# AI Recommendation Engines for Small and Mid-Sized Businesses: Cost, Data, and What to Skip

August 2026 • 13 min read 

_By the Silk Data engineering team. Based on production machine learning for retail, marketing, and publishing clients since 2015._

Most advice on AI recommendation engines is written for companies with a hundred million monthly visitors, then sold to companies with fifty thousand. The advice does not survive the translation, and the bill arrives anyway.

This guide is for owners and executives at companies from roughly twenty to five hundred people deciding whether an AI recommendation system is worth the investment. It covers the question that decides everything else, whether you have enough data for a model to learn anything, then the three routes to getting recommendations live and what each takes to run, where recommendations actually belong across your site and email, what changes now that customers shop through AI agents, the four pieces of common advice that will waste your money, and how to measure results so the number you report to your board is real.

Where the honest answer is that your existing platform already does this well enough, this guide says so.

## Key Takeaways

- Data volume decides feasibility before budget does. Below a few hundred thousand meaningful interactions, a custom recommendation system has nothing to learn from and platform-native features will beat it.
- There are three routes, not one: what your platform already includes, an ecommerce personalization tool you configure, or a custom build. Most growing businesses belong in the first two.
- The revenue mechanic is usually cross-sell into a larger basket and higher repeat purchase rate, not top-line conversion. Those are the numbers to instrument first.
- Measure with a holdout group. Without one you will attribute seasonal lift to the recommender and keep paying for it.
- Custom builds earn their cost above roughly a million meaningful interactions a year, or when your data cannot leave your infrastructure.
- Payback claims of three to six months assume you already have clean data and enough of it. Most companies do not, and closing that gap is the real project.

## What Do AI Recommendation Engines Actually Change in the Business?

Strip out the technology and a recommendation engine does three commercial things.

**It raises average order value through cross-sell and upsell.** A relevant complementary product shown at the right moment adds to a basket that was already converting. This is the fastest and most measurable effect, and for most catalogs it is where the majority of the return sits.

**It raises repeat purchase rate and customer lifetime value.** Customers who find things they want come back sooner. This effect is larger than the first one over a year and much harder to attribute, because it shows up as retention rather than as a line on an order. If you manage the business on LTV, this is the number that matters most and the one you will be tempted to stop waiting for.

**It makes the long tail visible.** Most catalogs have a handful of products that sell themselves and a long tail customers never reach. Recommendations surface that tail. For businesses with broad catalogs and thin margins on hero products, this is often the real prize, because the tail usually carries better margin.

The published evidence on personalization supports the direction while being vaguer than vendors imply. McKinsey's Next in Personalization research from 2021 found that faster-growing companies attribute a substantially larger share of revenue to personalization than slower-growing peers, on the order of 40 percent more, and that a large majority of consumers now expect personalized interactions and are frustrated when they are absent.

One number you will see everywhere deserves skepticism. The claim that Amazon draws 35 percent of its revenue from its recommendation engine is repeated constantly and has no primary source from Amazon. It appears to originate in early analyst estimates and has since become folklore. Treat it as an indication that recommendations matter at Amazon's scale, not as a benchmark for yours.

Two things a recommendation engine does not do. It does not fix a product range customers do not want. And it rarely moves top-line conversion much on its own, because the customers it influences were mostly already buying. If your problem is traffic that does not convert at all, recommendations are not the intervention.

## Do You Have Enough Data? Answer This Before Anything Else

This question decides whether the rest of the project is possible, and most vendor conversations skip it.

A recommendation system learns from interactions: views, clicks, add-to-carts, purchases, plays, reads. It needs enough of them to separate a genuine pattern from noise, and how many is enough depends on how rare the behavior you are predicting is.

These thresholds come from our own client work rather than from a benchmark. The pattern repeats across projects: where the behavior being predicted is rare, engineered features with a simple model beat a sophisticated one until interaction volume passes a threshold, and resampling tricks do not substitute for volume. Our [predictive analytics practice](https://silkdata.tech/predictive-analytics) is built around that constraint rather than around it.

For a business rather than an advertising platform, the practical thresholds look like this.

| Your situation | What is realistic | What to do |
|---|---|---|
| Under ~50k interactions/month, catalog under 500 items | Rules and manual curation outperform models | Use platform-native features. Curate collections by hand. Revisit in a year |
| ~50k-200k interactions/month, catalog 500-5,000 items | Basic collaborative filtering works; a custom build does not pay back | Configure what your platform or a specialist tool already gives you |
| ~200k-1M interactions/month, catalog 5,000-50,000 items | Specialist tools perform well; a custom build starts to make sense at the top of this band | Buy a specialist tool. Instrument properly. Reassess at 12 months |
| Over ~1M interactions/month, or data cannot leave your infrastructure | A custom build earns its cost | Run a scoped pilot before committing to a platform |

These bands are approximate and shift with how concentrated your behavior is. A business with 80,000 monthly interactions across 200 products has denser data than one with 300,000 across 40,000 products, and denser data works better. Count interactions per item, not interactions in total.

Two situations sit outside the table. If your catalog turns over heavily each season, as in fashion or seasonal goods, cold start is not an event you get through once but a permanent condition: a model trained on last spring has limited use this spring. The workable approach is to lean on product attributes rather than interaction history, so a new item inherits signal from similar past items, and to accept that the first few weeks of any season will be curated by humans. If you are a new business with no interaction history at all, no model applies yet. Spend the first year on attribute quality and category structure. That work is what makes recommendations possible later, and it improves conversion on its own in the meantime.

**The honest implication for smaller companies.** If you sit below the second band, the correct decision is almost always to use what you already have and spend the budget on product photography, on-site search, or category structure. Those produce more measurable lift at your size, and they are prerequisites for recommendations working later. For the discipline that makes any of this possible once you scale, see our [guide on data strategy for AI](https://silkdata.tech/blog/article/the-role-of-data-strategy-in-ai-success).

## Which Route Should You Take to Get Recommendations Live?

Almost every growing business chooses among three options. The differences that matter are cost, control, and where your customer data ends up.

| Route | Typical cost | Time to live | Control over logic | Best for |
|---|---|---|---|---|
| Platform-native (built into your ecommerce, CMS, or email platform) | Included or a small add-on | Days | Minimal, a few settings | Under ~200k monthly interactions, or anyone testing whether recommendations matter at all |
| Ecommerce personalization platform (specialist SaaS) | Monthly fee scaling with traffic or revenue | 2-6 weeks | Moderate, configurable rules and business constraints | Most mid-market businesses; broad catalogs; teams without ML staff |
| Custom build | Project cost plus ongoing running and monitoring | 3-6 months to production | Full | Over ~1M interactions; unusual catalog structure; data residency or regulatory constraints |

**Start with platform-native even if you expect to outgrow it.** It costs almost nothing to switch on and produces the one thing you cannot get any other way: a real measurement of what recommendations are worth in your specific business. That number justifies or kills the next step. Buying a specialist tool without it means negotiating from a position where you cannot value what you are buying.

**Specialist tools are where most mid-market businesses should land.** They handle the cold-start problem, business rules such as excluding out-of-stock items or enforcing margin floors, and the operational work of keeping a model fresh. The trade-off is lock-in and per-traffic pricing that grows with your success.

**Custom builds are justified by constraints, not by ambition.** Three constraints do it: volume beyond what specialist pricing makes sensible, a catalog whose structure no standard tool represents, or data that cannot leave your infrastructure. That last one is the most common reason clients come to us. For a marketing client we deployed a [local language model inside their own platform](https://silkdata.tech/case-studies/local-llm) so campaign data and customer records never crossed the perimeter, which mattered more to their legal team than any accuracy gain would have.

Before signing with any vendor, our [guide to evaluating AI vendors](https://silkdata.tech/blog/article/evaluating-ai-vendors-for-enterprise-2026-guide) covers the contract terms that decide whether you can leave later: data export, model version notification, and what happens to your configuration when you cancel.

**What it takes to run, whichever route you choose**

The question owners ask second, after cost, is who on the team does this. The honest answer depends on the route and it is smaller than people fear.

Platform-native features need a few hours to configure and someone looking at the numbers monthly. No new hire, no new skill.

A specialist tool needs an owner spending roughly two to four hours a week: reviewing what the model is promoting, adjusting business rules when merchandising priorities change, and checking that the feed of product data has not broken. That owner should sit in merchandising or marketing, not in IT, because most of the decisions are commercial. The failure mode is assigning it to whoever set it up and then losing them to another project.

A custom build needs engineering capacity permanently, not just during the project. Someone has to respond when the pipeline fails at two in the morning, and someone has to retrain when the data drifts. If you do not have or plan to have that capacity in house, either buy instead or budget for a support arrangement with whoever builds it. A custom recommender without an owner degrades quietly and takes about a year to visibly fail.

On integration, the practical question is whether your product data lives in one place that other systems can read. Businesses running a single ecommerce platform with a clean product feed connect a specialist tool in days. Businesses with product data split between an ERP, a spreadsheet a merchandiser maintains, and the platform itself spend most of the project reconciling those three. Which situation you are in is knowable in an afternoon and worth checking before you talk to vendors.

## Where Should Recommendations Actually Appear?

Most discussion of recommendations assumes a widget on a product page. For a growing business that is usually the third most valuable placement, not the first.

**Email and messaging is where the fastest return sits.** Abandoned cart, browse abandonment, post-purchase cross-sell, and replenishment reminders all use the same underlying signal as an on-site recommender, and they reach customers who have already left. Most email platforms include some version of this, which makes it the cheapest place to test whether personalization moves anything in your business. It also produces clean measurement, because email gives you a defined send list and a natural control group.

**On-site placement matters most on two pages.** The cart, where a complementary item is a genuine cross-sell rather than a distraction, and the empty-search or zero-results page, where a recommendation rescues a session that was otherwise lost. Homepage personalization is the placement vendors demo and the one that moves the least, because homepage visitors have not expressed intent yet.

**Category and search pages are the quiet win.** Reordering results by predicted relevance for the individual customer is technically the same problem as recommending, and it affects more sessions than any widget does. If your platform supports personalized sort order, that is often the single highest-leverage setting available to you.

A practical sequence for a business testing this for the first time: email flows, then cart, then personalized search ranking, then everything else. That order goes from cheapest and most measurable to most expensive and hardest to attribute.

## What Happens When Your Customer's AI Agent Does the Shopping?

Everything above assumes your customer is on your site. In 2026 that assumption stopped being safe, and the shift changes what you should build first.

By Adobe's Digital Economy Index, traffic reaching retailers from AI sources has grown roughly twelvefold while traditional search traffic declined about 10 percent year over year. A protocol layer formed around this quickly: the Agentic Commerce Protocol from Stripe and OpenAI behind ChatGPT Shopping, Google's Universal Commerce Protocol introduced at NRF 2026, and Anthropic's Model Context Protocol as the connectivity layer that lets an agent query live inventory, pricing, and product detail rather than scraping a rendered page. Shopify shipped Agentic Storefronts in January 2026, which makes a store discoverable inside ChatGPT, Perplexity, and Copilot from a single setting in the admin.

**What this means practically.** You now have two recommendation surfaces. One is your site, where the guidance above applies. The other is somebody else's agent, where your recommendation widget is invisible and the thing that decides whether your product appears at all is the quality of your structured data. Machine-readable attributes, accurate availability, real shipping timelines, and return terms exposed as fields rather than buried in prose. A product description written only for humans limits what an agent can match you against. If an agent cannot filter on a constraint the customer stated, your product does not enter the comparison.

**The honest counterweight.** Consumer behavior is ahead of consumer commitment. Surveys through 2026 put AI use somewhere in the journey for a large majority of shoppers, mostly for product ideas, review summaries, and price comparison, while the share who have actually completed a purchase through an AI referral remains in the low teens. Around seven in ten say they are at least somewhat comfortable delegating a purchase to an agent, which is a statement about willingness rather than behavior. Microsoft reported in January 2026 that Copilot users convert substantially faster when shopping intent is present, and that merchants remain the merchant of record, keeping the transaction, the customer data, and the relationship.

**So what should a growing business do now.** Not move budget from your site to agents. Two cheap things instead. First, check whether your platform already has an agentic setting and switch it on, the same logic as starting with platform-native recommendations. Second, audit your product data for machine readability, which is work you would eventually do anyway and which improves conventional search visibility at the same time.

## Four Pieces of Common Advice That Will Cost You Money

**"You don't need much data."** You do. This is the most expensive piece of optimism in the category, and it is repeated because it shortens the sales cycle. A model trained on thin data does not produce mediocre recommendations, it produces confident recommendations based on coincidence, which is worse because you cannot tell by looking. The thresholds above are honest ones. If a vendor tells you volume does not matter, ask what recall they achieved on a catalog your size and how they measured it.

**"Expect payback in three to six months."** Sometimes true, and it assumes your product data is clean, your customer identity is resolved across devices and sessions, and your interaction history is complete. In practice, data preparation consumes the majority of effort on real projects and is invisible to everyone outside the team doing it. Budget for it explicitly or it appears as a delay you cannot explain to your board.

**"More personalization is always better."** Recommendation models reward what already sells. Left alone they narrow your visible catalog to existing bestsellers and slowly erode discovery, which surfaces a year later as declining repeat purchase rate while your dashboards look healthy. Insist on diversity constraints and accept a small short-term drop in click-through as the price of long-term catalog coverage. That is a business decision, not a technical setting. There is a second limit worth naming: personalization that visibly demonstrates how closely you are watching makes customers uncomfortable, and the reaction lands on the brand rather than on the algorithm. Recommendations that feel helpful outperform recommendations that feel observant, and the difference is usually in how explicitly you reference the customer's own behavior back to them.

**"The algorithm is the hard part."** The hard parts are knowing which customer is which across sessions, keeping product attributes consistent, and having an honest measurement setup. Businesses that fix those with a simple model outperform businesses that buy a sophisticated model and feed it inconsistent data. In our [large-scale image search work](https://silkdata.tech/case-studies/large-scale-image-search-system) indexing over six million assets, consistent preprocessing of the inputs mattered more to the result than which model did the matching. The same holds for catalogs where the recommendable content is text rather than products, which our [guide on how AI handles unstructured data](https://silkdata.tech/blog/article/how-ai-handles-unstructured-data-for-better-decisions) covers in more depth.

## How to Measure It Without Fooling Yourself

Most reported recommendation ROI is inflated, usually not deliberately. The mechanism is simple: you launch, revenue rises, you attribute the rise to the launch. Seasonality, a marketing campaign, and a pricing change all hide inside that number.

**Use a holdout group.** Keep 5 to 10 percent of traffic without recommendations and compare the two groups on the same days. This single practice separates real measurement from storytelling, and every serious tool supports it. A vendor who cannot run a holdout is telling you something about how their other customers report results.

**Track four numbers, in this order:**

- **Average order value** for exposed versus holdout traffic. Fastest to move, easiest to attribute
- **Repeat purchase rate** over 90 days, which feeds customer lifetime value and is the number that determines whether this was worth doing
- **Catalog coverage**, meaning the share of your products shown to anyone. Falling coverage is the early warning that the model is collapsing onto bestsellers
- **Incremental revenue per thousand sessions**, the number to put in front of your board because it converts the effect into money without pretending precision you do not have

Click-through rate on the recommendation widget is the metric vendors report and the least useful of the set. It rises whenever recommendations get more aggressive and says nothing about whether total revenue moved.

**Give it a full business cycle.** Repeat purchase effects take a quarter to appear. Judging at four weeks measures novelty, not value.

## Why Do Recommendation Projects Fail Commercially?

The technical failures are documented elsewhere. These are the ones that cost money at your size.

**Nobody owns it.** The tool goes live during a project, the project ends, and eighteen months later nobody can say whether it is still helping. Recommendations need one named owner who reviews the numbers quarterly and has authority to turn it off. Without that you pay a subscription indefinitely for something nobody has evaluated since launch.

**Pricing scales faster than benefit.** Specialist tools commonly price on traffic or revenue. Your bill grows with your success while incremental benefit flattens, because the easy gains come first. Model the cost at three times your current traffic before signing, and put a review trigger in your calendar at that volume.

**The pilot becomes permanent by default.** A trial with ambiguous results often just continues, because cancelling requires a decision and renewal does not. Set the go or no-go date before the pilot starts and put a number on what success means.

**It gets bought instead of the thing that was actually broken.** Recommendations are an attractive purchase because they feel modern and the demo looks good. Weak on-site search, poor product photography, unclear category structure, and slow pages all suppress revenue more than missing recommendations do, and all are cheaper to fix. If more than one of those is true in your business, fix them first.

## What the Rules Require, Briefly

Two regulatory points matter for a personalization project in Europe, and neither is usually a blocker.

**[GDPR](https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng) applies to the profiling underneath.** Building behavioral profiles of identifiable customers is processing personal data, so you need a lawful basis, a privacy notice that actually describes what you do, and a retention period. Recommendations rarely reach the threshold of a solely automated decision with legal or significant effect, so the stricter provisions on automated decision-making typically do not apply, but record that assessment rather than assuming it. Our [guide to data privacy in AI deployment](https://silkdata.tech/blog/article/the-role-of-data-privacy-in-ai-deployment-for-leaders) covers how this translates into architecture.

**The [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), in most cases, does not classify a product recommender as high-risk.** The high-risk categories cover employment, credit, education access, and similar consequential decisions. A recommender suggesting products falls outside them. If your system also influences hiring, lending, or access to services, the classification changes and the timeline matters: the Digital Omnibus on AI, adopted in June 2026 and in force from July, moved standalone high-risk obligations to 2 December 2027, with AI embedded in already-regulated products moving to 2 August 2028. Transparency obligations under Article 50 were not deferred and applied from 2 August 2026.

Separately, the EU Digital Services Act imposes recommender transparency obligations on very large online platforms, including explaining the main parameters that shape what users see. Businesses at the scale this guide addresses fall outside that designation, but if you operate a marketplace with substantial reach in the EU, confirm your status with counsel rather than assuming you are exempt.

## Where Silk Data Fits

Two situations bring growing businesses to us on personalization work, and one situation sends them away.

The first is volume or catalog structure that specialist tools handle badly: broad catalogs with inconsistent attributes, marketplaces where the relationship between sellers, products, and buyers carries the signal, or media archives where the recommendable unit is not a product. Our work on [automatic news analysis](https://silkdata.tech/case-studies/automatic-news-analysis) for a publishing client and on [procurement software for an international retailer](https://silkdata.tech/case-studies/procurement-software) both dealt with catalogs that standard tooling represented poorly.

The second is data that cannot leave your infrastructure, whether for regulatory, contractual, or customer-commitment reasons. That is where on-prem deployment stops being a preference and becomes the only viable architecture.

The situation that sends businesses away is being below the data threshold. When a company arrives convinced it needs a custom recommendation engine and its interaction volume says otherwise, we say so and point them at their platform's built-in features plus a list of cheaper things that will move revenue more. That conversation has cost us projects. We prefer it to building something that cannot work.

If you want to know which band you are in before committing budget, our [AI proof of concept](https://silkdata.tech/ai-proof-of-concept) engagement answers that in weeks rather than quarters and ends in a documented go or no-go rather than a platform you now have to maintain. For the wider picture of what we build, see our [custom AI development](https://silkdata.tech/artificial-intelligence-development) and [AI consulting](https://silkdata.tech/ai-consulting) pages.

## Frequently Asked Questions

###   How much data do I need for an AI recommendation engine to work?  

Enough interactions per item for a pattern to separate from noise, which usually means at least a few hundred thousand meaningful interactions and a catalog small enough that each item accumulates history. Count interactions per product rather than in total: 80,000 monthly interactions across 200 products is denser and works better than 300,000 across 40,000. Below roughly 50,000 monthly interactions, hand-curated collections and platform-native features will outperform any model. 

###   How much does an AI recommendation system cost?  

Platform-native features are included in your existing subscription or offered as a small add-on and go live in days. Specialist ecommerce personalization platforms charge monthly, usually scaling with traffic or revenue, and take two to six weeks to configure. Custom builds run as a project over three to six months, plus ongoing running and monitoring costs that never end. Model the specialist-tool cost at three times your current traffic before signing, because the pricing grows with your success while the incremental benefit flattens. 

###   Do I still need on-site recommendations if customers use AI shopping assistants?  

Yes, and you need something additional. On-site recommendations serve the customers who reach your site, which remains the majority of purchases. AI agents serve customers who may never visit, and what determines whether your product appears in an agent's comparison is the quality of your structured product data rather than any recommendation model. The practical answer for a growing business is to do both cheaply: switch on whatever your platform offers for both, and prioritize making product attributes, availability, and shipping terms machine-readable. Moving budget away from your site would be premature, since the share of purchases actually completed through an AI referral is still small. 

###   Where should I put recommendations first, on my site or in email?  

Email, in most cases. Abandoned cart, browse abandonment, and post-purchase flows use the same signal as an on-site recommender, reach customers who have already left, and are usually included in the email platform you already pay for. Email also gives you cleaner measurement, because you have a defined send list and a natural control group. After email, the highest-value on-site placements are the cart and the empty-search page, followed by personalized sort order on category and search pages. Homepage personalization is the placement vendors demo most and the one that moves the least. 

###   Should a small business build a custom recommendation engine?  

Almost never. Custom builds earn their cost above roughly a million meaningful interactions a year, or when data cannot leave your own infrastructure for regulatory or contractual reasons. Below that, platform-native features cost nothing to switch on and a specialist tool covers most of what a custom build would do. Start with what your platform includes, measure the effect with a holdout group, and let that number decide the next step. 

###   How do I know whether recommendations actually increased revenue?  

Hold back 5 to 10 percent of traffic from recommendations and compare the two groups on the same days. Without a holdout you cannot separate the recommender's effect from seasonality, campaigns, or pricing changes. Track average order value and repeat purchase rate over 90 days rather than click-through on the widget, which rises whenever recommendations get more aggressive and says nothing about total revenue. 

###   Does GDPR allow personalized product recommendations?  

Yes, with the usual requirements. Profiling identifiable customers is processing personal data, so you need a lawful basis, a privacy notice describing what you actually do, and a defined retention period. Product recommendations rarely constitute a solely automated decision with legal or significant effect, so the stricter rules on automated decision-making typically do not apply, but record that assessment rather than assuming it. 

###   Will recommendations narrow what customers see over time?  

Yes, if nobody prevents it. Models reward what already sells, so left alone they concentrate visibility on existing bestsellers and reduce discovery. The effect appears a year later as declining repeat purchase rate while short-term metrics look fine. Track catalog coverage, meaning the share of products shown to anyone, and set diversity constraints even though they cost a little click-through in the short term. 

**Discuss your needs with our specialists!**  Contact us