AI for Product Descriptions

Status 2024-08-20

Stibo is finishing the implementation of the first iteration of the product descriptions, We will get an explanation and handover 240829. After that we will test and interate on this together with Product Managers (Produktschefer), Master Data managers, and SEO Manager.

I propose the initial evaluation phase to last a month. After that we will collect our impressions and make a decision on how to proceed.

Status 2024-05-24

We will collaborate with Stibo to get the text generation in STEP where all the master data and product managers work. The initial idea was to create a separate app for an initial pilot but it is probably better to get it into STEP as soon as possible. However, since things take some time we might have to work with what we can get hold of for the moment.

We have two different use cases, one is where a product manager is onboarding a new product and is working directly in STEP. Another is where we want to automate generation of thousands of product descriptions, which can possibly be done outside of STEP and the results later fed into STEP.

Big Room info: this initiaitve has had a slow start but we have now support to get this going for a proper discovery in Q3 with preparation before the summer.

Background

The product texts in STEP, which are shown in the web and other channels, are of varying quality. The reasons for this can be summarized as follows.

• Ahlsell largely relies on the supplier to provide the article text.

• A decentralized model is emplyed where the Product manager for a certain line of products is responsible for the quality.

• Retail, EMV, Proffsmagasinet have different requirements and therfore their own procedures to modify the text. For end users the text needs to be more selling

• SEO aspects are not considered, which make Ahlsell products not end up high up in a Google search. Addition of meta fields would help a lot here.

• As a connected note, products usually need to be translated to all the Nordic languages, as well as to English. This is investigated in a separate initiative (AI Translations)

What has been done already

Before the summer 2023 we presented a Proof of Concept that we showed could generate a text from the product attributes. We showed that the text could be written in sifferent languanges and showed that it can have differnt degrees of positivity. During the autumn of 2023 we did a small pilot together with EMV where we looked into 10 product texts that were not exclusively based on the attributed, and tried to recreate them using keywords and small phrases.

Business case

1. Efficiency: in the cases where new texts are written (EMV) or largely rewritten (Retail, PM) we hope to significantly reduce the time it takes to (re)write a product text. [Can this be estimated somehow?]

2. Quality. We can ensure a more even quality of our products,

3. Customer experience: better descriptions can be more informative and give more value to the customer.

4. Increased sales: it is possible that better product texts will increase sales

5. Findability: when using Google we need to make sure that our products end up high on the list to increase traffic to our pages compared to our competitors.

Stakeholders

POs Mathias Ljungberg, Tomas Jönsson, Christian Englén
Masterdata: Maria Fagrenius
Proffsmagasinet: Peer Dahlgren
Inköp: Karolina Hagberg-Chinell
SEO: Daniel Wellingford.

Proposed pilot

We propose the following.

• We create a tool, such as a web app (called App in the following), where the user can generate translations from provided texts and attributes.

  • We need an easy way to get attributes and text from the STEP portal. A straight forard way is to export an excel file that is then uploaded into the App.

  • Prompts and options should be possible to change. This is something we want to iterate on. The prompts should also preferrably be saved.

  • We should aim have the option to generate several types of texts at the same. For example, texts for long and short descriptions, metadata for SEO, possbilty texts for Ahslell and retail at the same time, or in all languanges. In the first pilot we would need to restrict ourselves.

  • Manual changes will have to be made in the descriptions. The generated text and the corrected one should be saved for future reference.

  • The text from the supplier should always be saved for reference.

  • It should also be possible to use the App sequatially by several persons that are responsible for their own texts.

• We propose a pilot where a number of persons working on onboarding products test the App. Preferrably, we should have test user from more than one out of the following groups because the use cases are slightly different.

  • Ahlsell

  • Retail

  • PM

  • EMV

• Time scope: Q2

• Resource requirements:

  • 0.5 developer in order to support the App and the testers.