AI Product Recommendations for an Ecommerce Catalog

A first recommendation feature is one catalog, one signal, and a result a merchant can explain and switch off. It is not a personalization platform.

By Muhammad Saif and Aashan Khan · Techtribe co-founders

A merchant searching for a Dubai company that builds AI product recommendations is usually past the slogan. The catalog is live. Shoppers already buy something. The question is whether a short list of "you may also want" can be added without inventing a second platform, and who can build that list so a merchandiser can understand it.

This note answers the build. It is not a ranking of agencies in Dubai or anywhere else. We are not a Dubai company. Techtribe is based in Islamabad and works remotely with teams in the UAE, which is described on the AI development agency UAE page. If you want a local office tour, this is the wrong engagement. If you want a recommendation feature in the store you already run, keep reading.

What the first version is

One catalog. The products you already sell, with the fields a recommendation is allowed to use: title, category, price, availability, and whatever text you trust. Not a new product information system.

One signal. Views, purchases, or a sentence the shopper types. Pick one for the first release. A signal you do not have data for is a fantasy. A signal you have but cannot join to a product id is a data project, and it should be named before anyone talks about a model.

A result a person can explain. "These are other in-stock jackets in the category this shopper just viewed." Or "these are the three products whose descriptions are closest to the sentence they typed." If the only explanation is "the model thought so," your merchandiser cannot correct it, and your customer cannot trust it when it is wrong.

An off switch. A flag that removes the list and leaves the product page alone. Recommendations will be wrong on the first week. The store has to survive that week. A feature you cannot turn off is not a feature. It is a coupling.

That is small enough to be a two-week LLM MVP when it is truly one flow: open a product, see the list, and the list comes from one signal. It stops being that package when you also want accounts for merchandisers, three data sources, a redesign of the product page, and an evaluation suite because the list is already in front of every shopper. That larger shape is the four-to-six-week Production MVP. Do not buy the small one and then narrate the large one.

What it is not

It is not a personalization platform. Real-time profiles, a dozen placements, email, the app, the ads, and a control room for marketers are a product line. They can be a later roadmap. They are not the first pull request.

It is not a promise about conversion. We will not tell you a percentage lift. Anyone who opens with a lift, before they have seen your catalog and your traffic, is guessing in your brand's voice. Measure the list after it ships: how often it renders, how often it is clicked, how often the click is in stock, how often a merchandiser hides an item. Those are facts you can collect. A forecast is not.

It is not a reason to replace the search you have, unless search is the actual wound. Related items on a product page and "search that understands a sentence" are different features. Do one.

It is not a scrape of someone else's store. The catalog is yours. The behavior, if you use behavior, is your shoppers', under whatever notice you already owe them. A recommendation feature that starts by copying a competitor's site is a legal problem wearing a model.

Where a model helps, and where a query is enough

Many "AI recommendation" pitches are a sort. Same category, in stock, not the current product, ordered by recent sales. That query is explainable, cheap, and often better than a model on a thin catalog. If your merchandiser can already describe the rule, build the rule. Use a model when the input is language you do not want to hand-code: "a waterproof jacket for a wet commute, under this price," asked in the shopper's words.

When the model is in the loop, ground it in the catalog. The list it returns should be product ids that exist and are purchasable, not names it invented. That is retrieval, and the service that describes grounding a model in your own material is RAG development. The catalog is the corpus. A paragraph the model wrote about a product you do not sell is a defect, not a flourish.

Keep the model out of price, tax, and stock. Those are columns in your system. The recommendation may filter on them. It must not invent them. A shopper who is shown an out-of-stock size because the prompt was confident will not be moved by an explanation of temperature.

What the merchandiser needs on day one

A way to see why a product was listed, in the words of the signal you picked. A way to hide a product from recommendations without deleting it from the store. A way to turn the whole feature off. A small set of example pages where the list is agreed to be right, and a small set where it is agreed to be wrong. Those examples are the eval, even if the first version is too small to have a formal suite.

Without the hide button, the first complaint becomes an engineering ticket, and the engineering ticket becomes a prompt, and the prompt changes the rule for every product. That is how a single bad pairing rewrites the feature. The hide button is the pressure valve.

A merchandiser in Dubai or anywhere else does not need a new login to a separate AI studio. They need the control on the store they already use, or on a single admin page next to it. A second product they have to learn is a second product you have to staff.

Data you must have before kickoff

Product ids that do not change every import. A category or a text field you trust. Availability that is true often enough to filter on. If the signal is purchases, an export or an API that joins an order line to a product id. If the signal is a typed query, a page where the query can be entered, even a plain one.

If the catalog is a spreadsheet someone updates by hand, say so. The first integration might be that sheet, and the fragility of the sheet is part of the scope. If the catalog is a platform with a clean API, the integration is smaller and the rule can be stricter. We would rather know which one it is on the call than discover a nightly CSV with renamed columns in week two.

Images are not required for a first list of titles and prices. They become required if the design you approved shows cards. Design is how a one-week integration becomes a longer package. Decide whether the first release is a simple list. You can make it beautiful after it is true.

How a UAE team usually wants this to run

The work is remote. Milestones and overlap are agreed up front, the same way we agree them with teams in other regions. You do not need a local entity on our side for a catalog feature. You do need a person who can give read access to the catalog and who can accept or reject the list on a staging page.

Bring the staging page, or accept that the first task is a safe place to show the list that is not the live homepage. Turning a model on in front of every shopper on day one is how a bad join becomes a public joke. Stage it. Let a merchandiser break it. Then put it on one template, not on every surface.

If the store is already a SaaS you operate for other merchants, the feature has a tenant boundary. One merchant's catalog and one merchant's behavior do not train or retrieve across the wall. That requirement alone can move the work from the two-week package to the Production MVP, because it is no longer one flow on one catalog. Say "we are the platform" on the call if that is what you are.

What to ignore in a sales conversation

A map of the UAE with pins. A claim that more data always makes the list better, before anyone has looked at your data. A required rebuild of checkout. A model brand presented as the feature. A percentage. A monthly retainer slipped into a feature that has not shipped.

The questions that matter are dull. Which signal. Which page. What "wrong" means. Who can hide a product. Where the code will live. Whether this is one catalog or many. Those questions fit in the thirty-minute scope call. The rest is decoration.

A first release you can describe in one paragraph

On the product page, under the fold, show four in-stock products from the same category, excluding the current one, ranked by a rule you can say out loud. If you add language, add a box that takes a sentence and returns ids from the catalog only. A merchandiser can hide a pairing. You can turn the block off. The code is in your repository. The catalog credentials are in your account.

If that paragraph is the product, we can scope it as the narrow MVP. If your paragraph includes several markets, a new app, and a customer data platform, we will say so and quote the larger scope or tell you to cut. The AI MVP development page is the service. The packages are the two fixed shapes. Neither includes a made-up conversion rate.

Ship the paragraph. Count clicks for a little while. Then decide whether the second signal is worth it. Most catalogs deserve the paragraph long before they deserve a platform.

A staging week that earns the live page

Put the list on a staging product page that uses a copy of the catalog, or on a live template only your staff can see. Click ten products you know well. For each, write whether the four results are acceptable. That sheet is more useful than a model score. If three of the ten are nonsense, do not debate temperature. Look at the join. Wrong category ids, stale stock, and duplicate colors dressed up as variety are the usual nonsense, and they are fixed in the data or the rule.

Ask one merchandiser who was not in the scoping call to do the same ten. If they cannot tell why an item appeared, the explanation string is missing, and the feature is not ready for shoppers. Add the string. It can be plain. "Same category. In stock. Often bought with this item." Plain beats a paragraph the model composed.

Only then turn it on for a slice of traffic, if your platform can slice, or for a single collection. Watch the hide button. If nobody uses it and nobody complains, you may have a dull, correct list, which is a win. If the hide button is used on the same pairing all day, exclude that pairing in the data and look for the pattern. A pattern is a rule you can add. A one-off is what the hide button is for.

Keep the model provider and the store credentials in the merchant's account. A recommendation feature whose key lives in an agency's dashboard will go dark when the contract does, and it will keep calling the model until someone notices. The key belongs to the merchant, in an account their own team can open.

When the catalog is messy

Messy is normal. Products with three names for one color. Categories that mean different things in different markets. A description field that is a paste of the manufacturer's legal line. Do not plan to clean the entire catalog as part of the first feature. Clean the fields the signal uses, for the categories you will show the list on, and leave the rest. A cleanup of forty thousand descriptions is a content operation. It can run beside the feature. It should not be disguised as the feature.

If the same product id is reused for a different item, stop. Recommendations on unstable ids recommend the wrong object and then learn from that mistake if you ever feed behavior back. Stability of the id is a prerequisite, not a nice-to-have. Your platform owner knows whether ids survive a re-import. Ask them before kickoff, not after the first wrong jacket.

Multi-language catalogs need a rule too. Recommend in the language of the page, from the fields in that language, or say you are not doing the other language yet. A model that translates on the fly will invent attributes. Invented attributes are how a careful category page becomes a complaint.

What you can decide without us

You can choose the signal this week. You can export ten products and write the list a human merchandiser would have shown. That human list is the acceptance test. If we cannot beat it or match it for a cheaper ongoing cost, the model does not belong in the first release, and a query does. Bring the human list to the call. It makes the scope honest, and it keeps the conversation on your catalog instead of on a platform diagram. If you cannot make the human list because nobody knows the catalog, the first week is merchandising, not a model, and we will say that before we quote a build. A list a human cannot explain is not a feature we should automate.

If this is your app, send the repo.

A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.

How Techtribe scopes a recommendation feature

If the work is one catalog, one signal, and a list you can explain and turn off, it can be the Basic MVP: two weeks, one AI or retrieval feature, code in your repo. If several catalogs, an admin, or evaluations for a list shoppers already depend on are in scope, it is the Production MVP. We will not quote a lift, and we will not pretend to be a Dubai agency. Teams in the UAE hire us remotely. That is described on the UAE AI development page.

The scope call is 30 minutes with a founder. Bring the catalog source, the page the list should sit on, and two examples of a good recommendation and one of a bad one. If you prototyped the storefront in a builder, Lovable, Bolt, v0, and Replit cover what those tools export. The service page is AI MVP development. When the list must be grounded in your own products, the related service is RAG development.

Questions

What is a first AI product-recommendation feature?

One catalog, one signal such as views, purchases, or a text query, and a short list the merchant can explain. It has an off switch. It is not a full personalization platform, and it does not require a new storefront.

Do you need to rebuild the store to add recommendations?

No. The first version should read the catalog you already have and render a list on a page you already have. A rebuild is a separate decision, made when the catalog or the checkout cannot support the feature at all.

Can a Dubai or UAE merchant hire you for this?

Yes. We work remotely with product teams in Dubai, Abu Dhabi, and across the UAE. We are based in Islamabad. The engagement is the recommendation feature, not a directory of local agencies, and the overlap and milestones are agreed on the scope call.

How long does the first version take?

If it is one flow and one model or retrieval job, it can fit the two-week Basic MVP. If it needs accounts, several integrations, and evaluations because customers depend on it, it is the four-to-six-week Production MVP. The timeline article explains that split.

Book a free call

Have a vibe-coded app in production?

30-minute free call. We'll read what you have, name what is dangerous, and quote a fixed scope.