss, ad blockers and consent refusals. If your tag fires on 40% of page views, the session is fiction.
• Retail output: a shopper views three espresso accessories and gets a grinder before they leave the category page, with nothing stored about them yet.

Research here is active and current. A 2026 paper on session-based recommendation with pure convolution (arXiv 2609.34802, September 2026) reports that transformer-based models dominate this subfield while questioning whether self-attention is necessary at all. For architecture, a useful survey is Graph and Sequential Neural Networks in Session-based Recommendation (arXiv 2408.14851, published August 2024).

Embedding and vector retrieval: matching meaning, not keywords

Vector retrieval converts products and queries into numeric vectors and returns the nearest neighbours by distance, rather than by matching keywords. This is what lets “something to keep drinks cold for camping” find a cooler bag.

• Data it needs: catalogue text and images, plus enough traffic to learn item vectors from behaviour.
• What it breaks on: homogenous catalogues. If all 4,000 SKUs are white cotton t-shirts, every vector is near every other vector.
• Retail output: a shopper types a sentence, gets products, not a zero-results page.

Algolia describes its NeuralSearch as combining its fastest semantic search with keyword search and says the underlying vector-search algorithm powers 30 billion searches a week, which is a vendor-reported figure. The tier placement matters more than the claim: on the cited pricing page, NeuralSearch is listed only under the annual Elevate plan. Bloomreach positions its semantic and AI capabilities inside the Loomi AI platform, and Recombee sells search:semantic as an add-on, with its docs describing it as using a large language model to improve search accuracy by understanding the intent and context behind queries.

Conversational and agentic discovery: the LLM layer

Conversational discovery puts a language model between the shopper and the catalogue, so discovery happens in sentences rather than filters. In 2026 this has split into two distinct products: the in-store assistant that guides shoppers, and the off-site agent that feeds AI shopping surfaces.

• Data it needs: a clean catalogue, because an LLM can only recommend from products you have described.
• What it breaks on: hallucinated attributes, long-tail catalogue gaps, and pricing errors that reach the customer.
• Retail output: a shopper asks for a gift under $60 that ships by Friday; the assistant returns four SKUs with reasons.

Athos Commerce is the clearest vendor example of the split. Its site lists a Conversational Assistant that “trains on your catalog to guide shoppers toward purchase”, a Channel Assistant that audits and optimises feeds across Google, social and marketplaces, and a GEO Assistant that “enriches catalogs and syndicates to Gemini, ChatGPT, and emerging AI platforms”. The GEO Assistant is worth flagging: if a meaningful slice of your discovery traffic is arriving from AI assistants rather than from a search engine, your catalogue structure becomes an AI-surface problem, and no recommendation engine on this list solves it.

This is the least settled mechanism commercially. Recent arXiv work includes a September 2026 study of how agentic AI turns human heuristics into a surrogate consumer (arXiv 2609.28372) and an August 2026 paper on learning from online user feedback for shopping agents (arXiv 2608.11604).

The cold-start problem: why most stores don’t need a machine learning platform

Cold start is the state where a recommender has no interactions to learn from, so it cannot personalise anything. It is the single most important technical fact in this category, and almost no vendor page leads with it.

The academic framing is worth one citation, and it is old: Deep Learning to Address Candidate Generation and Cold Start Challenges in Recommender Systems: A Survey (arXiv 1907.08674, published July 2019, so roughly seven years ago). Its central point still holds. Cold start has two independent halves. New items have no interaction history, which is a catalogue problem you solve with content and visual signals. New shoppers have no history, which is a session problem you solve with in-session behaviour and popularity priors.

Vendors acknowledge the second half more openly than the first. Recombee’s integration guide warns that if you model every size and colour as its own item, “it may take longer to train the models properly at lower traffic sites as the number of interactions is divided between all the variants”. That sentence is the most useful cold-start sentence written by any vendor in this category, and it should shape how you model your catalogue before you buy anything.

So: what should a small store actually buy?

If you run about 20 SKUs, buy nothing. Write the bundles yourself. A hand-curated frequently-bought-together block on a 20-SKU catalogue will be at least as good as any model and you can change it in an afternoon. Shopify’s predictive search, which the Storefront API exposes as the predictiveSearch query, will carry most of the search load.

If you run a few hundred SKUs with real traffic, use the free tiers that genuinely work: Algolia’s free plan covers 10,000 search requests, 50,000 records and 5,000 recommendation requests a month, and Recombee’s free tier covers 20,000 catalogue items, 20,000 active users and 100,000 interactions a month. Both are production-usable, not trials.

If you run tens of thousands of sessions a month with a catalogue that turns over, collaborative filtering finally earns its licence fee. That is where Constructor’s reinforcement-learning framing and Algolia’s Dynamic Re-ranking become worth the money, and where a $1,699/month Recombee Pro tier or a six-figure Bloomreach contract stops looking absurd.

The nine platforms compared, on data and architecture

Platform	Data it needs to work	Behaviour with no interaction data	Documented refresh model	Vendor-stated hard limits
Algolia	Catalogue records plus Insights events; personalised ranking needs a stable authenticatedUserToken	Recommend ships on the free tier at 5,000 requests/month; NeuralSearch is Elevate-only	Advanced Dynamic Re-ranking documented as reranking “up to 24 times more frequently, with hourly refresh rates”	5 browsing facets per index; 100,000 unique queries reranked; 100 hits per query; 10 million Insights actions/day
Constructor	Behavioural clickstream from your site, augmented with third-party market data	Vendor describes reinforcement learning, which by definition needs interaction history	Vendor states “real-time” on its product page	Not published
Recombee	Items, item properties and interactions; content models work off properties alone	Strongest documented cold start: similar-properties and visually-similar need no clicks	Real-time analytics on every tier; request-time API	Catalogue items 20,000 / 20,000 / 1,000,000 / 4,000,000 by tier; semantic search is an add-on
Nosto	Module-dependent; qualification runs a PoC against your own store data	No self-serve trial; structured Proof of Concept for qualified merchants	Vendor claims sub-millisecond edge delivery of personalised content	Not published
Bloomreach	Vendor states 500K+ monthly visitors for its smaller customers	Vendor explicitly excludes low-volume programmes from its fit criteria	Not published	Not published
Athos Commerce	Site search, catalogue and feed data; AI agents train on your catalogue	Not published; every plan is “Request Pricing”	Not published	Not published
Dynamic Yield (Mastercard)	Not published; the product now sits inside Mastercard’s personalisation business	Not verified	Not verified	Not verified
Fast Simon	Not published	Not published	Vendor states the algorithm “continuously learns”	Not published
Personyze	Pageviews, segments and 70+ visitor attributes	Works from day one on the free 5,000-pageview tier	Pageview-based tiers with overage billed per 100,000	5,000 to 10,000,000 monthly pageviews by tier

The nine platforms compared, commercially

Platform	Entry price verified 16 Oct 2026	Pricing model	Integration path	Published catalogue or traffic threshold	Genuinely for
Algolia	$0 (free plan)	Request volume: $0.50 per 1K search requests on Grow, $1.75 on Grow Plus, $0.60 per 1K Recommend requests	API, hosted UI libraries, crawler, Segment and GTM connectors	50,000 records free; 100,000 on paid plans; 100GB index	Teams that want search first and recommendations attached
Constructor	No public price	Quote only	API and SDK, enterprise onboarding	None	Large catalogues with deep behavioural clickstream
Recombee	$0 free, $99/mo Standard, $1,699/mo Pro, $4,499/mo Premium	Interaction volume plus extra traffic units at $0.70 / $0.25 / $0.20	No-code embed widgets, REST API, SDKs in JavaScript, Java, Ruby, Python, .NET, PHP	20,000 to 4,000,000 catalogue items by tier	Developers who want a free tier with no expiry
Nosto	No public price	Base platform fee plus a fixed fee on GMV turnover and traffic	App Store install, PoC-led	None	Mid-market brands with volume to justify GMV-linked fees
Bloomreach	“$60K+ a year for most brands”	Platform fee plus scaled usage	Enterprise implementation	500,000+ monthly visitors for smaller customers	Multi-channel brands replacing a point-tool stack
Athos Commerce	No public price	Quote only, tiered into Onsite, Offsite and Complete Discovery	Platform integration, AI agent add-ons	None	Brands wanting search, feeds and AI agents on one contract
Dynamic Yield (Mastercard)	No public price	Quote only	Enterprise	None	Enterprises already in Mastercard’s ecosystem
Fast Simon	No public price	Quote only	Shopify, BigCommerce, Magento, WooCommerce, Wix, custom	None	Platform-native mid-market stores wanting merchandising control
Personyze	$0, then $149 / $250 / $450 per month, from $1,450/month	Pageview volume plus AI task allowances	Single tag, marketplace apps, headless API on Enterprise	5,000 to 10,000,000 monthly pageviews	Small and mid-sized stores wanting 1-to-1 personalisation cheaply

What these platforms cost, in real numbers

Prices below are from vendor pricing pages dated 16 October 2026.

• Algolia. Free: 10,000 search requests per month, 50,000 records, 5,000 recommendation requests, 5,000 crawls, 7-day analytics retention, US/UK/EU-West hosting. Grow, pay as you go: 10,000 search requests included then $0.50 per additional 1,000, 100,000 records then $0.40 per additional 1,000. Grow Plus, labelled new on the page: same inclusions but $1.75 per additional 1,000 search requests, adding AI Ranking, Advanced Personalisation, Query Categorisation, Collections and 90-day analytics retention. Elevate is annual only, quoted, and is the only tier carrying NeuralSearch, AI Collections, Smart Groups and real-time personalisation. Index size is capped at 1GB free and 100GB on paid plans.
• Recombee. Unlimited 30-day trial, then Free at $0 forever, Standard from $99/month, Pro from $1,699/month, Premium from $4,499/month, Platinum custom. Price is calculated on the maximum of your plan limits across interactions, recommendation requests, active users and catalogue items. Extra traffic units cost $0.70, $0.25, $0.20 or custom depending on tier.
• Personyze. Free forever at $0 with 5,000 pageviews, one domain and three active campaigns; Starter $149/month with 50,000 pageviews; Growth $250/month with 100,000; Power $450/month with 1,000,000 and a named account manager; Enterprise from $1,450/month with 10,000,000 pageviews, iOS and Android SDKs and headless API access. Additional pageviews run $40 per 100,000 on Starter and Growth, $100 per million on Power and $80 per million on Enterprise. Every paid plan carries a 30-day money-back guarantee.
• Bloomreach. The pricing page states “Starting at $60K+ a year for most brands” and that “most smaller Bloomreach customers invest $60K+ annually and serve 500K+ monthly visitors”. That is a platform-wide figure spanning search, email and marketing, not a Discovery-only price.
• Everyone else. Constructor, Nosto, Athos Commerce, Fast Simon, Dynamic Yield and Bloomreach Discovery as a standalone line item publish nothing. Nosto at least states the formula: a base platform fee plus a fixed fee calculated on GMV turnover and traffic, and it says outright that it does not offer a self-service free trial.

On uplift claims, one number is quotable and one is not.

“56% of sites fail to adequately support users’ search needs.”: Baymard Institute, Search Bar UI, published 2 October 2026

Baymard is the most methodologically transparent source in this category. Its on-site search research covers 12 quantitative studies with 20,240 participants, and its cart abandonment list averages 50 separate studies to produce a 70.22% documented average, last updated 22 September 2025. That list also contains the most useful context figure in the article: 42% of US online shoppers say they abandoned a cart because “I was just browsing / not ready to buy”. If that is your abandonment profile, better search is not your problem.

Everything else in the “personalisation lifts conversion by X%” genre is vendor-reported, and Algolia’s own pages do not agree with each other. Its blog post of 13 March 2026 announces a commissioned Forrester Total Economic Impact study showing a 213% three-year ROI, a payback under six months and $3.1M net present value for the composite organisation; the report landing page repeats 213%, $4.5M present value of benefit and $3.1M net present value. Algolia’s own pricing page, checked the same day, still quotes a different Forrester result: “ROI as high as 382%”, $4.47M present value of benefit and $3.54M net present value. Those figures should not be treated as interchangeable. Fast Simon’s Ally Fashion case study, published 3 May 2025, reports a 32% sessions increase and 14% average-order-value increase from upsell recommendations. No retailer-published independent result is cited here.

How to roll out a recommendation engine: seven steps

1. Audit your data before you audit vendors. Count monthly sessions, monthly product-detail views and monthly purchases, then divide interactions by active SKUs. Under roughly 100 interactions per SKU per month, collaborative filtering is decorative. Write that number down; it decides everything after this step.
2. Fix catalogue data first. Content-based and vector retrieval are only as good as your titles, categories, attributes and images. This work pays off whether or not you ever buy a platform.
3. Ship the free tier on real traffic. Algolia’s free plan and Recombee’s free tier are both production-usable. Run them for 60 days with real shoppers before spending anything.
4. Instrument attribution properly. Both Algolia and Recombee support passing back a recommendation identifier so later interactions can be tied to the impression that caused them. Without that you cannot tell whether the engine earned its fee.
5. Model variants as one product. Recombee’s own guidance is to make the master product the item and send interactions from all variant pages, precisely because splitting variants divides your interaction count. Do this before you connect the feed, not after.
6. Layer merchandising rules on top and log every override. If a buyer pins an item and conversion moves, you have learned something. If nobody logs it, you will argue about it in six months.
7. Re-run the arithmetic at double your volume. Request-priced platforms bill you more as you succeed, and GMV-linked platforms bill you more as you succeed. Model the year-two bill before signing, not the month-one bill.

What breaks: failure modes and gotchas

“Real-time” is often a tier name. Algolia’s own policy documentation describes Advanced Dynamic Re-ranking as refreshing on hourly rates. Some layers of your stack genuinely recompute per request; others do not. Ask which is which, per feature.

Event loss quietly destroys session models. Session-based recommendation is the best value mechanism for most stores, and the most fragile. Blocked scripts, consent refusals and mobile webviews each delete rows from your stream.

Popularity bias is the default output. When the engine cannot personalise, most frameworks fall back to a trending or bestseller list. You will ship personalisation, measure it, and end up reporting a list that was never personal. Compare against a control group.

Variant fragmentation is the classic self-inflicted cold start. Already covered, and worth repeating because it is the single most common integration mistake.

Pricing models punish the stores that need them least. Recombee prices on the maximum of your plan limits across four separate dimensions, so you pay for the largest one you set. Algolia’s Recommend bill grows with recommendation requests, which grow with traffic. Both are fine; both surprise people in month three.

Nobody publishes their bad case. Vendor case studies lead with uplift. If a vendor will not name a customer who measured a negative result, treat the average as marketing.

Data residency is a real question, not a checkbox. Algolia’s AI Personalization Terms state that end-user events and profiles are hosted in the region the subscriber selects, which is the same choice as the analytics region. Know which region you pick before you connect the tag.

Data protection and EU rules for recommendation engines

Recommendation engines are not high-risk AI under the EU AI Act, but they are profiling under the GDPR. Both halves matter and they are frequently confused.

Under the AI Act, Annex III lists high-risk use cases: biometric identification and categorisation, emotion recognition, critical infrastructure, education, employment, essential public services, law enforcement, migration and justice. Product recommendation for a shop is not on that list, so the high-risk obligations do not apply to you. Article 50 bites in one specific case. If your recommendation slots contain AI-generated or manipulated imagery, video or audio, the provider of the generating system must mark the output as artificially generated in a machine-readable format and you, as deployer, must disclose it where it would otherwise appear authentic, which the Act defines as a deep fake. If your creative is photography, none of this triggers.

Under the GDPR, Article 4(4) defines profiling as any form of automated processing of personal data consisting of the use of personal data to evaluate certain personal aspects relating to a natural person, including personal preferences, interests, reliability, behaviour and movements. An engine that scores sessions against a visitor profile is doing exactly that. Article 22 gives a right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects the person, with narrow exceptions including necessity for entering into or performing a contract and explicit consent, plus safeguards including human intervention and the right to contest. A ranked product grid usually stays below that line. Personalised bundles, credit offers and personalised pricing do not, and that is where you need counsel rather than a blog post.

Practically, you need three things: a lawful basis and consent where you are relying on cookies or cross-site identifiers to build the profile, a data processing agreement with the vendor plus the right to object under Article 21 where you rely on legitimate interest, and an answer on whether your vendor’s subprocessors sit inside the EU.

The bottom line

Start with the data, not the demo. The platforms differ far less in algorithm than their feature grids suggest; they differ enormously in pricing model, and in how gracefully they behave on nothing but a catalogue.

If you have under 100 interactions per SKU per month, buy nothing and merchandise by hand, using the free tiers of Algolia or Recombee for search. If you have real volume, Recombee’s published pricing makes it the easiest to model and Algolia the most flexible to grow into, while Constructor and Bloomreach are enterprise conversations you should only start once search itself is already excellent. Treat every published uplift number as vendor-reported until you run your own holdout.

And check who you are actually contracting with. Four names on most shortlists now belong to someone else: Klevu and Searchspring to Athos Commerce, Dynamic Yield to Mastercard, and LiftIgniter’s domain to Paradium.AI.

FAQ

What is an AI product recommendation engine?

It is software that scores products against a specific shopper or session and returns a ranked list, normally by combining at least two of six mechanisms: content-based filtering, collaborative filtering, hybrid ranking, real-time session signals, embedding-based vector retrieval, and LLM-driven conversational retrieval. Recombee names these explicitly in its documentation, with logics such as ecommerce:similar-products for item-to-item work, ecommerce:similarly-purchasing for user-to-user work, recombee:visually-similar for image-embedding work, and search:semantic, which its docs describe as using a large language model to improve search accuracy by understanding the intent behind queries.

How much does an AI product recommendation engine cost?

Public figures verified on 16 October 2026: Algolia’s free plan includes 10,000 search requests, 50,000 records and 5,000 recommendation requests a month, then $0.50 per additional 1,000 search requests on Grow and $1.75 on Grow Plus, with Recommend at $0.60 per additional 1,000 requests. Recombee runs from a free tier to $99/month Standard, $1,699/month Pro and $4,499/month Premium, with an unlimited 30-day trial. Personyze runs $0, $149, $250, $450 per month and from $1,450 per month. Bloomreach states its entry point is $60K+ a year. Constructor, Nosto, Athos Commerce, Fast Simon and Dynamic Yield publish no price at all.

Do I need a recommendation engine if I only sell a few hundred products?

Almost certainly not. With few interactions per product, collaborative filtering has nothing to learn from and degenerates into a popularity list. Recombee’s integration guide notes that modelling each size and colour as a separate item means “it may take longer to train the models properly at lower traffic sites as the number of interactions is divided between all the variants”. At that scale a hand-curated frequently-bought-together block, best-sellers ranked by margin and a decent category page do most of the work, and they cost nothing to change.

What is the cold-start problem in recommendation systems?

It is the state where the engine has no interaction data for a new product, a new category or a new shopper, so it cannot personalise anything. Content-based and vector approaches degrade gracefully because they lean on catalogue attributes. Collaborative approaches degrade badly, because a product nobody has clicked has no learned representation. Vendor frameworks reflect this: Algolia puts NeuralSearch, AI Collections and real-time personalisation on the annual Elevate plan rather than the free tier, and Bloomreach tells prospects it is “not for” teams still building an audience and running a low-volume programme.

What is the difference between collaborative filtering and content-based filtering?

Collaborative filtering learns from what people did: shoppers who bought a linen shirt also bought chinos, so recommend chinos. It needs interaction data and fails on anything new. Content-based filtering reads what a product is: similar material, colour, price band and category get recommended. It needs catalogue data rather than traffic, and it works on a product that launched this morning. In production, serious engines blend both, because collaborative signal is more accurate once it exists and content signal is the only thing available before it does.

How is a recommendation engine different from a bought-together block?

A bought-together block is a fixed list, often written by a buyer or merchandising assistant, changed only when someone edits it. A recommendation engine computes the list per shopper per page load, drawing on that shopper’s session, history and the catalogue. The engine earns its fee when your traffic is high enough that per-shopper computation beats the curated guess, and when your catalogue depth and stock rotation make “what else did they buy” genuinely learnable.

Is Algolia a recommendation engine or just a search engine?

Both, and that is the point. Algolia’s Recommend API covers trending items, trending facets, related products, frequently bought together and “looking similar” recommendations, while NeuralSearch adds semantic ranking on top of keyword search. The catch is tier placement: Recommend is on the free plan at 5,000 requests a month, but NeuralSearch, AI Collections, Smart Groups and real-time personalisation sit on the annual Elevate plan, which has no published price. Price Algolia as a search platform with recommendations attached.

What happened to Dynamic Yield, Klevu, Searchspring and LiftIgniter?

All four changed hands. Checked on 16 October 2026: the klevu.com and searchspring.com pricing pages now both serve Athos Commerce content, following a merger announced 13 January 2025 combining Searchspring, Klevu and Intelligent Reach. The dynamicyield.com domain redirects to a Mastercard personalisation page. The liftigniter.com domain redirects to paradium.ai, where the personalisation engine is called Encore. Buy from the current entity, not the brand you remember.

Are the conversion-uplift figures these vendors publish real?

They are real numbers attached to specific customers, but they are vendor-commissioned and not independently reproducible. Algolia’s own pages disagree here, which is a useful warning. A Forrester Total Economic Impact study it announced on 13 March 2026 reports a 213% three-year ROI and $3.1M net present value for the composite organisation, while Algolia’s pricing page still quotes 382% ROI and $3.54M net present value from an earlier study. Fast Simon’s Ally Fashion page, published 3 May 2025, reports a 32% sessions increase and a 14% average-order-value increase from upsell recommendations. No retailer-published measurement is cited here, so treat every uplift figure as vendor-reported.

Do recommendation engines need cookie consent under GDPR?

If the engine builds a profile of the visitor, yes, and GDPR calls that profiling. Article 4(4) defines it as automated processing of personal data to evaluate personal aspects including preferences, interests and behaviour, and Article 22 gives a right not to be subject to solely automated decisions with legal or similarly significant effects. A product recommendation usually does not cross that line alone, but personalised bundles, credit offers and dynamic pricing can. Algolia’s AI Personalization Terms make the data relationship explicit, stating that the feature requires sending events to Algolia and implementing a consistent authenticated user token.

---

Sources

• Algolia pricing: free plan allowances, Grow and Grow Plus per-1K request pricing, Recommend at $0.60 per 1K, Elevate tier contents, index limits; its FAQ accordion also carries a conflicting Forrester ROI claim (382%, $4.47M present value of benefit, $3.54M net present value)
• Algolia AI Personalization Terms: requirement to send events and an authenticatedUserToken; 10 million daily Insights actions, five config changes per day
• Algolia Advanced Dynamic Re-ranking policy: hourly refresh rates, 24x re-ranking frequency, 5 facets, 100,000 queries, 100 hits
• Algolia Forrester TEI study post: commissioned Forrester study published 13 March 2026: 213% three-year ROI, payback under six months, $3.1M net present value
• Algolia Forrester TEI report landing page: 213% return on investment, $4.5M present value of benefit, $3.1M net present value
• Algolia NeuralSearch explained: vendor claim of 30 billion searches per week on vector search
• Recombee pricing: free/Standard/Pro/Premium/Platinum tiers, 20k to 4M catalogue items, extra traffic units at $0.70/$0.25/$0.20
• Recombee recommendation logics: named models including visually-similar, similar-properties, similarly-purchasing and search:semantic
• Recombee no-code widgets: embed-code widgets, grid/carousel/feed types, scenario-driven item-to-item and item-to-user endpoints
• Recombee introduction: separation of content-based properties, collaborative interactions, filters, boosters and constraints
• Recombee integration tips: guidance on master products versus variants and the low-traffic training warning
• Nosto pricing: GMV-and-traffic pricing formula, no self-service free trial, PoC process, uptime SLA claim
• Personyze pricing: five tiers from $0 to $1,450+/month, pageview and AI task allowances, overage rates, money-back guarantee
• Bloomreach pricing: “$60K+ a year for most brands”, 500K+ monthly visitor threshold, 1,400+ brands
• Bloomreach product recommendations: Loomi AI positioning and recommendation inputs
• Athos Commerce pricing: Onsite, Offsite and Complete Discovery plans, all request-quote; Conversational, Channel and GEO assistants
• Athos Commerce merger announcement: Searchspring, Klevu and Intelligent Reach combined; published 13 January 2025, updated 22 April 2025
• Athos Commerce homepage: GEO Assistant syndicating to Gemini and ChatGPT; 2,700+ brands claim
• Dynamic Yield on Mastercard: destination of the dynamicyield.com domain redirect
• Paradium.AI: destination of the liftigniter.com redirect; Encore personalisation engine
• Constructor product: Commerce Reasoning Engine, behavioural clickstream plus third-party data, reinforcement learning, real-time claim
• Fast Simon pricing: no public figures; sales-led model
• Fast Simon upsell and cross-sell: “continuously learns” vendor description
• Fast Simon Ally Fashion success story: 32% sessions, 14% AOV, 16% upsell sessions; published 3 May 2025, updated 1 January 2026
• Shopify predictiveSearch Storefront API: predictive search query, returned queries and products, limit scope and type
• Baymard cart abandonment statistics: 70.22% average across 50 studies, 42% “just browsing”, last updated 22 September 2025
• Baymard search bar UX research: 56% of sites failing search needs, published 2 October 2026
• Baymard ecommerce search research: methodology: 12 quantitative studies, 20,240 participants
• GDPR Article 4 definitions: Article 4(4) profiling definition
• GDPR Article 22: automated individual decision-making and its exceptions
• EU AI Act Annex III: high-risk categories; recommendation engines are not listed
• EU AI Act Article 50: machine-readable marking of synthetic content and deepfake disclosure
• Deep Learning to Address Candidate Generation and Cold Start Challenges in Recommender Systems: A Survey: cold-start framing; published July 2019
• Neural Collaborative Filtering: matrix factorisation background; submitted August 2017
• Self-Attentive Sequential Recommendation: SASRec; 2018
• Graph and Sequential Neural Networks in Session-based Recommendation: A Survey: session-based architecture survey; published August 2024
• No Attention, No Problem: Rethinking Session-based Recommendation with Pure Convolution: current session-based results; September 2026
• Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer: agentic commerce behaviour; September 2026
• Learning from Online User Feedback for Shopping Agents: shopping agent supervision; August 2026
• deployhyre.com: Best AI Product Recommendation Engines: 9 Tools: competitor comparison published 28 June 2026, updated 10 October 2026; still lists Klevu and Dynamic Yield as standalone vendors